Add spans

WIP add more logs
Make the merger multithreaded
2025-12-04 03:35:43 +00:00 · 2024-09-26 17:20:32 +02:00 · 2024-09-26 16:37:38 +02:00 · 2024-09-26 11:09:06 +02:00 · 2024-09-25 22:42:41 +02:00 · 2024-09-25 22:15:15 +02:00
118 changed files with 8159 additions and 3935 deletions
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -386,9 +386,8 @@ checksum = "96d30a06541fbafbc7f82ed10c06164cfbd2c401138f6addd8404629c4b16711"

 [[package]]
 name = "arroy"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dfc5f272f38fa063bbff0a7ab5219404e221493de005e2b4078c62d626ef567e"
+version = "0.4.0"
+source = "git+https://github.com/meilisearch/arroy/?rev=2386594dfb009ce08821a925ccc89fb8e30bf73d#2386594dfb009ce08821a925ccc89fb8e30bf73d"
 dependencies = [
 "bytemuck",
 "byteorder",
@@ -472,7 +471,7 @@ checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"

 [[package]]
 name = "benchmarks"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "anyhow",
 "bytes",
@@ -528,7 +527,7 @@ dependencies = [
 "proc-macro2",
 "quote",
 "regex",
- "rustc-hash 1.1.0",
+ "rustc-hash",
 "shlex",
 "syn 2.0.60",
 ]
@@ -653,7 +652,7 @@ dependencies = [

 [[package]]
 name = "build-info"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "anyhow",
 "time",
@@ -935,8 +934,7 @@ dependencies = [
 [[package]]
 name = "charabia"
 version = "0.9.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "55ff52497324e7d168505a16949ae836c14595606fab94687238d2f6c8d4c798"
+source = "git+https://github.com/meilisearch/charabia?branch=mutualize-char-normalizer#f8d8308cdb8db80819be7eeed5652cc4a995cc71"
 dependencies = [
 "aho-corasick",
 "csv",
@@ -1623,7 +1621,7 @@ dependencies = [

 [[package]]
 name = "dump"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "anyhow",
 "big_s",
@@ -1835,7 +1833,7 @@ checksum = "9fc0510504f03c51ada170672ac806f1f105a88aa97a5281117e1ddc3368e51a"

 [[package]]
 name = "file-store"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "tempfile",
 "thiserror",
@@ -1857,7 +1855,7 @@ dependencies = [

 [[package]]
 name = "filter-parser"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "insta",
 "nom",
@@ -1877,7 +1875,7 @@ dependencies = [

 [[package]]
 name = "flatten-serde-json"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "criterion",
 "serde_json",
@@ -2001,7 +1999,7 @@ dependencies = [

 [[package]]
 name = "fuzzers"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "arbitrary",
 "clap",
@@ -2222,11 +2220,11 @@ checksum = "d2fabcfbdc87f4758337ca535fb41a6d701b65693ce38287d856d1674551ec9b"
 [[package]]
 name = "grenad"
 version = "0.4.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "350d89047298d3b1b40050acd11ab76e487b854a104b760ebc5a7f375093de77"
+source = "git+https://github.com/meilisearch/grenad?branch=various-improvements#58ac87d852413571102f44c5e55ca13509a3f1a0"
 dependencies = [
 "bytemuck",
 "byteorder",
+ "either",
 "rayon",
 "tempfile",
 ]
@@ -2309,9 +2307,9 @@ dependencies = [

 [[package]]
 name = "hashbrown"
-version = "0.14.3"
+version = "0.14.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "290f1a1d9242c78d09ce40a5e87e7554ee637af1351968159f4952f028f75604"
+checksum = "e5274423e17b7c9fc20b6e7e208532f9b19825d82dfd615708b70edd83df41f1"
 dependencies = [
 "ahash 0.8.11",
 "allocator-api2",
@@ -2553,7 +2551,7 @@ checksum = "206ca75c9c03ba3d4ace2460e57b189f39f43de612c2f85836e65c929701bb2d"

 [[package]]
 name = "index-scheduler"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "anyhow",
 "arroy",
@@ -2571,6 +2569,7 @@ dependencies = [
 "meili-snap",
 "meilisearch-auth",
 "meilisearch-types",
+ "memmap2",
 "page_size",
 "rayon",
 "roaring",
@@ -2592,7 +2591,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "168fb715dda47215e360912c096649d23d58bf392ac62f73919e831745e40f26"
 dependencies = [
 "equivalent",
- "hashbrown 0.14.3",
+ "hashbrown 0.14.5",
 "serde",
 ]

@@ -2651,8 +2650,7 @@ checksum = "28b29a3cd74f0f4598934efe3aeba42bae0eb4680554128851ebbecb02af14e6"
 [[package]]
 name = "irg-kvariants"
 version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ef2af7c331f2536964a32b78a7d2e0963d78b42f4a76323b16cc7d94b1ddce26"
+source = "git+https://github.com/meilisearch/charabia?branch=mutualize-char-normalizer#f8d8308cdb8db80819be7eeed5652cc4a995cc71"
 dependencies = [
 "csv",
 "once_cell",
@@ -2747,7 +2745,7 @@ dependencies = [

 [[package]]
 name = "json-depth-checker"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "criterion",
 "serde_json",
@@ -3366,7 +3364,7 @@ checksum = "490cc448043f947bae3cbee9c203358d62dbee0db12107a74be5c30ccfd09771"

 [[package]]
 name = "meili-snap"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "insta",
 "md5",
@@ -3375,7 +3373,7 @@ dependencies = [

 [[package]]
 name = "meilisearch"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "actix-cors",
 "actix-http",
@@ -3415,7 +3413,6 @@ dependencies = [
 "meilisearch-types",
 "mimalloc",
 "mime",
- "mopa-maintained",
 "num_cpus",
 "obkv",
 "once_cell",
@@ -3465,7 +3462,7 @@ dependencies = [

 [[package]]
 name = "meilisearch-auth"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "base64 0.22.1",
 "enum-iterator",
@@ -3484,7 +3481,7 @@ dependencies = [

 [[package]]
 name = "meilisearch-types"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "actix-web",
 "anyhow",
@@ -3514,7 +3511,7 @@ dependencies = [

 [[package]]
 name = "meilitool"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "anyhow",
 "clap",
@@ -3545,7 +3542,7 @@ dependencies = [

 [[package]]
 name = "milli"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "arroy",
 "big_s",
@@ -3569,6 +3566,7 @@ dependencies = [
 "fxhash",
 "geoutils",
 "grenad",
+ "hashbrown 0.14.5",
 "heed",
 "hf-hub",
 "indexmap",
@@ -3682,12 +3680,6 @@ dependencies = [
 "syn 2.0.60",
 ]

-[[package]]
-name = "mopa-maintained"
-version = "0.2.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "79b7f3e22167862cc7c95b21a6f326c22e4bf40da59cbf000b368a310173ba11"
-
 [[package]]
 name = "mutually_exclusive_features"
 version = "0.0.3"
@@ -3844,9 +3836,8 @@ dependencies = [

 [[package]]
 name = "obkv"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a2e27bcfe835a379d32352112f6b8dbae2d99d16a5fff42abe6e5ba5386c1e5a"
+version = "0.3.0"
+source = "git+https://github.com/kerollmops/obkv?branch=unsized-kvreader#ce535874008ecac554f02e0c670e6caf62134d6b"

 [[package]]
 name = "once_cell"
@@ -3991,7 +3982,7 @@ checksum = "e3148f5046208a5d56bcfc03053e3ca6334e51da8dfb19b6cdc8b306fae3283e"

 [[package]]
 name = "permissive-json-pointer"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "big_s",
 "serde_json",
@@ -4322,7 +4313,7 @@ dependencies = [
 "pin-project-lite",
 "quinn-proto",
 "quinn-udp",
- "rustc-hash 1.1.0",
+ "rustc-hash",
 "rustls",
 "thiserror",
 "tokio",
@@ -4331,14 +4322,14 @@ dependencies = [

 [[package]]
 name = "quinn-proto"
-version = "0.11.8"
+version = "0.11.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fadfaed2cd7f389d0161bb73eeb07b7b78f8691047a6f3e73caaeae55310a4a6"
+checksum = "ddf517c03a109db8100448a4be38d498df8a210a99fe0e1b9eaf39e78c640efe"
 dependencies = [
 "bytes",
 "rand",
 "ring",
- "rustc-hash 2.0.0",
+ "rustc-hash",
 "rustls",
 "slab",
 "thiserror",
@@ -4589,8 +4580,9 @@ dependencies = [

 [[package]]
 name = "rhai"
-version = "1.20.0"
-source = "git+https://github.com/rhaiscript/rhai?rev=ef3df63121d27aacd838f366f2b83fd65f20a1e4#ef3df63121d27aacd838f366f2b83fd65f20a1e4"
+version = "1.19.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "61797318be89b1a268a018a92a7657096d83f3ecb31418b9e9c16dcbb043b702"
 dependencies = [
 "ahash 0.8.11",
 "bitflags 2.6.0",
@@ -4607,7 +4599,8 @@ dependencies = [
 [[package]]
 name = "rhai_codegen"
 version = "2.2.0"
-source = "git+https://github.com/rhaiscript/rhai?rev=ef3df63121d27aacd838f366f2b83fd65f20a1e4#ef3df63121d27aacd838f366f2b83fd65f20a1e4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a5a11a05ee1ce44058fa3d5961d05194fdbe3ad6b40f904af764d81b86450e6b"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -4661,8 +4654,7 @@ dependencies = [
 [[package]]
 name = "roaring"
 version = "0.10.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8f4b84ba6e838ceb47b41de5194a60244fac43d9fe03b71dbe8c5a201081d6d1"
+source = "git+https://github.com/RoaringBitmap/roaring-rs?branch=clone-iter-slice#6bba84b1a47da1d6e52d5c4dc0ce8593ae4646a5"
 dependencies = [
 "bytemuck",
 "byteorder",
@@ -4709,12 +4701,6 @@ version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "08d43f7aa6b08d49f382cde6a7982047c3426db949b1424bc4b7ec9ae12c6ce2"

-[[package]]
-name = "rustc-hash"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "583034fd73374156e66797ed8e5b0d5690409c9226b22d87cb7f19821c05d152"
-
 [[package]]
 name = "rustc_version"
 version = "0.4.0"
@@ -4853,9 +4839,9 @@ checksum = "a3f0bf26fd526d2a95683cd0f87bf103b8539e2ca1ef48ce002d67aad59aa0b4"

 [[package]]
 name = "serde"
-version = "1.0.209"
+version = "1.0.210"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "99fce0ffe7310761ca6bf9faf5115afbc19688edd00171d81b1bb1b116c63e09"
+checksum = "c8e3592472072e6e22e0a54d5904d9febf8508f65fb8552499a1abc7d1078c3a"
 dependencies = [
 "serde_derive",
 ]
@@ -4871,9 +4857,9 @@ dependencies = [

 [[package]]
 name = "serde_derive"
-version = "1.0.209"
+version = "1.0.210"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a5831b979fd7b5439637af1752d535ff49f4860c0f341d1baeb6faf0f4242170"
+checksum = "243902eda00fad750862fc144cea25caca5e20d615af0a81bee94ca738f1df1f"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -5366,7 +5352,7 @@ dependencies = [
 "fancy-regex 0.12.0",
 "lazy_static",
 "parking_lot",
- "rustc-hash 1.1.0",
+ "rustc-hash",
 ]

 [[package]]
@@ -6054,7 +6040,7 @@ version = "0.16.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "471d1c1645d361eb782a1650b1786a8fb58dd625e681a04c09f5ff7c8764a7b0"
 dependencies = [
- "hashbrown 0.14.3",
+ "hashbrown 0.14.5",
 "once_cell",
 ]

@@ -6380,7 +6366,7 @@ dependencies = [

 [[package]]
 name = "xtask"
-version = "1.11.3"
+version = "1.11.0"
 dependencies = [
 "anyhow",
 "build-info",
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -22,7 +22,7 @@ members = [
 ]

 [workspace.package]
-version = "1.11.3"
+version = "1.11.0"
 authors = [
    "Quentin de Quelen <quentin@dequelen.me>",
    "Clément Renault <clement@meilisearch.com>",
@@ -44,23 +44,5 @@ opt-level = 3
 [profile.dev.package.roaring]
 opt-level = 3

-[profile.dev.package.lindera-ipadic-builder]
-opt-level = 3
-[profile.dev.package.encoding]
-opt-level = 3
-[profile.dev.package.yada]
-opt-level = 3
-
-[profile.release.package.lindera-ipadic-builder]
-opt-level = 3
-[profile.release.package.encoding]
-opt-level = 3
-[profile.release.package.yada]
-opt-level = 3
-
-[profile.bench.package.lindera-ipadic-builder]
-opt-level = 3
-[profile.bench.package.encoding]
-opt-level = 3
-[profile.bench.package.yada]
-opt-level = 3
+[patch.crates-io]
+roaring = { git = "https://github.com/RoaringBitmap/roaring-rs", branch = "clone-iter-slice" }
--- a/index-scheduler/Cargo.toml
+++ b/index-scheduler/Cargo.toml
@@ -29,6 +29,7 @@ serde_json = { version = "1.0.120", features = ["preserve_order"] }
 synchronoise = "1.0.1"
 tempfile = "3.10.1"
 thiserror = "1.0.61"
+memmap2 = "0.9.4"
 time = { version = "0.3.36", features = [
    "serde-well-known",
    "formatting",
@@ -40,7 +41,7 @@ ureq = "2.10.0"
 uuid = { version = "1.10.0", features = ["serde", "v4"] }

 [dev-dependencies]
-arroy = "0.5.0"
+arroy = { git = "https://github.com/meilisearch/arroy/", rev = "2386594dfb009ce08821a925ccc89fb8e30bf73d" }
 big_s = "1.0.2"
 crossbeam = "0.8.4"
 insta = { version = "1.39.0", features = ["json", "redactions"] }
--- a/index-scheduler/src/batch.rs
+++ b/index-scheduler/src/batch.rs
@@ -28,6 +28,9 @@ use meilisearch_types::error::Code;
 use meilisearch_types::heed::{RoTxn, RwTxn};
 use meilisearch_types::milli::documents::{obkv_to_object, DocumentsBatchReader};
 use meilisearch_types::milli::heed::CompactionOption;
+use meilisearch_types::milli::update::new::indexer::{
+    self, retrieve_or_guess_primary_key, DocumentChanges,
+};
 use meilisearch_types::milli::update::{
    IndexDocumentsConfig, IndexDocumentsMethod, IndexerConfig, Settings as MilliSettings,
 };
@@ -875,10 +878,8 @@ impl IndexScheduler {
                            while let Some(doc) =
                                cursor.next_document().map_err(milli::Error::from)?
                            {
-                                dump_content_file.push_document(&obkv_to_object(
-                                    &doc,
-                                    &documents_batch_index,
-                                )?)?;
+                                dump_content_file
+                                    .push_document(&obkv_to_object(doc, &documents_batch_index)?)?;
                            }
                            dump_content_file.flush()?;
                        }
@@ -1252,58 +1253,52 @@ impl IndexScheduler {
                let must_stop_processing = self.must_stop_processing.clone();
                let indexer_config = self.index_mapper.indexer_config();

-                if let Some(primary_key) = primary_key {
-                    match index.primary_key(index_wtxn)? {
-                        // if a primary key was set AND had already been defined in the index
-                        // but to a different value, we can make the whole batch fail.
-                        Some(pk) => {
-                            if primary_key != pk {
-                                return Err(milli::Error::from(
-                                    milli::UserError::PrimaryKeyCannotBeChanged(pk.to_string()),
-                                )
-                                .into());
-                            }
-                        }
-                        // if the primary key was set and there was no primary key set for this index
-                        // we set it to the received value before starting the indexing process.
-                        None => {
-                            let mut builder =
-                                milli::update::Settings::new(index_wtxn, index, indexer_config);
-                            builder.set_primary_key(primary_key);
-                            builder.execute(
-                                |indexing_step| tracing::debug!(update = ?indexing_step),
-                                || must_stop_processing.clone().get(),
-                            )?;
-                            primary_key_has_been_set = true;
+                /// TODO manage errors correctly
+                let rtxn = index.read_txn()?;
+                let first_addition_uuid = operations
+                    .iter()
+                    .find_map(|op| match op {
+                        DocumentOperation::Add(content_uuid) => Some(content_uuid),
+                        _ => None,
+                    })
+                    .unwrap();
+
+                let mut content_files = Vec::new();
+                for operation in &operations {
+                    if let DocumentOperation::Add(content_uuid) = operation {
+                        let content_file = self.file_store.get_update(*content_uuid)?;
+                        let mmap = unsafe { memmap2::Mmap::map(&content_file)? };
+                        if !mmap.is_empty() {
+                            content_files.push(mmap);
                        }
                    }
                }

-                let config = IndexDocumentsConfig { update_method: method, ..Default::default() };
+                let mut fields_ids_map = index.fields_ids_map(&rtxn)?;
+                let first_document = match content_files.first() {
+                    Some(mmap) => {
+                        let mut iter = serde_json::Deserializer::from_slice(mmap).into_iter();
+                        iter.next().transpose().map_err(|e| e.into()).map_err(Error::IoError)?
+                    }
+                    None => None,
+                };

-                let embedder_configs = index.embedding_configs(index_wtxn)?;
-                // TODO: consider Arc'ing the map too (we only need read access + we'll be cloning it multiple times, so really makes sense)
-                let embedders = self.embedders(embedder_configs)?;
-
-                let mut builder = milli::update::IndexDocuments::new(
-                    index_wtxn,
+                let primary_key = retrieve_or_guess_primary_key(
+                    &rtxn,
                    index,
-                    indexer_config,
-                    config,
-                    |indexing_step| tracing::trace!(?indexing_step, "Update"),
-                    || must_stop_processing.get(),
-                )?;
+                    &mut fields_ids_map,
+                    first_document.as_ref(),
+                )?
+                .unwrap();

+                let mut content_files_iter = content_files.iter();
+                let mut indexer = indexer::DocumentOperation::new(method);
                for (operation, task) in operations.into_iter().zip(tasks.iter_mut()) {
                    match operation {
-                        DocumentOperation::Add(content_uuid) => {
-                            let content_file = self.file_store.get_update(content_uuid)?;
-                            let reader = DocumentsBatchReader::from_reader(content_file)
-                                .map_err(milli::Error::from)?;
-                            let (new_builder, user_result) = builder.add_documents(reader)?;
-                            builder = new_builder;
-
-                            builder = builder.with_embedders(embedders.clone());
+                        DocumentOperation::Add(_content_uuid) => {
+                            let mmap = content_files_iter.next().unwrap();
+                            let stats = indexer.add_documents(mmap)?;
+                            // builder = builder.with_embedders(embedders.clone());

                            let received_documents =
                                if let Some(Details::DocumentAdditionOrUpdate {
@@ -1317,30 +1312,17 @@ impl IndexScheduler {
                                    unreachable!();
                                };

-                            match user_result {
-                                Ok(count) => {
-                                    task.status = Status::Succeeded;
-                                    task.details = Some(Details::DocumentAdditionOrUpdate {
-                                        received_documents,
-                                        indexed_documents: Some(count),
-                                    })
-                                }
-                                Err(e) => {
-                                    task.status = Status::Failed;
-                                    task.details = Some(Details::DocumentAdditionOrUpdate {
-                                        received_documents,
-                                        indexed_documents: Some(0),
-                                    });
-                                    task.error = Some(milli::Error::from(e).into());
-                                }
-                            }
+                            task.status = Status::Succeeded;
+                            task.details = Some(Details::DocumentAdditionOrUpdate {
+                                received_documents,
+                                indexed_documents: Some(stats.document_count as u64),
+                            })
                        }
                        DocumentOperation::Delete(document_ids) => {
-                            let (new_builder, user_result) =
-                                builder.remove_documents(document_ids)?;
-                            builder = new_builder;
+                            let count = document_ids.len();
+                            indexer.delete_documents(document_ids);
                            // Uses Invariant: remove documents actually always returns Ok for the inner result
-                            let count = user_result.unwrap();
+                            // let count = user_result.unwrap();
                            let provided_ids =
                                if let Some(Details::DocumentDeletion { provided_ids, .. }) =
                                    task.details
@@ -1354,26 +1336,35 @@ impl IndexScheduler {
                            task.status = Status::Succeeded;
                            task.details = Some(Details::DocumentDeletion {
                                provided_ids,
-                                deleted_documents: Some(count),
+                                deleted_documents: Some(count as u64),
                            });
                        }
                    }
                }

                if !tasks.iter().all(|res| res.error.is_some()) {
-                    let addition = builder.execute()?;
-                    tracing::info!(indexing_result = ?addition, processed_in = ?started_processing_at.elapsed(), "document indexing done");
-                } else if primary_key_has_been_set {
-                    // Everything failed but we've set a primary key.
-                    // We need to remove it.
-                    let mut builder =
-                        milli::update::Settings::new(index_wtxn, index, indexer_config);
-                    builder.reset_primary_key();
-                    builder.execute(
-                        |indexing_step| tracing::trace!(update = ?indexing_step),
-                        || must_stop_processing.clone().get(),
-                    )?;
+                    /// TODO create a pool if needed
+                    // let pool = indexer_config.thread_pool.unwrap();
+                    let pool = rayon::ThreadPoolBuilder::new().build().unwrap();
+
+                    let param = (index, &rtxn, &primary_key);
+                    let document_changes = indexer.document_changes(&mut fields_ids_map, param)?;
+                    /// TODO pass/write the FieldsIdsMap
+                    indexer::index(index_wtxn, index, fields_ids_map, &pool, document_changes)?;
+
+                    // tracing::info!(indexing_result = ?addition, processed_in = ?started_processing_at.elapsed(), "document indexing done");
                }
+                // else if primary_key_has_been_set {
+                //     // Everything failed but we've set a primary key.
+                //     // We need to remove it.
+                //     let mut builder =
+                //         milli::update::Settings::new(index_wtxn, index, indexer_config);
+                //     builder.reset_primary_key();
+                //     builder.execute(
+                //         |indexing_step| tracing::trace!(update = ?indexing_step),
+                //         || must_stop_processing.clone().get(),
+                //     )?;
+                // }

                Ok(tasks)
            }
--- a/index-scheduler/src/lib.rs
+++ b/index-scheduler/src/lib.rs
@@ -1263,7 +1263,7 @@ impl IndexScheduler {
                    #[cfg(test)]
                    self.maybe_fail(tests::FailureLocation::UpdatingTaskAfterProcessBatchFailure)?;

-                    tracing::error!("Batch failed {}", error);
+                    tracing::info!("Batch failed {}", error);

                    self.update_task(&mut wtxn, &task)
                        .map_err(|e| Error::TaskDatabaseUpdate(Box::new(e)))?;
@@ -5201,10 +5201,9 @@ mod tests {

            let configs = index_scheduler.embedders(configs).unwrap();
            let (hf_embedder, _, _) = configs.get(&simple_hf_name).unwrap();
-            let beagle_embed =
-                hf_embedder.embed_one(S("Intel the beagle best doggo"), None).unwrap();
-            let lab_embed = hf_embedder.embed_one(S("Max the lab best doggo"), None).unwrap();
-            let patou_embed = hf_embedder.embed_one(S("kefir the patou best doggo"), None).unwrap();
+            let beagle_embed = hf_embedder.embed_one(S("Intel the beagle best doggo")).unwrap();
+            let lab_embed = hf_embedder.embed_one(S("Max the lab best doggo")).unwrap();
+            let patou_embed = hf_embedder.embed_one(S("kefir the patou best doggo")).unwrap();
            (fakerest_name, simple_hf_name, beagle_embed, lab_embed, patou_embed)
        };

--- a/meilisearch-types/Cargo.toml
+++ b/meilisearch-types/Cargo.toml
@@ -66,8 +66,5 @@ khmer = ["milli/khmer"]
 vietnamese = ["milli/vietnamese"]
 # force swedish character recomposition
 swedish-recomposition = ["milli/swedish-recomposition"]
-# allow german tokenization
+# force german character recomposition
 german = ["milli/german"]
-# allow turkish normalization
-turkish = ["milli/turkish"]
-
--- a/meilisearch-types/src/document_formats.rs
+++ b/meilisearch-types/src/document_formats.rs
@@ -1,20 +1,22 @@
 use std::fmt::{self, Debug, Display};
 use std::fs::File;
-use std::io::{self, BufWriter, Write};
+use std::io::{self, BufWriter};
 use std::marker::PhantomData;

-use memmap2::MmapOptions;
-use milli::documents::{DocumentsBatchBuilder, Error};
+use memmap2::Mmap;
+use milli::documents::Error;
+use milli::update::new::TopLevelMap;
 use milli::Object;
 use serde::de::{SeqAccess, Visitor};
 use serde::{Deserialize, Deserializer};
 use serde_json::error::Category;
+use serde_json::{to_writer, Map, Value};

 use crate::error::{Code, ErrorCode};

 type Result<T> = std::result::Result<T, DocumentFormatError>;

-#[derive(Debug)]
+#[derive(Debug, Clone, Copy)]
 pub enum PayloadType {
    Ndjson,
    Json,
@@ -88,6 +90,26 @@ impl From<(PayloadType, Error)> for DocumentFormatError {
    }
 }

+impl From<(PayloadType, serde_json::Error)> for DocumentFormatError {
+    fn from((ty, error): (PayloadType, serde_json::Error)) -> Self {
+        if error.classify() == Category::Data {
+            Self::Io(error.into())
+        } else {
+            Self::MalformedPayload(Error::Json(error), ty)
+        }
+    }
+}
+
+impl From<(PayloadType, csv::Error)> for DocumentFormatError {
+    fn from((ty, error): (PayloadType, csv::Error)) -> Self {
+        if error.is_io_error() {
+            Self::Io(error.into())
+        } else {
+            Self::MalformedPayload(Error::Csv(error), ty)
+        }
+    }
+}
+
 impl From<io::Error> for DocumentFormatError {
    fn from(error: io::Error) -> Self {
        Self::Io(error)
@@ -103,67 +125,140 @@ impl ErrorCode for DocumentFormatError {
    }
 }

-/// Reads CSV from input and write an obkv batch to writer.
-pub fn read_csv(file: &File, writer: impl Write, delimiter: u8) -> Result<u64> {
-    let mut builder = DocumentsBatchBuilder::new(BufWriter::new(writer));
-    let mmap = unsafe { MmapOptions::new().map(file)? };
-    let csv = csv::ReaderBuilder::new().delimiter(delimiter).from_reader(mmap.as_ref());
-    builder.append_csv(csv).map_err(|e| (PayloadType::Csv { delimiter }, e))?;
-
-    let count = builder.documents_count();
-    let _ = builder.into_inner().map_err(DocumentFormatError::Io)?;
-
-    Ok(count as u64)
+// TODO remove that from the place I've borrowed it
+#[derive(Debug)]
+enum AllowedType {
+    String,
+    Boolean,
+    Number,
 }

-/// Reads JSON from temporary file and write an obkv batch to writer.
-pub fn read_json(file: &File, writer: impl Write) -> Result<u64> {
-    let mut builder = DocumentsBatchBuilder::new(BufWriter::new(writer));
-    let mmap = unsafe { MmapOptions::new().map(file)? };
-    let mut deserializer = serde_json::Deserializer::from_slice(&mmap);
+fn parse_csv_header(header: &str) -> (&str, AllowedType) {
+    // if there are several separators we only split on the last one.
+    match header.rsplit_once(':') {
+        Some((field_name, field_type)) => match field_type {
+            "string" => (field_name, AllowedType::String),
+            "boolean" => (field_name, AllowedType::Boolean),
+            "number" => (field_name, AllowedType::Number),
+            // if the pattern isn't recognized, we keep the whole field.
+            _otherwise => (header, AllowedType::String),
+        },
+        None => (header, AllowedType::String),
+    }
+}

-    match array_each(&mut deserializer, |obj| builder.append_json_object(&obj)) {
+/// Reads CSV from file and write it in NDJSON in a file checking it along the way.
+pub fn read_csv(input: &File, output: impl io::Write, delimiter: u8) -> Result<u64> {
+    let ptype = PayloadType::Csv { delimiter };
+    let mut output = BufWriter::new(output);
+    let mut reader = csv::ReaderBuilder::new().delimiter(delimiter).from_reader(input);
+
+    let headers = reader.headers().map_err(|e| DocumentFormatError::from((ptype, e)))?.clone();
+    let typed_fields: Vec<_> = headers.iter().map(parse_csv_header).collect();
+    let mut object: Map<_, _> =
+        typed_fields.iter().map(|(k, _)| (k.to_string(), Value::Null)).collect();
+
+    let mut line = 0;
+    let mut record = csv::StringRecord::new();
+    while reader.read_record(&mut record).map_err(|e| DocumentFormatError::from((ptype, e)))? {
+        // We increment here and not at the end of the loop
+        // to take the header offset into account.
+        line += 1;
+
+        // Reset the document values
+        object.iter_mut().for_each(|(_, v)| *v = Value::Null);
+
+        for (i, (name, atype)) in typed_fields.iter().enumerate() {
+            let value = &record[i];
+            let trimmed_value = value.trim();
+            let value = match atype {
+                AllowedType::Number if trimmed_value.is_empty() => Value::Null,
+                AllowedType::Number => match trimmed_value.parse::<i64>() {
+                    Ok(integer) => Value::from(integer),
+                    Err(_) => match trimmed_value.parse::<f64>() {
+                        Ok(float) => Value::from(float),
+                        Err(error) => {
+                            return Err(DocumentFormatError::MalformedPayload(
+                                Error::ParseFloat { error, line, value: value.to_string() },
+                                ptype,
+                            ))
+                        }
+                    },
+                },
+                AllowedType::Boolean if trimmed_value.is_empty() => Value::Null,
+                AllowedType::Boolean => match trimmed_value.parse::<bool>() {
+                    Ok(bool) => Value::from(bool),
+                    Err(error) => {
+                        return Err(DocumentFormatError::MalformedPayload(
+                            Error::ParseBool { error, line, value: value.to_string() },
+                            ptype,
+                        ))
+                    }
+                },
+                AllowedType::String if value.is_empty() => Value::Null,
+                AllowedType::String => Value::from(value),
+            };
+
+            *object.get_mut(*name).expect("encountered an unknown field") = value;
+        }
+
+        to_writer(&mut output, &object).map_err(|e| DocumentFormatError::from((ptype, e)))?;
+    }
+
+    Ok(line as u64)
+}
+
+/// Reads JSON from file and write it in NDJSON in a file checking it along the way.
+pub fn read_json(input: &File, output: impl io::Write) -> Result<u64> {
+    // We memory map to be able to deserailize into a TopLevelMap<'pl> that
+    // does not allocate when possible and only materialize the first/top level.
+    let input = unsafe { Mmap::map(input).map_err(DocumentFormatError::Io)? };
+
+    let mut out = BufWriter::new(output);
+    let mut deserializer = serde_json::Deserializer::from_slice(&input);
+    let count = match array_each(&mut deserializer, |obj: TopLevelMap| to_writer(&mut out, &obj)) {
        // The json data has been deserialized and does not need to be processed again.
        // The data has been transferred to the writer during the deserialization process.
-        Ok(Ok(_)) => (),
-        Ok(Err(e)) => return Err(DocumentFormatError::Io(e)),
+        Ok(Ok(count)) => count,
+        Ok(Err(e)) => return Err(DocumentFormatError::from((PayloadType::Json, e))),
        Err(e) => {
            // Attempt to deserialize a single json string when the cause of the exception is not Category.data
            // Other types of deserialisation exceptions are returned directly to the front-end
-            if e.classify() != serde_json::error::Category::Data {
-                return Err(DocumentFormatError::MalformedPayload(
-                    Error::Json(e),
-                    PayloadType::Json,
-                ));
+            if e.classify() != Category::Data {
+                return Err(DocumentFormatError::from((PayloadType::Json, e)));
            }

-            let content: Object = serde_json::from_slice(&mmap)
+            let content: Object = serde_json::from_slice(&input)
                .map_err(Error::Json)
                .map_err(|e| (PayloadType::Json, e))?;
-            builder.append_json_object(&content).map_err(DocumentFormatError::Io)?;
+            to_writer(&mut out, &content)
+                .map(|_| 1)
+                .map_err(|e| DocumentFormatError::from((PayloadType::Json, e)))?
        }
+    };
+
+    match out.into_inner() {
+        Ok(_) => Ok(count),
+        Err(ie) => Err(DocumentFormatError::Io(ie.into_error())),
    }
-
-    let count = builder.documents_count();
-    let _ = builder.into_inner().map_err(DocumentFormatError::Io)?;
-
-    Ok(count as u64)
 }

-/// Reads JSON from temporary file  and write an obkv batch to writer.
-pub fn read_ndjson(file: &File, writer: impl Write) -> Result<u64> {
-    let mut builder = DocumentsBatchBuilder::new(BufWriter::new(writer));
-    let mmap = unsafe { MmapOptions::new().map(file)? };
+/// Reads NDJSON from file and write it in NDJSON in a file checking it along the way.
+pub fn read_ndjson(input: &File, output: impl io::Write) -> Result<u64> {
+    // We memory map to be able to deserailize into a TopLevelMap<'pl> that
+    // does not allocate when possible and only materialize the first/top level.
+    let input = unsafe { Mmap::map(input).map_err(DocumentFormatError::Io)? };
+    let mut output = BufWriter::new(output);

-    for result in serde_json::Deserializer::from_slice(&mmap).into_iter() {
-        let object = result.map_err(Error::Json).map_err(|e| (PayloadType::Ndjson, e))?;
-        builder.append_json_object(&object).map_err(Into::into).map_err(DocumentFormatError::Io)?;
+    let mut count = 0;
+    for result in serde_json::Deserializer::from_slice(&input).into_iter() {
+        count += 1;
+        result
+            .and_then(|map: TopLevelMap| to_writer(&mut output, &map))
+            .map_err(|e| DocumentFormatError::from((PayloadType::Ndjson, e)))?;
    }

-    let count = builder.documents_count();
-    let _ = builder.into_inner().map_err(Into::into).map_err(DocumentFormatError::Io)?;
-
-    Ok(count as u64)
+    Ok(count)
 }

 /// The actual handling of the deserialization process in serde
@@ -172,20 +267,23 @@ pub fn read_ndjson(file: &File, writer: impl Write) -> Result<u64> {
 /// ## References
 /// <https://serde.rs/stream-array.html>
 /// <https://github.com/serde-rs/json/issues/160>
-fn array_each<'de, D, T, F>(deserializer: D, f: F) -> std::result::Result<io::Result<u64>, D::Error>
+fn array_each<'de, D, T, F>(
+    deserializer: D,
+    f: F,
+) -> std::result::Result<serde_json::Result<u64>, D::Error>
 where
    D: Deserializer<'de>,
    T: Deserialize<'de>,
-    F: FnMut(T) -> io::Result<()>,
+    F: FnMut(T) -> serde_json::Result<()>,
 {
    struct SeqVisitor<T, F>(F, PhantomData<T>);

    impl<'de, T, F> Visitor<'de> for SeqVisitor<T, F>
    where
        T: Deserialize<'de>,
-        F: FnMut(T) -> io::Result<()>,
+        F: FnMut(T) -> serde_json::Result<()>,
    {
-        type Value = io::Result<u64>;
+        type Value = serde_json::Result<u64>;

        fn expecting(&self, formatter: &mut fmt::Formatter) -> fmt::Result {
            formatter.write_str("a nonempty sequence")
@@ -194,7 +292,7 @@ where
        fn visit_seq<A>(
            mut self,
            mut seq: A,
-        ) -> std::result::Result<io::Result<u64>, <A as SeqAccess<'de>>::Error>
+        ) -> std::result::Result<serde_json::Result<u64>, <A as SeqAccess<'de>>::Error>
        where
            A: SeqAccess<'de>,
        {
@@ -203,7 +301,7 @@ where
                match self.0(value) {
                    Ok(()) => max += 1,
                    Err(e) => return Ok(Err(e)),
-                };
+                }
            }
            Ok(Ok(max))
        }
--- a/meilisearch/Cargo.toml
+++ b/meilisearch/Cargo.toml
@@ -57,7 +57,7 @@ meilisearch-types = { path = "../meilisearch-types" }
 mimalloc = { version = "0.1.43", default-features = false }
 mime = "0.3.17"
 num_cpus = "1.16.0"
-obkv = "0.2.2"
+obkv = { git = "https://github.com/kerollmops/obkv", branch = "unsized-kvreader" }
 once_cell = "1.19.0"
 ordered-float = "4.2.1"
 parking_lot = "0.12.3"
@@ -75,7 +75,7 @@ reqwest = { version = "0.12.5", features = [
 rustls = { version = "0.23.11", features = ["ring"], default-features = false }
 rustls-pki-types = { version = "1.7.0", features = ["alloc"] }
 rustls-pemfile = "2.1.2"
-segment = { version = "0.2.4" }
+segment = { version = "0.2.4", optional = true }
 serde = { version = "1.0.204", features = ["derive"] }
 serde_json = { version = "1.0.120", features = ["preserve_order"] }
 sha2 = "0.10.8"
@@ -104,7 +104,6 @@ tracing-trace = { version = "0.1.0", path = "../tracing-trace" }
 tracing-actix-web = "0.7.11"
 build-info = { version = "1.7.0", path = "../build-info" }
 roaring = "0.10.2"
-mopa-maintained = "0.2.3"

 [dev-dependencies]
 actix-rt = "2.10.0"
@@ -132,7 +131,8 @@ tempfile = { version = "3.10.1", optional = true }
 zip = { version = "2.1.3", optional = true }

 [features]
-default = ["meilisearch-types/all-tokenizations", "mini-dashboard"]
+default = ["analytics", "meilisearch-types/all-tokenizations", "mini-dashboard"]
+analytics = ["segment"]
 mini-dashboard = [
    "static-files",
    "anyhow",
@@ -154,8 +154,7 @@ khmer = ["meilisearch-types/khmer"]
 vietnamese = ["meilisearch-types/vietnamese"]
 swedish-recomposition = ["meilisearch-types/swedish-recomposition"]
 german = ["meilisearch-types/german"]
-turkish = ["meilisearch-types/turkish"]

 [package.metadata.mini-dashboard]
-assets-url = "https://github.com/meilisearch/mini-dashboard/releases/download/v0.2.15/build.zip"
-sha1 = "d057600b4a839a2e0c0be7a372cd1b2683f3ca7e"
+assets-url = "https://github.com/meilisearch/mini-dashboard/releases/download/v0.2.14/build.zip"
+sha1 = "592d1b5a3459d621d0aae1dded8fe3154f5c38fe"
--- a/meilisearch/src/analytics/mock_analytics.rs
+++ b/meilisearch/src/analytics/mock_analytics.rs
@@ -0,0 +1,109 @@
+use std::any::Any;
+use std::sync::Arc;
+
+use actix_web::HttpRequest;
+use meilisearch_types::InstanceUid;
+use serde_json::Value;
+
+use super::{find_user_id, Analytics, DocumentDeletionKind, DocumentFetchKind};
+use crate::routes::indexes::documents::{DocumentEditionByFunction, UpdateDocumentsQuery};
+use crate::Opt;
+
+pub struct MockAnalytics {
+    instance_uid: Option<InstanceUid>,
+}
+
+#[derive(Default)]
+pub struct SearchAggregator;
+
+#[allow(dead_code)]
+impl SearchAggregator {
+    pub fn from_query(_: &dyn Any, _: &dyn Any) -> Self {
+        Self
+    }
+
+    pub fn succeed(&mut self, _: &dyn Any) {}
+}
+
+#[derive(Default)]
+pub struct SimilarAggregator;
+
+#[allow(dead_code)]
+impl SimilarAggregator {
+    pub fn from_query(_: &dyn Any, _: &dyn Any) -> Self {
+        Self
+    }
+
+    pub fn succeed(&mut self, _: &dyn Any) {}
+}
+
+#[derive(Default)]
+pub struct MultiSearchAggregator;
+
+#[allow(dead_code)]
+impl MultiSearchAggregator {
+    pub fn from_federated_search(_: &dyn Any, _: &dyn Any) -> Self {
+        Self
+    }
+
+    pub fn succeed(&mut self) {}
+}
+
+#[derive(Default)]
+pub struct FacetSearchAggregator;
+
+#[allow(dead_code)]
+impl FacetSearchAggregator {
+    pub fn from_query(_: &dyn Any, _: &dyn Any) -> Self {
+        Self
+    }
+
+    pub fn succeed(&mut self, _: &dyn Any) {}
+}
+
+impl MockAnalytics {
+    #[allow(clippy::new_ret_no_self)]
+    pub fn new(opt: &Opt) -> Arc<dyn Analytics> {
+        let instance_uid = find_user_id(&opt.db_path);
+        Arc::new(Self { instance_uid })
+    }
+}
+
+impl Analytics for MockAnalytics {
+    fn instance_uid(&self) -> Option<&meilisearch_types::InstanceUid> {
+        self.instance_uid.as_ref()
+    }
+
+    // These methods are noop and should be optimized out
+    fn publish(&self, _event_name: String, _send: Value, _request: Option<&HttpRequest>) {}
+    fn get_search(&self, _aggregate: super::SearchAggregator) {}
+    fn post_search(&self, _aggregate: super::SearchAggregator) {}
+    fn get_similar(&self, _aggregate: super::SimilarAggregator) {}
+    fn post_similar(&self, _aggregate: super::SimilarAggregator) {}
+    fn post_multi_search(&self, _aggregate: super::MultiSearchAggregator) {}
+    fn post_facet_search(&self, _aggregate: super::FacetSearchAggregator) {}
+    fn add_documents(
+        &self,
+        _documents_query: &UpdateDocumentsQuery,
+        _index_creation: bool,
+        _request: &HttpRequest,
+    ) {
+    }
+    fn delete_documents(&self, _kind: DocumentDeletionKind, _request: &HttpRequest) {}
+    fn update_documents(
+        &self,
+        _documents_query: &UpdateDocumentsQuery,
+        _index_creation: bool,
+        _request: &HttpRequest,
+    ) {
+    }
+    fn update_documents_by_function(
+        &self,
+        _documents_query: &DocumentEditionByFunction,
+        _index_creation: bool,
+        _request: &HttpRequest,
+    ) {
+    }
+    fn get_fetch_documents(&self, _documents_query: &DocumentFetchKind, _request: &HttpRequest) {}
+    fn post_fetch_documents(&self, _documents_query: &DocumentFetchKind, _request: &HttpRequest) {}
+}
--- a/meilisearch/src/analytics/mod.rs
+++ b/meilisearch/src/analytics/mod.rs
@@ -1,45 +1,44 @@
-pub mod segment_analytics;
+mod mock_analytics;
+#[cfg(feature = "analytics")]
+mod segment_analytics;

 use std::fs;
 use std::path::{Path, PathBuf};
 use std::str::FromStr;
-use std::sync::Arc;

 use actix_web::HttpRequest;
-use index_scheduler::IndexScheduler;
-use meilisearch_auth::AuthController;
 use meilisearch_types::InstanceUid;
-use mopa::mopafy;
+pub use mock_analytics::MockAnalytics;
 use once_cell::sync::Lazy;
 use platform_dirs::AppDirs;
+use serde_json::Value;
+
+use crate::routes::indexes::documents::{DocumentEditionByFunction, UpdateDocumentsQuery};
+
+// if the analytics feature is disabled
+// the `SegmentAnalytics` point to the mock instead of the real analytics
+#[cfg(not(feature = "analytics"))]
+pub type SegmentAnalytics = mock_analytics::MockAnalytics;
+#[cfg(not(feature = "analytics"))]
+pub type SearchAggregator = mock_analytics::SearchAggregator;
+#[cfg(not(feature = "analytics"))]
+pub type SimilarAggregator = mock_analytics::SimilarAggregator;
+#[cfg(not(feature = "analytics"))]
+pub type MultiSearchAggregator = mock_analytics::MultiSearchAggregator;
+#[cfg(not(feature = "analytics"))]
+pub type FacetSearchAggregator = mock_analytics::FacetSearchAggregator;

 // if the feature analytics is enabled we use the real analytics
+#[cfg(feature = "analytics")]
 pub type SegmentAnalytics = segment_analytics::SegmentAnalytics;
-
-use crate::Opt;
-
-/// A macro used to quickly define events that don't aggregate or send anything besides an empty event with its name.
-#[macro_export]
-macro_rules! empty_analytics {
-    ($struct_name:ident, $event_name:literal) => {
-        #[derive(Default)]
-        struct $struct_name {}
-
-        impl $crate::analytics::Aggregate for $struct_name {
-            fn event_name(&self) -> &'static str {
-                $event_name
-            }
-
-            fn aggregate(self: Box<Self>, _other: Box<Self>) -> Box<Self> {
-                self
-            }
-
-            fn into_event(self: Box<Self>) -> serde_json::Value {
-                serde_json::json!({})
-            }
-        }
-    };
-}
+#[cfg(feature = "analytics")]
+pub type SearchAggregator = segment_analytics::SearchAggregator;
+#[cfg(feature = "analytics")]
+pub type SimilarAggregator = segment_analytics::SimilarAggregator;
+#[cfg(feature = "analytics")]
+pub type MultiSearchAggregator = segment_analytics::MultiSearchAggregator;
+#[cfg(feature = "analytics")]
+pub type FacetSearchAggregator = segment_analytics::FacetSearchAggregator;

 /// The Meilisearch config dir:
 /// `~/.config/Meilisearch` on *NIX or *BSD.
@@ -79,88 +78,60 @@ pub enum DocumentFetchKind {
    Normal { with_filter: bool, limit: usize, offset: usize, retrieve_vectors: bool },
 }

-/// To send an event to segment, your event must be able to aggregate itself with another event of the same type.
-pub trait Aggregate: 'static + mopa::Any + Send {
-    /// The name of the event that will be sent to segment.
-    fn event_name(&self) -> &'static str;
-
-    /// Will be called every time an event has been used twice before segment flushed its buffer.
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self>
-    where
-        Self: Sized;
-
-    /// Converts your structure to the final event that'll be sent to segment.
-    fn into_event(self: Box<Self>) -> serde_json::Value;
-}
-
-mopafy!(Aggregate);
-
-/// Helper trait to define multiple aggregates with the same content but a different name.
-/// Commonly used when you must aggregate a search with POST or with GET, for example.
-pub trait AggregateMethod: 'static + Default + Send {
-    fn event_name() -> &'static str;
-}
-
-/// A macro used to quickly define multiple aggregate method with their name
-/// Usage:
-/// ```rust
-/// use meilisearch::aggregate_methods;
-///
-/// aggregate_methods!(
-///     SearchGET => "Documents Searched GET",
-///     SearchPOST => "Documents Searched POST",
-/// );
-/// ```
-#[macro_export]
-macro_rules! aggregate_methods {
-    ($method:ident => $event_name:literal) => {
-        #[derive(Default)]
-        pub struct $method {}
-
-        impl $crate::analytics::AggregateMethod for $method {
-            fn event_name() -> &'static str {
-                $event_name
-            }
-        }
-    };
-    ($($method:ident => $event_name:literal,)+) => {
-        $(
-            aggregate_methods!($method => $event_name);
-        )+
-
-    };
-}
-
-#[derive(Clone)]
-pub struct Analytics {
-    segment: Option<Arc<SegmentAnalytics>>,
-}
-
-impl Analytics {
-    pub async fn new(
-        opt: &Opt,
-        index_scheduler: Arc<IndexScheduler>,
-        auth_controller: Arc<AuthController>,
-    ) -> Self {
-        if opt.no_analytics {
-            Self { segment: None }
-        } else {
-            Self { segment: SegmentAnalytics::new(opt, index_scheduler, auth_controller).await }
-        }
-    }
-
-    pub fn no_analytics() -> Self {
-        Self { segment: None }
-    }
-
-    pub fn instance_uid(&self) -> Option<&InstanceUid> {
-        self.segment.as_ref().map(|segment| segment.instance_uid.as_ref())
-    }
+pub trait Analytics: Sync + Send {
+    fn instance_uid(&self) -> Option<&InstanceUid>;

    /// The method used to publish most analytics that do not need to be batched every hours
-    pub fn publish<T: Aggregate>(&self, event: T, request: &HttpRequest) {
-        if let Some(ref segment) = self.segment {
-            let _ = segment.sender.try_send(segment_analytics::Message::new(event, request));
-        }
-    }
+    fn publish(&self, event_name: String, send: Value, request: Option<&HttpRequest>);
+
+    /// This method should be called to aggregate a get search
+    fn get_search(&self, aggregate: SearchAggregator);
+
+    /// This method should be called to aggregate a post search
+    fn post_search(&self, aggregate: SearchAggregator);
+
+    /// This method should be called to aggregate a get similar request
+    fn get_similar(&self, aggregate: SimilarAggregator);
+
+    /// This method should be called to aggregate a post similar request
+    fn post_similar(&self, aggregate: SimilarAggregator);
+
+    /// This method should be called to aggregate a post array of searches
+    fn post_multi_search(&self, aggregate: MultiSearchAggregator);
+
+    /// This method should be called to aggregate post facet values searches
+    fn post_facet_search(&self, aggregate: FacetSearchAggregator);
+
+    // this method should be called to aggregate an add documents request
+    fn add_documents(
+        &self,
+        documents_query: &UpdateDocumentsQuery,
+        index_creation: bool,
+        request: &HttpRequest,
+    );
+
+    // this method should be called to aggregate a fetch documents request
+    fn get_fetch_documents(&self, documents_query: &DocumentFetchKind, request: &HttpRequest);
+
+    // this method should be called to aggregate a fetch documents request
+    fn post_fetch_documents(&self, documents_query: &DocumentFetchKind, request: &HttpRequest);
+
+    // this method should be called to aggregate a add documents request
+    fn delete_documents(&self, kind: DocumentDeletionKind, request: &HttpRequest);
+
+    // this method should be called to batch an update documents request
+    fn update_documents(
+        &self,
+        documents_query: &UpdateDocumentsQuery,
+        index_creation: bool,
+        request: &HttpRequest,
+    );
+
+    // this method should be called to batch an update documents by function request
+    fn update_documents_by_function(
+        &self,
+        documents_query: &DocumentEditionByFunction,
+        index_creation: bool,
+        request: &HttpRequest,
+    );
 }
--- a/meilisearch/src/analytics/segment_analytics.rs
+++ b/meilisearch/src/analytics/segment_analytics.rs
--- a/meilisearch/src/lib.rs
+++ b/meilisearch/src/lib.rs
@@ -120,7 +120,7 @@ pub fn create_app(
    search_queue: Data<SearchQueue>,
    opt: Opt,
    logs: (LogRouteHandle, LogStderrHandle),
-    analytics: Data<Analytics>,
+    analytics: Arc<dyn Analytics>,
    enable_dashboard: bool,
 ) -> actix_web::App<
    impl ServiceFactory<
@@ -473,14 +473,14 @@ pub fn configure_data(
    search_queue: Data<SearchQueue>,
    opt: &Opt,
    (logs_route, logs_stderr): (LogRouteHandle, LogStderrHandle),
-    analytics: Data<Analytics>,
+    analytics: Arc<dyn Analytics>,
 ) {
    let http_payload_size_limit = opt.http_payload_size_limit.as_u64() as usize;
    config
        .app_data(index_scheduler)
        .app_data(auth)
        .app_data(search_queue)
-        .app_data(analytics)
+        .app_data(web::Data::from(analytics))
        .app_data(web::Data::new(logs_route))
        .app_data(web::Data::new(logs_stderr))
        .app_data(web::Data::new(opt.clone()))
--- a/meilisearch/src/main.rs
+++ b/meilisearch/src/main.rs
@@ -5,7 +5,6 @@ use std::path::PathBuf;
 use std::str::FromStr;
 use std::sync::Arc;
 use std::thread::available_parallelism;
-use std::time::Duration;

 use actix_web::http::KeepAlive;
 use actix_web::web::Data;
@@ -124,12 +123,19 @@ async fn try_main() -> anyhow::Result<()> {

    let (index_scheduler, auth_controller) = setup_meilisearch(&opt)?;

-    let analytics =
-        analytics::Analytics::new(&opt, index_scheduler.clone(), auth_controller.clone()).await;
+    #[cfg(all(not(debug_assertions), feature = "analytics"))]
+    let analytics = if !opt.no_analytics {
+        analytics::SegmentAnalytics::new(&opt, index_scheduler.clone(), auth_controller.clone())
+            .await
+    } else {
+        analytics::MockAnalytics::new(&opt)
+    };
+    #[cfg(any(debug_assertions, not(feature = "analytics")))]
+    let analytics = analytics::MockAnalytics::new(&opt);

    print_launch_resume(&opt, analytics.clone(), config_read_from);

-    run_http(index_scheduler, auth_controller, opt, log_handle, Arc::new(analytics)).await?;
+    run_http(index_scheduler, auth_controller, opt, log_handle, analytics).await?;

    Ok(())
 }
@@ -139,23 +145,16 @@ async fn run_http(
    auth_controller: Arc<AuthController>,
    opt: Opt,
    logs: (LogRouteHandle, LogStderrHandle),
-    analytics: Arc<Analytics>,
+    analytics: Arc<dyn Analytics>,
 ) -> anyhow::Result<()> {
    let enable_dashboard = &opt.env == "development";
    let opt_clone = opt.clone();
    let index_scheduler = Data::from(index_scheduler);
    let auth_controller = Data::from(auth_controller);
-    let analytics = Data::from(analytics);
    let search_queue = SearchQueue::new(
        opt.experimental_search_queue_size,
-        available_parallelism()
-            .unwrap_or(NonZeroUsize::new(2).unwrap())
-            .checked_mul(opt.experimental_nb_searches_per_core)
-            .unwrap_or(NonZeroUsize::MAX),
-    )
-    .with_time_to_abort(Duration::from_secs(
-        usize::from(opt.experimental_drop_search_after) as u64
-    ));
+        available_parallelism().unwrap_or(NonZeroUsize::new(2).unwrap()),
+    );
    let search_queue = Data::new(search_queue);

    let http_server = HttpServer::new(move || {
@@ -181,7 +180,11 @@ async fn run_http(
    Ok(())
 }

-pub fn print_launch_resume(opt: &Opt, analytics: Analytics, config_read_from: Option<PathBuf>) {
+pub fn print_launch_resume(
+    opt: &Opt,
+    analytics: Arc<dyn Analytics>,
+    config_read_from: Option<PathBuf>,
+) {
    let build_info = build_info::BuildInfo::from_build();

    let protocol =
@@ -223,6 +226,7 @@ pub fn print_launch_resume(opt: &Opt, analytics: Analytics, config_read_from: Op
        eprintln!("Prototype:\t\t{:?}", prototype);
    }

+    #[cfg(all(not(debug_assertions), feature = "analytics"))]
    {
        if !opt.no_analytics {
            eprintln!(
--- a/meilisearch/src/option.rs
+++ b/meilisearch/src/option.rs
@@ -2,7 +2,7 @@ use std::env::VarError;
 use std::ffi::OsStr;
 use std::fmt::Display;
 use std::io::{BufReader, Read};
-use std::num::{NonZeroUsize, ParseIntError};
+use std::num::ParseIntError;
 use std::ops::Deref;
 use std::path::PathBuf;
 use std::str::FromStr;
@@ -29,6 +29,7 @@ const MEILI_MASTER_KEY: &str = "MEILI_MASTER_KEY";
 const MEILI_ENV: &str = "MEILI_ENV";
 const MEILI_TASK_WEBHOOK_URL: &str = "MEILI_TASK_WEBHOOK_URL";
 const MEILI_TASK_WEBHOOK_AUTHORIZATION_HEADER: &str = "MEILI_TASK_WEBHOOK_AUTHORIZATION_HEADER";
+#[cfg(feature = "analytics")]
 const MEILI_NO_ANALYTICS: &str = "MEILI_NO_ANALYTICS";
 const MEILI_HTTP_PAYLOAD_SIZE_LIMIT: &str = "MEILI_HTTP_PAYLOAD_SIZE_LIMIT";
 const MEILI_SSL_CERT_PATH: &str = "MEILI_SSL_CERT_PATH";
@@ -54,8 +55,6 @@ const MEILI_EXPERIMENTAL_ENABLE_LOGS_ROUTE: &str = "MEILI_EXPERIMENTAL_ENABLE_LO
 const MEILI_EXPERIMENTAL_CONTAINS_FILTER: &str = "MEILI_EXPERIMENTAL_CONTAINS_FILTER";
 const MEILI_EXPERIMENTAL_ENABLE_METRICS: &str = "MEILI_EXPERIMENTAL_ENABLE_METRICS";
 const MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE: &str = "MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE";
-const MEILI_EXPERIMENTAL_DROP_SEARCH_AFTER: &str = "MEILI_EXPERIMENTAL_DROP_SEARCH_AFTER";
-const MEILI_EXPERIMENTAL_NB_SEARCHES_PER_CORE: &str = "MEILI_EXPERIMENTAL_NB_SEARCHES_PER_CORE";
 const MEILI_EXPERIMENTAL_REDUCE_INDEXING_MEMORY_USAGE: &str =
    "MEILI_EXPERIMENTAL_REDUCE_INDEXING_MEMORY_USAGE";
 const MEILI_EXPERIMENTAL_MAX_NUMBER_OF_BATCHED_TASKS: &str =
@@ -209,6 +208,7 @@ pub struct Opt {
    /// Meilisearch automatically collects data from all instances that do not opt out using this flag.
    /// All gathered data is used solely for the purpose of improving Meilisearch, and can be deleted
    /// at any time.
+    #[cfg(feature = "analytics")]
    #[serde(default)] // we can't send true
    #[clap(long, env = MEILI_NO_ANALYTICS)]
    pub no_analytics: bool,
@@ -357,26 +357,10 @@ pub struct Opt {
    /// Lets you customize the size of the search queue. Meilisearch processes your search requests as fast as possible but once the
    /// queue is full it starts returning HTTP 503, Service Unavailable.
    /// The default value is 1000.
-    #[clap(long, env = MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE, default_value_t = default_experimental_search_queue_size())]
-    #[serde(default = "default_experimental_search_queue_size")]
+    #[clap(long, env = MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE, default_value_t = 1000)]
+    #[serde(default)]
    pub experimental_search_queue_size: usize,

-    /// Experimental drop search after. For more information, see: <https://github.com/orgs/meilisearch/discussions/783>
-    ///
-    /// Let you customize after how many seconds Meilisearch should consider a search request irrelevant and drop it.
-    /// The default value is 60.
-    #[clap(long, env = MEILI_EXPERIMENTAL_DROP_SEARCH_AFTER, default_value_t = default_drop_search_after())]
-    #[serde(default = "default_drop_search_after")]
-    pub experimental_drop_search_after: NonZeroUsize,
-
-    /// Experimental number of searches per core. For more information, see: <https://github.com/orgs/meilisearch/discussions/784>
-    ///
-    /// Lets you customize how many search requests can run on each core concurrently.
-    /// The default value is 4.
-    #[clap(long, env = MEILI_EXPERIMENTAL_NB_SEARCHES_PER_CORE, default_value_t = default_nb_searches_per_core())]
-    #[serde(default = "default_nb_searches_per_core")]
-    pub experimental_nb_searches_per_core: NonZeroUsize,
-
    /// Experimental logs mode feature. For more information, see: <https://github.com/orgs/meilisearch/discussions/723>
    ///
    /// Change the mode of the logs on the console.
@@ -423,6 +407,7 @@ pub struct Opt {

 impl Opt {
    /// Whether analytics should be enabled or not.
+    #[cfg(all(not(debug_assertions), feature = "analytics"))]
    pub fn analytics(&self) -> bool {
        !self.no_analytics
    }
@@ -502,12 +487,11 @@ impl Opt {
            ignore_missing_dump: _,
            ignore_dump_if_db_exists: _,
            config_file_path: _,
+            #[cfg(feature = "analytics")]
            no_analytics,
            experimental_contains_filter,
            experimental_enable_metrics,
            experimental_search_queue_size,
-            experimental_drop_search_after,
-            experimental_nb_searches_per_core,
            experimental_logs_mode,
            experimental_enable_logs_route,
            experimental_replication_parameters,
@@ -529,7 +513,10 @@ impl Opt {
            );
        }

-        export_to_env_if_not_present(MEILI_NO_ANALYTICS, no_analytics.to_string());
+        #[cfg(feature = "analytics")]
+        {
+            export_to_env_if_not_present(MEILI_NO_ANALYTICS, no_analytics.to_string());
+        }
        export_to_env_if_not_present(
            MEILI_HTTP_PAYLOAD_SIZE_LIMIT,
            http_payload_size_limit.to_string(),
@@ -572,14 +559,6 @@ impl Opt {
            MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE,
            experimental_search_queue_size.to_string(),
        );
-        export_to_env_if_not_present(
-            MEILI_EXPERIMENTAL_DROP_SEARCH_AFTER,
-            experimental_drop_search_after.to_string(),
-        );
-        export_to_env_if_not_present(
-            MEILI_EXPERIMENTAL_NB_SEARCHES_PER_CORE,
-            experimental_nb_searches_per_core.to_string(),
-        );
        export_to_env_if_not_present(
            MEILI_EXPERIMENTAL_LOGS_MODE,
            experimental_logs_mode.to_string(),
@@ -911,18 +890,6 @@ fn default_dump_dir() -> PathBuf {
    PathBuf::from(DEFAULT_DUMP_DIR)
 }

-fn default_experimental_search_queue_size() -> usize {
-    1000
-}
-
-fn default_drop_search_after() -> NonZeroUsize {
-    NonZeroUsize::new(60).unwrap()
-}
-
-fn default_nb_searches_per_core() -> NonZeroUsize {
-    NonZeroUsize::new(4).unwrap()
-}
-
 /// Indicates if a snapshot was scheduled, and if yes with which interval.
 #[derive(Debug, Default, Copy, Clone, Deserialize, Serialize)]
 pub enum ScheduleSnapshot {
--- a/meilisearch/src/routes/dump.rs
+++ b/meilisearch/src/routes/dump.rs
@@ -4,6 +4,7 @@ use index_scheduler::IndexScheduler;
 use meilisearch_auth::AuthController;
 use meilisearch_types::error::ResponseError;
 use meilisearch_types::tasks::KindWithContent;
+use serde_json::json;
 use tracing::debug;

 use crate::analytics::Analytics;
@@ -17,16 +18,14 @@ pub fn configure(cfg: &mut web::ServiceConfig) {
    cfg.service(web::resource("").route(web::post().to(SeqHandler(create_dump))));
 }

-crate::empty_analytics!(DumpAnalytics, "Dump Created");
-
 pub async fn create_dump(
    index_scheduler: GuardedData<ActionPolicy<{ actions::DUMPS_CREATE }>, Data<IndexScheduler>>,
    auth_controller: GuardedData<ActionPolicy<{ actions::DUMPS_CREATE }>, Data<AuthController>>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
-    analytics.publish(DumpAnalytics::default(), &req);
+    analytics.publish("Dump Created".to_string(), json!({}), Some(&req));

    let task = KindWithContent::DumpCreation {
        keys: auth_controller.list_keys()?,
--- a/meilisearch/src/routes/features.rs
+++ b/meilisearch/src/routes/features.rs
@@ -6,10 +6,10 @@ use index_scheduler::IndexScheduler;
 use meilisearch_types::deserr::DeserrJsonError;
 use meilisearch_types::error::ResponseError;
 use meilisearch_types::keys::actions;
-use serde::Serialize;
+use serde_json::json;
 use tracing::debug;

-use crate::analytics::{Aggregate, Analytics};
+use crate::analytics::Analytics;
 use crate::extractors::authentication::policies::ActionPolicy;
 use crate::extractors::authentication::GuardedData;
 use crate::extractors::sequential_extractor::SeqHandler;
@@ -17,7 +17,7 @@ use crate::extractors::sequential_extractor::SeqHandler;
 pub fn configure(cfg: &mut web::ServiceConfig) {
    cfg.service(
        web::resource("")
-            .route(web::get().to(get_features))
+            .route(web::get().to(SeqHandler(get_features)))
            .route(web::patch().to(SeqHandler(patch_features))),
    );
 }
@@ -27,9 +27,12 @@ async fn get_features(
        ActionPolicy<{ actions::EXPERIMENTAL_FEATURES_GET }>,
        Data<IndexScheduler>,
    >,
+    req: HttpRequest,
+    analytics: Data<dyn Analytics>,
 ) -> HttpResponse {
    let features = index_scheduler.features();

+    analytics.publish("Experimental features Seen".to_string(), json!(null), Some(&req));
    let features = features.runtime_features();
    debug!(returns = ?features, "Get features");
    HttpResponse::Ok().json(features)
@@ -50,35 +53,6 @@ pub struct RuntimeTogglableFeatures {
    pub contains_filter: Option<bool>,
 }

-#[derive(Serialize)]
-pub struct PatchExperimentalFeatureAnalytics {
-    vector_store: bool,
-    metrics: bool,
-    logs_route: bool,
-    edit_documents_by_function: bool,
-    contains_filter: bool,
-}
-
-impl Aggregate for PatchExperimentalFeatureAnalytics {
-    fn event_name(&self) -> &'static str {
-        "Experimental features Updated"
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self {
-            vector_store: new.vector_store,
-            metrics: new.metrics,
-            logs_route: new.logs_route,
-            edit_documents_by_function: new.edit_documents_by_function,
-            contains_filter: new.contains_filter,
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
-
 async fn patch_features(
    index_scheduler: GuardedData<
        ActionPolicy<{ actions::EXPERIMENTAL_FEATURES_UPDATE }>,
@@ -86,7 +60,7 @@ async fn patch_features(
    >,
    new_features: AwebJson<RuntimeTogglableFeatures, DeserrJsonError>,
    req: HttpRequest,
-    analytics: Data<Analytics>,
+    analytics: Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let features = index_scheduler.features();
    debug!(parameters = ?new_features, "Patch features");
@@ -115,14 +89,15 @@ async fn patch_features(
    } = new_features;

    analytics.publish(
-        PatchExperimentalFeatureAnalytics {
-            vector_store,
-            metrics,
-            logs_route,
-            edit_documents_by_function,
-            contains_filter,
-        },
-        &req,
+        "Experimental features Updated".to_string(),
+        json!({
+            "vector_store": vector_store,
+            "metrics": metrics,
+            "logs_route": logs_route,
+            "edit_documents_by_function": edit_documents_by_function,
+            "contains_filter": contains_filter,
+        }),
+        Some(&req),
    );
    index_scheduler.put_runtime_features(new_features)?;
    debug!(returns = ?new_features, "Patch features");
--- a/meilisearch/src/routes/indexes/documents.rs
+++ b/meilisearch/src/routes/indexes/documents.rs
@@ -1,6 +1,4 @@
-use std::collections::HashSet;
 use std::io::ErrorKind;
-use std::marker::PhantomData;

 use actix_web::http::header::CONTENT_TYPE;
 use actix_web::web::Data;
@@ -25,14 +23,14 @@ use meilisearch_types::tasks::KindWithContent;
 use meilisearch_types::{milli, Document, Index};
 use mime::Mime;
 use once_cell::sync::Lazy;
-use serde::{Deserialize, Serialize};
+use serde::Deserialize;
 use serde_json::Value;
 use tempfile::tempfile;
 use tokio::fs::File;
 use tokio::io::{AsyncSeekExt, AsyncWriteExt, BufWriter};
 use tracing::debug;

-use crate::analytics::{Aggregate, AggregateMethod, Analytics};
+use crate::analytics::{Analytics, DocumentDeletionKind, DocumentFetchKind};
 use crate::error::MeilisearchHttpError;
 use crate::error::PayloadError::ReceivePayload;
 use crate::extractors::authentication::policies::*;
@@ -43,7 +41,7 @@ use crate::routes::{
    get_task_id, is_dry_run, PaginationView, SummarizedTaskView, PAGINATION_DEFAULT_LIMIT,
 };
 use crate::search::{parse_filter, RetrieveVectors};
-use crate::{aggregate_methods, Opt};
+use crate::Opt;

 static ACCEPTED_CONTENT_TYPE: Lazy<Vec<String>> = Lazy::new(|| {
    vec!["application/json".to_string(), "application/x-ndjson".to_string(), "text/csv".to_string()]
@@ -102,84 +100,12 @@ pub struct GetDocument {
    retrieve_vectors: Param<bool>,
 }

-aggregate_methods!(
-    DocumentsGET => "Documents Fetched GET",
-    DocumentsPOST => "Documents Fetched POST",
-);
-
-#[derive(Serialize)]
-pub struct DocumentsFetchAggregator<Method: AggregateMethod> {
-    // a call on ../documents/:doc_id
-    per_document_id: bool,
-    // if a filter was used
-    per_filter: bool,
-
-    #[serde(rename = "vector.retrieve_vectors")]
-    retrieve_vectors: bool,
-
-    // pagination
-    #[serde(rename = "pagination.max_limit")]
-    max_limit: usize,
-    #[serde(rename = "pagination.max_offset")]
-    max_offset: usize,
-
-    marker: std::marker::PhantomData<Method>,
-}
-
-#[derive(Copy, Clone, Debug, PartialEq, Eq)]
-pub enum DocumentFetchKind {
-    PerDocumentId { retrieve_vectors: bool },
-    Normal { with_filter: bool, limit: usize, offset: usize, retrieve_vectors: bool },
-}
-
-impl<Method: AggregateMethod> DocumentsFetchAggregator<Method> {
-    pub fn from_query(query: &DocumentFetchKind) -> Self {
-        let (limit, offset, retrieve_vectors) = match query {
-            DocumentFetchKind::PerDocumentId { retrieve_vectors } => (1, 0, *retrieve_vectors),
-            DocumentFetchKind::Normal { limit, offset, retrieve_vectors, .. } => {
-                (*limit, *offset, *retrieve_vectors)
-            }
-        };
-
-        Self {
-            per_document_id: matches!(query, DocumentFetchKind::PerDocumentId { .. }),
-            per_filter: matches!(query, DocumentFetchKind::Normal { with_filter, .. } if *with_filter),
-            max_limit: limit,
-            max_offset: offset,
-            retrieve_vectors,
-
-            marker: PhantomData,
-        }
-    }
-}
-
-impl<Method: AggregateMethod> Aggregate for DocumentsFetchAggregator<Method> {
-    fn event_name(&self) -> &'static str {
-        Method::event_name()
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self {
-            per_document_id: self.per_document_id | new.per_document_id,
-            per_filter: self.per_filter | new.per_filter,
-            retrieve_vectors: self.retrieve_vectors | new.retrieve_vectors,
-            max_limit: self.max_limit.max(new.max_limit),
-            max_offset: self.max_offset.max(new.max_offset),
-            marker: PhantomData,
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
-
 pub async fn get_document(
    index_scheduler: GuardedData<ActionPolicy<{ actions::DOCUMENTS_GET }>, Data<IndexScheduler>>,
    document_param: web::Path<DocumentParam>,
    params: AwebQueryParameter<GetDocument, DeserrQueryParamError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let DocumentParam { index_uid, document_id } = document_param.into_inner();
    debug!(parameters = ?params, "Get document");
@@ -191,15 +117,8 @@ pub async fn get_document(
    let features = index_scheduler.features();
    let retrieve_vectors = RetrieveVectors::new(param_retrieve_vectors.0, features)?;

-    analytics.publish(
-        DocumentsFetchAggregator::<DocumentsGET> {
-            retrieve_vectors: param_retrieve_vectors.0,
-            per_document_id: true,
-            per_filter: false,
-            max_limit: 0,
-            max_offset: 0,
-            marker: PhantomData,
-        },
+    analytics.get_fetch_documents(
+        &DocumentFetchKind::PerDocumentId { retrieve_vectors: param_retrieve_vectors.0 },
        &req,
    );

@@ -210,52 +129,17 @@ pub async fn get_document(
    Ok(HttpResponse::Ok().json(document))
 }

-#[derive(Serialize)]
-pub struct DocumentsDeletionAggregator {
-    per_document_id: bool,
-    clear_all: bool,
-    per_batch: bool,
-    per_filter: bool,
-}
-
-impl Aggregate for DocumentsDeletionAggregator {
-    fn event_name(&self) -> &'static str {
-        "Documents Deleted"
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self {
-            per_document_id: self.per_document_id | new.per_document_id,
-            clear_all: self.clear_all | new.clear_all,
-            per_batch: self.per_batch | new.per_batch,
-            per_filter: self.per_filter | new.per_filter,
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
-
 pub async fn delete_document(
    index_scheduler: GuardedData<ActionPolicy<{ actions::DOCUMENTS_DELETE }>, Data<IndexScheduler>>,
    path: web::Path<DocumentParam>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let DocumentParam { index_uid, document_id } = path.into_inner();
    let index_uid = IndexUid::try_from(index_uid)?;

-    analytics.publish(
-        DocumentsDeletionAggregator {
-            per_document_id: true,
-            clear_all: false,
-            per_batch: false,
-            per_filter: false,
-        },
-        &req,
-    );
+    analytics.delete_documents(DocumentDeletionKind::PerDocumentId, &req);

    let task = KindWithContent::DocumentDeletion {
        index_uid: index_uid.to_string(),
@@ -306,19 +190,17 @@ pub async fn documents_by_query_post(
    index_uid: web::Path<String>,
    body: AwebJson<BrowseQuery, DeserrJsonError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let body = body.into_inner();
    debug!(parameters = ?body, "Get documents POST");

-    analytics.publish(
-        DocumentsFetchAggregator::<DocumentsPOST> {
-            per_filter: body.filter.is_some(),
+    analytics.post_fetch_documents(
+        &DocumentFetchKind::Normal {
+            with_filter: body.filter.is_some(),
+            limit: body.limit,
+            offset: body.offset,
            retrieve_vectors: body.retrieve_vectors,
-            max_limit: body.limit,
-            max_offset: body.offset,
-            per_document_id: false,
-            marker: PhantomData,
        },
        &req,
    );
@@ -331,7 +213,7 @@ pub async fn get_documents(
    index_uid: web::Path<String>,
    params: AwebQueryParameter<BrowseQueryGet, DeserrQueryParamError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    debug!(parameters = ?params, "Get documents GET");

@@ -353,14 +235,12 @@ pub async fn get_documents(
        filter,
    };

-    analytics.publish(
-        DocumentsFetchAggregator::<DocumentsGET> {
-            per_filter: query.filter.is_some(),
+    analytics.get_fetch_documents(
+        &DocumentFetchKind::Normal {
+            with_filter: query.filter.is_some(),
+            limit: query.limit,
+            offset: query.offset,
            retrieve_vectors: query.retrieve_vectors,
-            max_limit: query.limit,
-            max_offset: query.offset,
-            per_document_id: false,
-            marker: PhantomData,
        },
        &req,
    );
@@ -418,39 +298,6 @@ fn from_char_csv_delimiter(
    }
 }

-aggregate_methods!(
-    Replaced => "Documents Added",
-    Updated => "Documents Updated",
-);
-
-#[derive(Serialize)]
-pub struct DocumentsAggregator<T: AggregateMethod> {
-    payload_types: HashSet<String>,
-    primary_key: HashSet<String>,
-    index_creation: bool,
-    #[serde(skip)]
-    method: PhantomData<T>,
-}
-
-impl<Method: AggregateMethod> Aggregate for DocumentsAggregator<Method> {
-    fn event_name(&self) -> &'static str {
-        Method::event_name()
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self {
-            payload_types: self.payload_types.union(&new.payload_types).cloned().collect(),
-            primary_key: self.primary_key.union(&new.primary_key).cloned().collect(),
-            index_creation: self.index_creation | new.index_creation,
-            method: PhantomData,
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(self).unwrap_or_default()
-    }
-}
-
 pub async fn replace_documents(
    index_scheduler: GuardedData<ActionPolicy<{ actions::DOCUMENTS_ADD }>, Data<IndexScheduler>>,
    index_uid: web::Path<String>,
@@ -458,32 +305,16 @@ pub async fn replace_documents(
    body: Payload,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;

    debug!(parameters = ?params, "Replace documents");
    let params = params.into_inner();

-    let mut content_types = HashSet::new();
-    let content_type = req
-        .headers()
-        .get(CONTENT_TYPE)
-        .and_then(|s| s.to_str().ok())
-        .unwrap_or("unknown")
-        .to_string();
-    content_types.insert(content_type);
-    let mut primary_keys = HashSet::new();
-    if let Some(primary_key) = params.primary_key.clone() {
-        primary_keys.insert(primary_key);
-    }
-    analytics.publish(
-        DocumentsAggregator::<Replaced> {
-            payload_types: content_types,
-            primary_key: primary_keys,
-            index_creation: index_scheduler.index_exists(&index_uid).map_or(true, |x| !x),
-            method: PhantomData,
-        },
+    analytics.add_documents(
+        &params,
+        index_scheduler.index_exists(&index_uid).map_or(true, |x| !x),
        &req,
    );

@@ -515,32 +346,16 @@ pub async fn update_documents(
    body: Payload,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;

    let params = params.into_inner();
    debug!(parameters = ?params, "Update documents");

-    let mut content_types = HashSet::new();
-    let content_type = req
-        .headers()
-        .get(CONTENT_TYPE)
-        .and_then(|s| s.to_str().ok())
-        .unwrap_or("unknown")
-        .to_string();
-    content_types.insert(content_type);
-    let mut primary_keys = HashSet::new();
-    if let Some(primary_key) = params.primary_key.clone() {
-        primary_keys.insert(primary_key);
-    }
-    analytics.publish(
-        DocumentsAggregator::<Updated> {
-            payload_types: content_types,
-            primary_key: primary_keys,
-            index_creation: index_scheduler.index_exists(&index_uid).map_or(true, |x| !x),
-            method: PhantomData,
-        },
+    analytics.add_documents(
+        &params,
+        index_scheduler.index_exists(&index_uid).map_or(true, |x| !x),
        &req,
    );

@@ -709,20 +524,12 @@ pub async fn delete_documents_batch(
    body: web::Json<Vec<Value>>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    debug!(parameters = ?body, "Delete documents by batch");
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;

-    analytics.publish(
-        DocumentsDeletionAggregator {
-            per_batch: true,
-            per_document_id: false,
-            clear_all: false,
-            per_filter: false,
-        },
-        &req,
-    );
+    analytics.delete_documents(DocumentDeletionKind::PerBatch, &req);

    let ids = body
        .iter()
@@ -755,22 +562,14 @@ pub async fn delete_documents_by_filter(
    body: AwebJson<DocumentDeletionByFilter, DeserrJsonError>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    debug!(parameters = ?body, "Delete documents by filter");
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;
    let index_uid = index_uid.into_inner();
    let filter = body.into_inner().filter;

-    analytics.publish(
-        DocumentsDeletionAggregator {
-            per_filter: true,
-            per_document_id: false,
-            clear_all: false,
-            per_batch: false,
-        },
-        &req,
-    );
+    analytics.delete_documents(DocumentDeletionKind::PerFilter, &req);

    // we ensure the filter is well formed before enqueuing it
    crate::search::parse_filter(&filter, Code::InvalidDocumentFilter, index_scheduler.features())?
@@ -800,41 +599,13 @@ pub struct DocumentEditionByFunction {
    pub function: String,
 }

-#[derive(Serialize)]
-struct EditDocumentsByFunctionAggregator {
-    // Set to true if at least one request was filtered
-    filtered: bool,
-    // Set to true if at least one request contained a context
-    with_context: bool,
-
-    index_creation: bool,
-}
-
-impl Aggregate for EditDocumentsByFunctionAggregator {
-    fn event_name(&self) -> &'static str {
-        "Documents Edited By Function"
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self {
-            filtered: self.filtered | new.filtered,
-            with_context: self.with_context | new.with_context,
-            index_creation: self.index_creation | new.index_creation,
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
-
 pub async fn edit_documents_by_function(
    index_scheduler: GuardedData<ActionPolicy<{ actions::DOCUMENTS_ALL }>, Data<IndexScheduler>>,
    index_uid: web::Path<String>,
    params: AwebJson<DocumentEditionByFunction, DeserrJsonError>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    debug!(parameters = ?params, "Edit documents by function");

@@ -846,12 +617,9 @@ pub async fn edit_documents_by_function(
    let index_uid = index_uid.into_inner();
    let params = params.into_inner();

-    analytics.publish(
-        EditDocumentsByFunctionAggregator {
-            filtered: params.filter.is_some(),
-            with_context: params.context.is_some(),
-            index_creation: index_scheduler.index(&index_uid).is_err(),
-        },
+    analytics.update_documents_by_function(
+        &params,
+        index_scheduler.index(&index_uid).is_err(),
        &req,
    );

@@ -902,18 +670,10 @@ pub async fn clear_all_documents(
    index_uid: web::Path<String>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;
-    analytics.publish(
-        DocumentsDeletionAggregator {
-            clear_all: true,
-            per_document_id: false,
-            per_batch: false,
-            per_filter: false,
-        },
-        &req,
-    );
+    analytics.delete_documents(DocumentDeletionKind::ClearAll, &req);

    let task = KindWithContent::DocumentClear { index_uid: index_uid.to_string() };
    let uid = get_task_id(&req, &opt)?;
--- a/meilisearch/src/routes/indexes/facet_search.rs
+++ b/meilisearch/src/routes/indexes/facet_search.rs
@@ -1,5 +1,3 @@
-use std::collections::{BinaryHeap, HashSet};
-
 use actix_web::web::Data;
 use actix_web::{web, HttpRequest, HttpResponse};
 use deserr::actix_web::AwebJson;
@@ -12,15 +10,14 @@ use meilisearch_types::locales::Locale;
 use serde_json::Value;
 use tracing::debug;

-use crate::analytics::{Aggregate, Analytics};
+use crate::analytics::{Analytics, FacetSearchAggregator};
 use crate::extractors::authentication::policies::*;
 use crate::extractors::authentication::GuardedData;
 use crate::routes::indexes::search::search_kind;
 use crate::search::{
-    add_search_rules, perform_facet_search, FacetSearchResult, HybridQuery, MatchingStrategy,
-    RankingScoreThreshold, SearchQuery, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER,
-    DEFAULT_HIGHLIGHT_POST_TAG, DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT,
-    DEFAULT_SEARCH_OFFSET,
+    add_search_rules, perform_facet_search, HybridQuery, MatchingStrategy, RankingScoreThreshold,
+    SearchQuery, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER, DEFAULT_HIGHLIGHT_POST_TAG,
+    DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET,
 };
 use crate::search_queue::SearchQueue;

@@ -56,122 +53,20 @@ pub struct FacetSearchQuery {
    pub locales: Option<Vec<Locale>>,
 }

-#[derive(Default)]
-pub struct FacetSearchAggregator {
-    // requests
-    total_received: usize,
-    total_succeeded: usize,
-    time_spent: BinaryHeap<usize>,
-
-    // The set of all facetNames that were used
-    facet_names: HashSet<String>,
-
-    // As there been any other parameter than the facetName or facetQuery ones?
-    additional_search_parameters_provided: bool,
-}
-
-impl FacetSearchAggregator {
-    #[allow(clippy::field_reassign_with_default)]
-    pub fn from_query(query: &FacetSearchQuery) -> Self {
-        let FacetSearchQuery {
-            facet_query: _,
-            facet_name,
-            vector,
-            q,
-            filter,
-            matching_strategy,
-            attributes_to_search_on,
-            hybrid,
-            ranking_score_threshold,
-            locales,
-        } = query;
-
-        Self {
-            total_received: 1,
-            facet_names: Some(facet_name.clone()).into_iter().collect(),
-            additional_search_parameters_provided: q.is_some()
-                || vector.is_some()
-                || filter.is_some()
-                || *matching_strategy != MatchingStrategy::default()
-                || attributes_to_search_on.is_some()
-                || hybrid.is_some()
-                || ranking_score_threshold.is_some()
-                || locales.is_some(),
-            ..Default::default()
-        }
-    }
-
-    pub fn succeed(&mut self, result: &FacetSearchResult) {
-        let FacetSearchResult { facet_hits: _, facet_query: _, processing_time_ms } = result;
-        self.total_succeeded = 1;
-        self.time_spent.push(*processing_time_ms as usize);
-    }
-}
-
-impl Aggregate for FacetSearchAggregator {
-    fn event_name(&self) -> &'static str {
-        "Facet Searched POST"
-    }
-
-    fn aggregate(mut self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        for time in new.time_spent {
-            self.time_spent.push(time);
-        }
-
-        Box::new(Self {
-            total_received: self.total_received.saturating_add(new.total_received),
-            total_succeeded: self.total_succeeded.saturating_add(new.total_succeeded),
-            time_spent: self.time_spent,
-            facet_names: self.facet_names.union(&new.facet_names).cloned().collect(),
-            additional_search_parameters_provided: self.additional_search_parameters_provided
-                | new.additional_search_parameters_provided,
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        let Self {
-            total_received,
-            total_succeeded,
-            time_spent,
-            facet_names,
-            additional_search_parameters_provided,
-        } = *self;
-        // the index of the 99th percentage of value
-        let percentile_99th = 0.99 * (total_succeeded as f64 - 1.) + 1.;
-        // we get all the values in a sorted manner
-        let time_spent = time_spent.into_sorted_vec();
-        // We are only interested by the slowest value of the 99th fastest results
-        let time_spent = time_spent.get(percentile_99th as usize);
-
-        serde_json::json!({
-            "requests": {
-                "99th_response_time":  time_spent.map(|t| format!("{:.2}", t)),
-                "total_succeeded": total_succeeded,
-                "total_failed": total_received.saturating_sub(total_succeeded), // just to be sure we never panics
-                "total_received": total_received,
-            },
-            "facets": {
-                "total_distinct_facet_count": facet_names.len(),
-                "additional_search_parameters_provided": additional_search_parameters_provided,
-            },
-        })
-    }
-}
-
 pub async fn search(
    index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
    search_queue: Data<SearchQueue>,
    index_uid: web::Path<String>,
    params: AwebJson<FacetSearchQuery, DeserrJsonError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;

    let query = params.into_inner();
    debug!(parameters = ?query, "Facet search");

-    let mut aggregate = FacetSearchAggregator::from_query(&query);
+    let mut aggregate = FacetSearchAggregator::from_query(&query, &req);

    let facet_query = query.facet_query.clone();
    let facet_name = query.facet_name.clone();
@@ -205,7 +100,7 @@ pub async fn search(
    if let Ok(ref search_result) = search_result {
        aggregate.succeed(search_result);
    }
-    analytics.publish(aggregate, &req);
+    analytics.post_facet_search(aggregate);

    let search_result = search_result?;

--- a/meilisearch/src/routes/indexes/mod.rs
+++ b/meilisearch/src/routes/indexes/mod.rs
@@ -1,4 +1,3 @@
-use std::collections::BTreeSet;
 use std::convert::Infallible;

 use actix_web::web::Data;
@@ -14,11 +13,12 @@ use meilisearch_types::index_uid::IndexUid;
 use meilisearch_types::milli::{self, FieldDistribution, Index};
 use meilisearch_types::tasks::KindWithContent;
 use serde::Serialize;
+use serde_json::json;
 use time::OffsetDateTime;
 use tracing::debug;

 use super::{get_task_id, Pagination, SummarizedTaskView, PAGINATION_DEFAULT_LIMIT};
-use crate::analytics::{Aggregate, Analytics};
+use crate::analytics::Analytics;
 use crate::extractors::authentication::policies::*;
 use crate::extractors::authentication::{AuthenticationError, GuardedData};
 use crate::extractors::sequential_extractor::SeqHandler;
@@ -28,11 +28,8 @@ use crate::Opt;
 pub mod documents;
 pub mod facet_search;
 pub mod search;
-mod search_analytics;
 pub mod settings;
-mod settings_analytics;
 pub mod similar;
-mod similar_analytics;

 pub fn configure(cfg: &mut web::ServiceConfig) {
    cfg.service(
@@ -126,31 +123,12 @@ pub struct IndexCreateRequest {
    primary_key: Option<String>,
 }

-#[derive(Serialize)]
-struct IndexCreatedAggregate {
-    primary_key: BTreeSet<String>,
-}
-
-impl Aggregate for IndexCreatedAggregate {
-    fn event_name(&self) -> &'static str {
-        "Index Created"
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self { primary_key: self.primary_key.union(&new.primary_key).cloned().collect() })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
-
 pub async fn create_index(
    index_scheduler: GuardedData<ActionPolicy<{ actions::INDEXES_CREATE }>, Data<IndexScheduler>>,
    body: AwebJson<IndexCreateRequest, DeserrJsonError>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    debug!(parameters = ?body, "Create index");
    let IndexCreateRequest { primary_key, uid } = body.into_inner();
@@ -158,8 +136,9 @@ pub async fn create_index(
    let allow_index_creation = index_scheduler.filters().allow_index_creation(&uid);
    if allow_index_creation {
        analytics.publish(
-            IndexCreatedAggregate { primary_key: primary_key.iter().cloned().collect() },
-            &req,
+            "Index Created".to_string(),
+            json!({ "primary_key": primary_key }),
+            Some(&req),
        );

        let task = KindWithContent::IndexCreation { index_uid: uid.to_string(), primary_key };
@@ -215,38 +194,21 @@ pub async fn get_index(
    Ok(HttpResponse::Ok().json(index_view))
 }

-#[derive(Serialize)]
-struct IndexUpdatedAggregate {
-    primary_key: BTreeSet<String>,
-}
-
-impl Aggregate for IndexUpdatedAggregate {
-    fn event_name(&self) -> &'static str {
-        "Index Updated"
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self { primary_key: self.primary_key.union(&new.primary_key).cloned().collect() })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
 pub async fn update_index(
    index_scheduler: GuardedData<ActionPolicy<{ actions::INDEXES_UPDATE }>, Data<IndexScheduler>>,
    index_uid: web::Path<String>,
    body: AwebJson<UpdateIndexRequest, DeserrJsonError>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    debug!(parameters = ?body, "Update index");
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;
    let body = body.into_inner();
    analytics.publish(
-        IndexUpdatedAggregate { primary_key: body.primary_key.iter().cloned().collect() },
-        &req,
+        "Index Updated".to_string(),
+        json!({ "primary_key": body.primary_key }),
+        Some(&req),
    );

    let task = KindWithContent::IndexUpdate {
--- a/meilisearch/src/routes/indexes/search.rs
+++ b/meilisearch/src/routes/indexes/search.rs
@@ -13,13 +13,12 @@ use meilisearch_types::serde_cs::vec::CS;
 use serde_json::Value;
 use tracing::debug;

-use crate::analytics::Analytics;
+use crate::analytics::{Analytics, SearchAggregator};
 use crate::error::MeilisearchHttpError;
 use crate::extractors::authentication::policies::*;
 use crate::extractors::authentication::GuardedData;
 use crate::extractors::sequential_extractor::SeqHandler;
 use crate::metrics::MEILISEARCH_DEGRADED_SEARCH_REQUESTS;
-use crate::routes::indexes::search_analytics::{SearchAggregator, SearchGET, SearchPOST};
 use crate::search::{
    add_search_rules, perform_search, HybridQuery, MatchingStrategy, RankingScoreThreshold,
    RetrieveVectors, SearchKind, SearchQuery, SemanticRatio, DEFAULT_CROP_LENGTH,
@@ -226,7 +225,7 @@ pub async fn search_with_url_query(
    index_uid: web::Path<String>,
    params: AwebQueryParameter<SearchQueryGet, DeserrQueryParamError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    debug!(parameters = ?params, "Search get");
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;
@@ -238,7 +237,7 @@ pub async fn search_with_url_query(
        add_search_rules(&mut query.filter, search_rules);
    }

-    let mut aggregate = SearchAggregator::<SearchGET>::from_query(&query);
+    let mut aggregate = SearchAggregator::from_query(&query, &req);

    let index = index_scheduler.index(&index_uid)?;
    let features = index_scheduler.features();
@@ -255,7 +254,7 @@ pub async fn search_with_url_query(
    if let Ok(ref search_result) = search_result {
        aggregate.succeed(search_result);
    }
-    analytics.publish(aggregate, &req);
+    analytics.get_search(aggregate);

    let search_result = search_result?;

@@ -269,7 +268,7 @@ pub async fn search_with_post(
    index_uid: web::Path<String>,
    params: AwebJson<SearchQuery, DeserrJsonError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;

@@ -281,7 +280,7 @@ pub async fn search_with_post(
        add_search_rules(&mut query.filter, search_rules);
    }

-    let mut aggregate = SearchAggregator::<SearchPOST>::from_query(&query);
+    let mut aggregate = SearchAggregator::from_query(&query, &req);

    let index = index_scheduler.index(&index_uid)?;

@@ -303,7 +302,7 @@ pub async fn search_with_post(
            MEILISEARCH_DEGRADED_SEARCH_REQUESTS.inc();
        }
    }
-    analytics.publish(aggregate, &req);
+    analytics.post_search(aggregate);

    let search_result = search_result?;

--- a/meilisearch/src/routes/indexes/search_analytics.rs
+++ b/meilisearch/src/routes/indexes/search_analytics.rs
@@ -1,485 +0,0 @@
-use once_cell::sync::Lazy;
-use regex::Regex;
-use serde_json::{json, Value};
-use std::collections::{BTreeSet, BinaryHeap, HashMap};
-
-use meilisearch_types::locales::Locale;
-
-use crate::{
-    aggregate_methods,
-    analytics::{Aggregate, AggregateMethod},
-    search::{
-        SearchQuery, SearchResult, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER,
-        DEFAULT_HIGHLIGHT_POST_TAG, DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT,
-        DEFAULT_SEMANTIC_RATIO,
-    },
-};
-
-aggregate_methods!(
-    SearchGET => "Documents Searched GET",
-    SearchPOST => "Documents Searched POST",
-);
-
-#[derive(Default)]
-pub struct SearchAggregator<Method: AggregateMethod> {
-    // requests
-    total_received: usize,
-    total_succeeded: usize,
-    total_degraded: usize,
-    total_used_negative_operator: usize,
-    time_spent: BinaryHeap<usize>,
-
-    // sort
-    sort_with_geo_point: bool,
-    // every time a request has a filter, this field must be incremented by the number of terms it contains
-    sort_sum_of_criteria_terms: usize,
-    // every time a request has a filter, this field must be incremented by one
-    sort_total_number_of_criteria: usize,
-
-    // distinct
-    distinct: bool,
-
-    // filter
-    filter_with_geo_radius: bool,
-    filter_with_geo_bounding_box: bool,
-    // every time a request has a filter, this field must be incremented by the number of terms it contains
-    filter_sum_of_criteria_terms: usize,
-    // every time a request has a filter, this field must be incremented by one
-    filter_total_number_of_criteria: usize,
-    used_syntax: HashMap<String, usize>,
-
-    // attributes_to_search_on
-    // every time a search is done using attributes_to_search_on
-    attributes_to_search_on_total_number_of_uses: usize,
-
-    // q
-    // The maximum number of terms in a q request
-    max_terms_number: usize,
-
-    // vector
-    // The maximum number of floats in a vector request
-    max_vector_size: usize,
-    // Whether the semantic ratio passed to a hybrid search equals the default ratio.
-    semantic_ratio: bool,
-    hybrid: bool,
-    retrieve_vectors: bool,
-
-    // every time a search is done, we increment the counter linked to the used settings
-    matching_strategy: HashMap<String, usize>,
-
-    // List of the unique Locales passed as parameter
-    locales: BTreeSet<Locale>,
-
-    // pagination
-    max_limit: usize,
-    max_offset: usize,
-    finite_pagination: usize,
-
-    // formatting
-    max_attributes_to_retrieve: usize,
-    max_attributes_to_highlight: usize,
-    highlight_pre_tag: bool,
-    highlight_post_tag: bool,
-    max_attributes_to_crop: usize,
-    crop_marker: bool,
-    show_matches_position: bool,
-    crop_length: bool,
-
-    // facets
-    facets_sum_of_terms: usize,
-    facets_total_number_of_facets: usize,
-
-    // scoring
-    show_ranking_score: bool,
-    show_ranking_score_details: bool,
-    ranking_score_threshold: bool,
-
-    marker: std::marker::PhantomData<Method>,
-}
-
-impl<Method: AggregateMethod> SearchAggregator<Method> {
-    #[allow(clippy::field_reassign_with_default)]
-    pub fn from_query(query: &SearchQuery) -> Self {
-        let SearchQuery {
-            q,
-            vector,
-            offset,
-            limit,
-            page,
-            hits_per_page,
-            attributes_to_retrieve: _,
-            retrieve_vectors,
-            attributes_to_crop: _,
-            crop_length,
-            attributes_to_highlight: _,
-            show_matches_position,
-            show_ranking_score,
-            show_ranking_score_details,
-            filter,
-            sort,
-            distinct,
-            facets: _,
-            highlight_pre_tag,
-            highlight_post_tag,
-            crop_marker,
-            matching_strategy,
-            attributes_to_search_on,
-            hybrid,
-            ranking_score_threshold,
-            locales,
-        } = query;
-
-        let mut ret = Self::default();
-
-        ret.total_received = 1;
-
-        if let Some(ref sort) = sort {
-            ret.sort_total_number_of_criteria = 1;
-            ret.sort_with_geo_point = sort.iter().any(|s| s.contains("_geoPoint("));
-            ret.sort_sum_of_criteria_terms = sort.len();
-        }
-
-        ret.distinct = distinct.is_some();
-
-        if let Some(ref filter) = filter {
-            static RE: Lazy<Regex> = Lazy::new(|| Regex::new("AND | OR").unwrap());
-            ret.filter_total_number_of_criteria = 1;
-
-            let syntax = match filter {
-                Value::String(_) => "string".to_string(),
-                Value::Array(values) => {
-                    if values.iter().map(|v| v.to_string()).any(|s| RE.is_match(&s)) {
-                        "mixed".to_string()
-                    } else {
-                        "array".to_string()
-                    }
-                }
-                _ => "none".to_string(),
-            };
-            // convert the string to a HashMap
-            ret.used_syntax.insert(syntax, 1);
-
-            let stringified_filters = filter.to_string();
-            ret.filter_with_geo_radius = stringified_filters.contains("_geoRadius(");
-            ret.filter_with_geo_bounding_box = stringified_filters.contains("_geoBoundingBox(");
-            ret.filter_sum_of_criteria_terms = RE.split(&stringified_filters).count();
-        }
-
-        // attributes_to_search_on
-        if attributes_to_search_on.is_some() {
-            ret.attributes_to_search_on_total_number_of_uses = 1;
-        }
-
-        if let Some(ref q) = q {
-            ret.max_terms_number = q.split_whitespace().count();
-        }
-
-        if let Some(ref vector) = vector {
-            ret.max_vector_size = vector.len();
-        }
-        ret.retrieve_vectors |= retrieve_vectors;
-
-        if query.is_finite_pagination() {
-            let limit = hits_per_page.unwrap_or_else(DEFAULT_SEARCH_LIMIT);
-            ret.max_limit = limit;
-            ret.max_offset = page.unwrap_or(1).saturating_sub(1) * limit;
-            ret.finite_pagination = 1;
-        } else {
-            ret.max_limit = *limit;
-            ret.max_offset = *offset;
-            ret.finite_pagination = 0;
-        }
-
-        ret.matching_strategy.insert(format!("{:?}", matching_strategy), 1);
-
-        if let Some(locales) = locales {
-            ret.locales = locales.iter().copied().collect();
-        }
-
-        ret.highlight_pre_tag = *highlight_pre_tag != DEFAULT_HIGHLIGHT_PRE_TAG();
-        ret.highlight_post_tag = *highlight_post_tag != DEFAULT_HIGHLIGHT_POST_TAG();
-        ret.crop_marker = *crop_marker != DEFAULT_CROP_MARKER();
-        ret.crop_length = *crop_length != DEFAULT_CROP_LENGTH();
-        ret.show_matches_position = *show_matches_position;
-
-        ret.show_ranking_score = *show_ranking_score;
-        ret.show_ranking_score_details = *show_ranking_score_details;
-        ret.ranking_score_threshold = ranking_score_threshold.is_some();
-
-        if let Some(hybrid) = hybrid {
-            ret.semantic_ratio = hybrid.semantic_ratio != DEFAULT_SEMANTIC_RATIO();
-            ret.hybrid = true;
-        }
-
-        ret
-    }
-
-    pub fn succeed(&mut self, result: &SearchResult) {
-        let SearchResult {
-            hits: _,
-            query: _,
-            processing_time_ms,
-            hits_info: _,
-            semantic_hit_count: _,
-            facet_distribution: _,
-            facet_stats: _,
-            degraded,
-            used_negative_operator,
-        } = result;
-
-        self.total_succeeded = self.total_succeeded.saturating_add(1);
-        if *degraded {
-            self.total_degraded = self.total_degraded.saturating_add(1);
-        }
-        if *used_negative_operator {
-            self.total_used_negative_operator = self.total_used_negative_operator.saturating_add(1);
-        }
-        self.time_spent.push(*processing_time_ms as usize);
-    }
-}
-
-impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
-    fn event_name(&self) -> &'static str {
-        Method::event_name()
-    }
-
-    fn aggregate(mut self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        let Self {
-            total_received,
-            total_succeeded,
-            mut time_spent,
-            sort_with_geo_point,
-            sort_sum_of_criteria_terms,
-            sort_total_number_of_criteria,
-            distinct,
-            filter_with_geo_radius,
-            filter_with_geo_bounding_box,
-            filter_sum_of_criteria_terms,
-            filter_total_number_of_criteria,
-            used_syntax,
-            attributes_to_search_on_total_number_of_uses,
-            max_terms_number,
-            max_vector_size,
-            retrieve_vectors,
-            matching_strategy,
-            max_limit,
-            max_offset,
-            finite_pagination,
-            max_attributes_to_retrieve,
-            max_attributes_to_highlight,
-            highlight_pre_tag,
-            highlight_post_tag,
-            max_attributes_to_crop,
-            crop_marker,
-            show_matches_position,
-            crop_length,
-            facets_sum_of_terms,
-            facets_total_number_of_facets,
-            show_ranking_score,
-            show_ranking_score_details,
-            semantic_ratio,
-            hybrid,
-            total_degraded,
-            total_used_negative_operator,
-            ranking_score_threshold,
-            mut locales,
-            marker: _,
-        } = *new;
-
-        // request
-        self.total_received = self.total_received.saturating_add(total_received);
-        self.total_succeeded = self.total_succeeded.saturating_add(total_succeeded);
-        self.total_degraded = self.total_degraded.saturating_add(total_degraded);
-        self.total_used_negative_operator =
-            self.total_used_negative_operator.saturating_add(total_used_negative_operator);
-        self.time_spent.append(&mut time_spent);
-
-        // sort
-        self.sort_with_geo_point |= sort_with_geo_point;
-        self.sort_sum_of_criteria_terms =
-            self.sort_sum_of_criteria_terms.saturating_add(sort_sum_of_criteria_terms);
-        self.sort_total_number_of_criteria =
-            self.sort_total_number_of_criteria.saturating_add(sort_total_number_of_criteria);
-
-        // distinct
-        self.distinct |= distinct;
-
-        // filter
-        self.filter_with_geo_radius |= filter_with_geo_radius;
-        self.filter_with_geo_bounding_box |= filter_with_geo_bounding_box;
-        self.filter_sum_of_criteria_terms =
-            self.filter_sum_of_criteria_terms.saturating_add(filter_sum_of_criteria_terms);
-        self.filter_total_number_of_criteria =
-            self.filter_total_number_of_criteria.saturating_add(filter_total_number_of_criteria);
-        for (key, value) in used_syntax.into_iter() {
-            let used_syntax = self.used_syntax.entry(key).or_insert(0);
-            *used_syntax = used_syntax.saturating_add(value);
-        }
-
-        // attributes_to_search_on
-        self.attributes_to_search_on_total_number_of_uses = self
-            .attributes_to_search_on_total_number_of_uses
-            .saturating_add(attributes_to_search_on_total_number_of_uses);
-
-        // q
-        self.max_terms_number = self.max_terms_number.max(max_terms_number);
-
-        // vector
-        self.max_vector_size = self.max_vector_size.max(max_vector_size);
-        self.retrieve_vectors |= retrieve_vectors;
-        self.semantic_ratio |= semantic_ratio;
-        self.hybrid |= hybrid;
-
-        // pagination
-        self.max_limit = self.max_limit.max(max_limit);
-        self.max_offset = self.max_offset.max(max_offset);
-        self.finite_pagination += finite_pagination;
-
-        // formatting
-        self.max_attributes_to_retrieve =
-            self.max_attributes_to_retrieve.max(max_attributes_to_retrieve);
-        self.max_attributes_to_highlight =
-            self.max_attributes_to_highlight.max(max_attributes_to_highlight);
-        self.highlight_pre_tag |= highlight_pre_tag;
-        self.highlight_post_tag |= highlight_post_tag;
-        self.max_attributes_to_crop = self.max_attributes_to_crop.max(max_attributes_to_crop);
-        self.crop_marker |= crop_marker;
-        self.show_matches_position |= show_matches_position;
-        self.crop_length |= crop_length;
-
-        // facets
-        self.facets_sum_of_terms = self.facets_sum_of_terms.saturating_add(facets_sum_of_terms);
-        self.facets_total_number_of_facets =
-            self.facets_total_number_of_facets.saturating_add(facets_total_number_of_facets);
-
-        // matching strategy
-        for (key, value) in matching_strategy.into_iter() {
-            let matching_strategy = self.matching_strategy.entry(key).or_insert(0);
-            *matching_strategy = matching_strategy.saturating_add(value);
-        }
-
-        // scoring
-        self.show_ranking_score |= show_ranking_score;
-        self.show_ranking_score_details |= show_ranking_score_details;
-        self.ranking_score_threshold |= ranking_score_threshold;
-
-        // locales
-        self.locales.append(&mut locales);
-
-        self
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        let Self {
-            total_received,
-            total_succeeded,
-            time_spent,
-            sort_with_geo_point,
-            sort_sum_of_criteria_terms,
-            sort_total_number_of_criteria,
-            distinct,
-            filter_with_geo_radius,
-            filter_with_geo_bounding_box,
-            filter_sum_of_criteria_terms,
-            filter_total_number_of_criteria,
-            used_syntax,
-            attributes_to_search_on_total_number_of_uses,
-            max_terms_number,
-            max_vector_size,
-            retrieve_vectors,
-            matching_strategy,
-            max_limit,
-            max_offset,
-            finite_pagination,
-            max_attributes_to_retrieve,
-            max_attributes_to_highlight,
-            highlight_pre_tag,
-            highlight_post_tag,
-            max_attributes_to_crop,
-            crop_marker,
-            show_matches_position,
-            crop_length,
-            facets_sum_of_terms,
-            facets_total_number_of_facets,
-            show_ranking_score,
-            show_ranking_score_details,
-            semantic_ratio,
-            hybrid,
-            total_degraded,
-            total_used_negative_operator,
-            ranking_score_threshold,
-            locales,
-            marker: _,
-        } = *self;
-
-        // we get all the values in a sorted manner
-        let time_spent = time_spent.into_sorted_vec();
-        // the index of the 99th percentage of value
-        let percentile_99th = time_spent.len() * 99 / 100;
-        // We are only interested by the slowest value of the 99th fastest results
-        let time_spent = time_spent.get(percentile_99th);
-
-        json!({
-            "requests": {
-                "99th_response_time": time_spent.map(|t| format!("{:.2}", t)),
-                "total_succeeded": total_succeeded,
-                "total_failed": total_received.saturating_sub(total_succeeded), // just to be sure we never panics
-                "total_received": total_received,
-                "total_degraded": total_degraded,
-                "total_used_negative_operator": total_used_negative_operator,
-            },
-            "sort": {
-                "with_geoPoint": sort_with_geo_point,
-                "avg_criteria_number": format!("{:.2}", sort_sum_of_criteria_terms as f64 / sort_total_number_of_criteria as f64),
-            },
-            "distinct": distinct,
-            "filter": {
-               "with_geoRadius": filter_with_geo_radius,
-               "with_geoBoundingBox": filter_with_geo_bounding_box,
-               "avg_criteria_number": format!("{:.2}", filter_sum_of_criteria_terms as f64 / filter_total_number_of_criteria as f64),
-               "most_used_syntax": used_syntax.iter().max_by_key(|(_, v)| *v).map(|(k, _)| json!(k)).unwrap_or_else(|| json!(null)),
-            },
-            "attributes_to_search_on": {
-               "total_number_of_uses": attributes_to_search_on_total_number_of_uses,
-            },
-            "q": {
-               "max_terms_number": max_terms_number,
-            },
-            "vector": {
-                "max_vector_size": max_vector_size,
-                "retrieve_vectors": retrieve_vectors,
-            },
-            "hybrid": {
-                "enabled": hybrid,
-                "semantic_ratio": semantic_ratio,
-            },
-            "pagination": {
-               "max_limit": max_limit,
-               "max_offset": max_offset,
-               "most_used_navigation": if finite_pagination > (total_received / 2) { "exhaustive" } else { "estimated" },
-            },
-            "formatting": {
-                "max_attributes_to_retrieve": max_attributes_to_retrieve,
-                "max_attributes_to_highlight": max_attributes_to_highlight,
-                "highlight_pre_tag": highlight_pre_tag,
-                "highlight_post_tag": highlight_post_tag,
-                "max_attributes_to_crop": max_attributes_to_crop,
-                "crop_marker": crop_marker,
-                "show_matches_position": show_matches_position,
-                "crop_length": crop_length,
-            },
-            "facets": {
-                "avg_facets_number": format!("{:.2}", facets_sum_of_terms as f64 / facets_total_number_of_facets as f64),
-            },
-            "matching_strategy": {
-                "most_used_strategy": matching_strategy.iter().max_by_key(|(_, v)| *v).map(|(k, _)| json!(k)).unwrap_or_else(|| json!(null)),
-            },
-            "locales": locales,
-            "scoring": {
-                "show_ranking_score": show_ranking_score,
-                "show_ranking_score_details": show_ranking_score_details,
-                "ranking_score_threshold": ranking_score_threshold,
-            },
-        })
-    }
-}
--- a/meilisearch/src/routes/indexes/settings.rs
+++ b/meilisearch/src/routes/indexes/settings.rs
@@ -1,14 +1,15 @@
-use super::settings_analytics::*;
 use actix_web::web::Data;
 use actix_web::{web, HttpRequest, HttpResponse};
 use deserr::actix_web::AwebJson;
 use index_scheduler::IndexScheduler;
 use meilisearch_types::deserr::DeserrJsonError;
 use meilisearch_types::error::ResponseError;
+use meilisearch_types::facet_values_sort::FacetValuesSort;
 use meilisearch_types::index_uid::IndexUid;
 use meilisearch_types::milli::update::Setting;
-use meilisearch_types::settings::{settings, SecretPolicy, Settings, Unchecked};
+use meilisearch_types::settings::{settings, RankingRuleView, SecretPolicy, Settings, Unchecked};
 use meilisearch_types::tasks::KindWithContent;
+use serde_json::json;
 use tracing::debug;

 use crate::analytics::Analytics;
@@ -19,7 +20,7 @@ use crate::Opt;

 #[macro_export]
 macro_rules! make_setting_route {
-    ($route:literal, $update_verb:ident, $type:ty, $err_ty:ty, $attr:ident, $camelcase_attr:literal, $analytics:ident) => {
+    ($route:literal, $update_verb:ident, $type:ty, $err_ty:ty, $attr:ident, $camelcase_attr:literal, $analytics_var:ident, $analytics:expr) => {
        pub mod $attr {
            use actix_web::web::Data;
            use actix_web::{web, HttpRequest, HttpResponse, Resource};
@@ -79,7 +80,7 @@ macro_rules! make_setting_route {
                body: deserr::actix_web::AwebJson<Option<$type>, $err_ty>,
                req: HttpRequest,
                opt: web::Data<Opt>,
-                analytics: web::Data<Analytics>,
+                $analytics_var: web::Data<dyn Analytics>,
            ) -> std::result::Result<HttpResponse, ResponseError> {
                let index_uid = IndexUid::try_from(index_uid.into_inner())?;

@@ -87,10 +88,7 @@ macro_rules! make_setting_route {
                debug!(parameters = ?body, "Update settings");

                #[allow(clippy::redundant_closure_call)]
-                analytics.publish(
-                    $crate::routes::indexes::settings_analytics::$analytics::new(body.as_ref()).into_settings(),
-                    &req,
-                );
+                $analytics(&body, &req);

                let new_settings = Settings {
                    $attr: match body {
@@ -162,7 +160,21 @@ make_setting_route!(
    >,
    filterable_attributes,
    "filterableAttributes",
-    FilterableAttributesAnalytics
+    analytics,
+    |setting: &Option<std::collections::BTreeSet<String>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "FilterableAttributes Updated".to_string(),
+            json!({
+                "filterable_attributes": {
+                    "total": setting.as_ref().map(|filter| filter.len()).unwrap_or(0),
+                    "has_geo": setting.as_ref().map(|filter| filter.contains("_geo")).unwrap_or(false),
+                }
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -174,7 +186,21 @@ make_setting_route!(
    >,
    sortable_attributes,
    "sortableAttributes",
-    SortableAttributesAnalytics
+    analytics,
+    |setting: &Option<std::collections::BTreeSet<String>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "SortableAttributes Updated".to_string(),
+            json!({
+                "sortable_attributes": {
+                    "total": setting.as_ref().map(|sort| sort.len()),
+                    "has_geo": setting.as_ref().map(|sort| sort.contains("_geo")),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -186,7 +212,21 @@ make_setting_route!(
    >,
    displayed_attributes,
    "displayedAttributes",
-    DisplayedAttributesAnalytics
+    analytics,
+    |displayed: &Option<Vec<String>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "DisplayedAttributes Updated".to_string(),
+            json!({
+                "displayed_attributes": {
+                    "total": displayed.as_ref().map(|displayed| displayed.len()),
+                    "with_wildcard": displayed.as_ref().map(|displayed| displayed.iter().any(|displayed| displayed == "*")),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -198,7 +238,40 @@ make_setting_route!(
    >,
    typo_tolerance,
    "typoTolerance",
-    TypoToleranceAnalytics
+    analytics,
+    |setting: &Option<meilisearch_types::settings::TypoSettings>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "TypoTolerance Updated".to_string(),
+            json!({
+                "typo_tolerance": {
+                    "enabled": setting.as_ref().map(|s| !matches!(s.enabled, Setting::Set(false))),
+                    "disable_on_attributes": setting
+                        .as_ref()
+                        .and_then(|s| s.disable_on_attributes.as_ref().set().map(|m| !m.is_empty())),
+                    "disable_on_words": setting
+                        .as_ref()
+                        .and_then(|s| s.disable_on_words.as_ref().set().map(|m| !m.is_empty())),
+                    "min_word_size_for_one_typo": setting
+                        .as_ref()
+                        .and_then(|s| s.min_word_size_for_typos
+                            .as_ref()
+                            .set()
+                            .map(|s| s.one_typo.set()))
+                        .flatten(),
+                    "min_word_size_for_two_typos": setting
+                        .as_ref()
+                        .and_then(|s| s.min_word_size_for_typos
+                            .as_ref()
+                            .set()
+                            .map(|s| s.two_typos.set()))
+                        .flatten(),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -210,7 +283,21 @@ make_setting_route!(
    >,
    searchable_attributes,
    "searchableAttributes",
-    SearchableAttributesAnalytics
+    analytics,
+    |setting: &Option<Vec<String>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "SearchableAttributes Updated".to_string(),
+            json!({
+                "searchable_attributes": {
+                    "total": setting.as_ref().map(|searchable| searchable.len()),
+                    "with_wildcard": setting.as_ref().map(|searchable| searchable.iter().any(|searchable| searchable == "*")),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -222,7 +309,20 @@ make_setting_route!(
    >,
    stop_words,
    "stopWords",
-    StopWordsAnalytics
+    analytics,
+    |stop_words: &Option<std::collections::BTreeSet<String>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "StopWords Updated".to_string(),
+            json!({
+                "stop_words": {
+                    "total": stop_words.as_ref().map(|stop_words| stop_words.len()),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -234,7 +334,20 @@ make_setting_route!(
    >,
    non_separator_tokens,
    "nonSeparatorTokens",
-    NonSeparatorTokensAnalytics
+    analytics,
+    |non_separator_tokens: &Option<std::collections::BTreeSet<String>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "nonSeparatorTokens Updated".to_string(),
+            json!({
+                "non_separator_tokens": {
+                    "total": non_separator_tokens.as_ref().map(|non_separator_tokens| non_separator_tokens.len()),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -246,7 +359,20 @@ make_setting_route!(
    >,
    separator_tokens,
    "separatorTokens",
-    SeparatorTokensAnalytics
+    analytics,
+    |separator_tokens: &Option<std::collections::BTreeSet<String>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "separatorTokens Updated".to_string(),
+            json!({
+                "separator_tokens": {
+                    "total": separator_tokens.as_ref().map(|separator_tokens| separator_tokens.len()),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -258,7 +384,20 @@ make_setting_route!(
    >,
    dictionary,
    "dictionary",
-    DictionaryAnalytics
+    analytics,
+    |dictionary: &Option<std::collections::BTreeSet<String>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "dictionary Updated".to_string(),
+            json!({
+                "dictionary": {
+                    "total": dictionary.as_ref().map(|dictionary| dictionary.len()),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -270,7 +409,20 @@ make_setting_route!(
    >,
    synonyms,
    "synonyms",
-    SynonymsAnalytics
+    analytics,
+    |synonyms: &Option<std::collections::BTreeMap<String, Vec<String>>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "Synonyms Updated".to_string(),
+            json!({
+                "synonyms": {
+                    "total": synonyms.as_ref().map(|synonyms| synonyms.len()),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -282,7 +434,19 @@ make_setting_route!(
    >,
    distinct_attribute,
    "distinctAttribute",
-    DistinctAttributeAnalytics
+    analytics,
+    |distinct: &Option<String>, req: &HttpRequest| {
+        use serde_json::json;
+        analytics.publish(
+            "DistinctAttribute Updated".to_string(),
+            json!({
+                "distinct_attribute": {
+                    "set": distinct.is_some(),
+                }
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -294,7 +458,20 @@ make_setting_route!(
    >,
    proximity_precision,
    "proximityPrecision",
-    ProximityPrecisionAnalytics
+    analytics,
+    |precision: &Option<meilisearch_types::settings::ProximityPrecisionView>, req: &HttpRequest| {
+        use serde_json::json;
+        analytics.publish(
+            "ProximityPrecision Updated".to_string(),
+            json!({
+                "proximity_precision": {
+                    "set": precision.is_some(),
+                    "value": precision.unwrap_or_default(),
+                }
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -306,7 +483,17 @@ make_setting_route!(
    >,
    localized_attributes,
    "localizedAttributes",
-    LocalesAnalytics
+    analytics,
+    |rules: &Option<Vec<meilisearch_types::locales::LocalizedAttributesRuleView>>, req: &HttpRequest| {
+        use serde_json::json;
+        analytics.publish(
+            "LocalizedAttributesRules Updated".to_string(),
+            json!({
+                "locales": rules.as_ref().map(|rules| rules.iter().flat_map(|rule| rule.locales.iter().cloned()).collect::<std::collections::BTreeSet<_>>())
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -318,7 +505,26 @@ make_setting_route!(
    >,
    ranking_rules,
    "rankingRules",
-    RankingRulesAnalytics
+    analytics,
+    |setting: &Option<Vec<meilisearch_types::settings::RankingRuleView>>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "RankingRules Updated".to_string(),
+            json!({
+                "ranking_rules": {
+                    "words_position": setting.as_ref().map(|rr| rr.iter().position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Words))),
+                    "typo_position": setting.as_ref().map(|rr| rr.iter().position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Typo))),
+                    "proximity_position": setting.as_ref().map(|rr| rr.iter().position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Proximity))),
+                    "attribute_position": setting.as_ref().map(|rr| rr.iter().position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Attribute))),
+                    "sort_position": setting.as_ref().map(|rr| rr.iter().position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Sort))),
+                    "exactness_position": setting.as_ref().map(|rr| rr.iter().position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Exactness))),
+                    "values": setting.as_ref().map(|rr| rr.iter().filter(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Asc(_) | meilisearch_types::settings::RankingRuleView::Desc(_)) ).map(|x| x.to_string()).collect::<Vec<_>>().join(", ")),
+                }
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -330,7 +536,25 @@ make_setting_route!(
    >,
    faceting,
    "faceting",
-    FacetingAnalytics
+    analytics,
+    |setting: &Option<meilisearch_types::settings::FacetingSettings>, req: &HttpRequest| {
+        use serde_json::json;
+        use meilisearch_types::facet_values_sort::FacetValuesSort;
+
+        analytics.publish(
+            "Faceting Updated".to_string(),
+            json!({
+                "faceting": {
+                    "max_values_per_facet": setting.as_ref().and_then(|s| s.max_values_per_facet.set()),
+                    "sort_facet_values_by_star_count": setting.as_ref().and_then(|s| {
+                        s.sort_facet_values_by.as_ref().set().map(|s| s.iter().any(|(k, v)| k == "*" && v == &FacetValuesSort::Count))
+                    }),
+                    "sort_facet_values_by_total": setting.as_ref().and_then(|s| s.sort_facet_values_by.as_ref().set().map(|s| s.len())),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -342,7 +566,20 @@ make_setting_route!(
    >,
    pagination,
    "pagination",
-    PaginationAnalytics
+    analytics,
+    |setting: &Option<meilisearch_types::settings::PaginationSettings>, req: &HttpRequest| {
+        use serde_json::json;
+
+        analytics.publish(
+            "Pagination Updated".to_string(),
+            json!({
+                "pagination": {
+                    "max_total_hits": setting.as_ref().and_then(|s| s.max_total_hits.set()),
+                },
+            }),
+            Some(req),
+        );
+    }
 );

 make_setting_route!(
@@ -354,9 +591,75 @@ make_setting_route!(
    >,
    embedders,
    "embedders",
-    EmbeddersAnalytics
+    analytics,
+    |setting: &Option<std::collections::BTreeMap<String, Setting<meilisearch_types::milli::vector::settings::EmbeddingSettings>>>, req: &HttpRequest| {
+
+
+        analytics.publish(
+            "Embedders Updated".to_string(),
+            serde_json::json!({"embedders": crate::routes::indexes::settings::embedder_analytics(setting.as_ref())}),
+            Some(req),
+        );
+    }
 );

+fn embedder_analytics(
+    setting: Option<
+        &std::collections::BTreeMap<
+            String,
+            Setting<meilisearch_types::milli::vector::settings::EmbeddingSettings>,
+        >,
+    >,
+) -> serde_json::Value {
+    let mut sources = std::collections::HashSet::new();
+
+    if let Some(s) = &setting {
+        for source in s
+            .values()
+            .filter_map(|config| config.clone().set())
+            .filter_map(|config| config.source.set())
+        {
+            use meilisearch_types::milli::vector::settings::EmbedderSource;
+            match source {
+                EmbedderSource::OpenAi => sources.insert("openAi"),
+                EmbedderSource::HuggingFace => sources.insert("huggingFace"),
+                EmbedderSource::UserProvided => sources.insert("userProvided"),
+                EmbedderSource::Ollama => sources.insert("ollama"),
+                EmbedderSource::Rest => sources.insert("rest"),
+            };
+        }
+    };
+
+    let document_template_used = setting.as_ref().map(|map| {
+        map.values()
+            .filter_map(|config| config.clone().set())
+            .any(|config| config.document_template.set().is_some())
+    });
+
+    let document_template_max_bytes = setting.as_ref().and_then(|map| {
+        map.values()
+            .filter_map(|config| config.clone().set())
+            .filter_map(|config| config.document_template_max_bytes.set())
+            .max()
+    });
+
+    let binary_quantization_used = setting.as_ref().map(|map| {
+        map.values()
+            .filter_map(|config| config.clone().set())
+            .any(|config| config.binary_quantized.set().is_some())
+    });
+
+    json!(
+        {
+            "total": setting.as_ref().map(|s| s.len()),
+            "sources": sources,
+            "document_template_used": document_template_used,
+            "document_template_max_bytes": document_template_max_bytes,
+            "binary_quantization_used": binary_quantization_used,
+        }
+    )
+}
+
 make_setting_route!(
    "/search-cutoff-ms",
    put,
@@ -366,7 +669,14 @@ make_setting_route!(
    >,
    search_cutoff_ms,
    "searchCutoffMs",
-    SearchCutoffMsAnalytics
+    analytics,
+    |setting: &Option<u64>, req: &HttpRequest| {
+        analytics.publish(
+            "Search Cutoff Updated".to_string(),
+            serde_json::json!({"search_cutoff_ms": setting }),
+            Some(req),
+        );
+    }
 );

 macro_rules! generate_configure {
@@ -410,7 +720,7 @@ pub async fn update_all(
    body: AwebJson<Settings<Unchecked>, DeserrJsonError>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;

@@ -419,45 +729,104 @@ pub async fn update_all(
    let new_settings = validate_settings(new_settings, &index_scheduler)?;

    analytics.publish(
-        SettingsAnalytics {
-            ranking_rules: RankingRulesAnalytics::new(new_settings.ranking_rules.as_ref().set()),
-            searchable_attributes: SearchableAttributesAnalytics::new(
-                new_settings.searchable_attributes.as_ref().set(),
-            ),
-            displayed_attributes: DisplayedAttributesAnalytics::new(
-                new_settings.displayed_attributes.as_ref().set(),
-            ),
-            sortable_attributes: SortableAttributesAnalytics::new(
-                new_settings.sortable_attributes.as_ref().set(),
-            ),
-            filterable_attributes: FilterableAttributesAnalytics::new(
-                new_settings.filterable_attributes.as_ref().set(),
-            ),
-            distinct_attribute: DistinctAttributeAnalytics::new(
-                new_settings.distinct_attribute.as_ref().set(),
-            ),
-            proximity_precision: ProximityPrecisionAnalytics::new(
-                new_settings.proximity_precision.as_ref().set(),
-            ),
-            typo_tolerance: TypoToleranceAnalytics::new(new_settings.typo_tolerance.as_ref().set()),
-            faceting: FacetingAnalytics::new(new_settings.faceting.as_ref().set()),
-            pagination: PaginationAnalytics::new(new_settings.pagination.as_ref().set()),
-            stop_words: StopWordsAnalytics::new(new_settings.stop_words.as_ref().set()),
-            synonyms: SynonymsAnalytics::new(new_settings.synonyms.as_ref().set()),
-            embedders: EmbeddersAnalytics::new(new_settings.embedders.as_ref().set()),
-            search_cutoff_ms: SearchCutoffMsAnalytics::new(
-                new_settings.search_cutoff_ms.as_ref().set(),
-            ),
-            locales: LocalesAnalytics::new(new_settings.localized_attributes.as_ref().set()),
-            dictionary: DictionaryAnalytics::new(new_settings.dictionary.as_ref().set()),
-            separator_tokens: SeparatorTokensAnalytics::new(
-                new_settings.separator_tokens.as_ref().set(),
-            ),
-            non_separator_tokens: NonSeparatorTokensAnalytics::new(
-                new_settings.non_separator_tokens.as_ref().set(),
-            ),
-        },
-        &req,
+        "Settings Updated".to_string(),
+        json!({
+           "ranking_rules": {
+                "words_position": new_settings.ranking_rules.as_ref().set().map(|rr| rr.iter().position(|s| matches!(s, RankingRuleView::Words))),
+                "typo_position": new_settings.ranking_rules.as_ref().set().map(|rr| rr.iter().position(|s| matches!(s, RankingRuleView::Typo))),
+                "proximity_position": new_settings.ranking_rules.as_ref().set().map(|rr| rr.iter().position(|s| matches!(s, RankingRuleView::Proximity))),
+                "attribute_position": new_settings.ranking_rules.as_ref().set().map(|rr| rr.iter().position(|s| matches!(s, RankingRuleView::Attribute))),
+                "sort_position": new_settings.ranking_rules.as_ref().set().map(|rr| rr.iter().position(|s| matches!(s, RankingRuleView::Sort))),
+                "exactness_position": new_settings.ranking_rules.as_ref().set().map(|rr| rr.iter().position(|s| matches!(s, RankingRuleView::Exactness))),
+                "values": new_settings.ranking_rules.as_ref().set().map(|rr| rr.iter().filter(|s| !matches!(s, RankingRuleView::Asc(_) | RankingRuleView::Desc(_)) ).map(|x| x.to_string()).collect::<Vec<_>>().join(", ")),
+            },
+            "searchable_attributes": {
+                "total": new_settings.searchable_attributes.as_ref().set().map(|searchable| searchable.len()),
+                "with_wildcard": new_settings.searchable_attributes.as_ref().set().map(|searchable| searchable.iter().any(|searchable| searchable == "*")),
+            },
+            "displayed_attributes": {
+                "total": new_settings.displayed_attributes.as_ref().set().map(|displayed| displayed.len()),
+                "with_wildcard": new_settings.displayed_attributes.as_ref().set().map(|displayed| displayed.iter().any(|displayed| displayed == "*")),
+            },
+           "sortable_attributes": {
+                "total": new_settings.sortable_attributes.as_ref().set().map(|sort| sort.len()),
+                "has_geo": new_settings.sortable_attributes.as_ref().set().map(|sort| sort.iter().any(|s| s == "_geo")),
+            },
+           "filterable_attributes": {
+                "total": new_settings.filterable_attributes.as_ref().set().map(|filter| filter.len()),
+                "has_geo": new_settings.filterable_attributes.as_ref().set().map(|filter| filter.iter().any(|s| s == "_geo")),
+            },
+            "distinct_attribute": {
+                "set": new_settings.distinct_attribute.as_ref().set().is_some()
+            },
+            "proximity_precision": {
+                "set": new_settings.proximity_precision.as_ref().set().is_some(),
+                "value": new_settings.proximity_precision.as_ref().set().copied().unwrap_or_default()
+            },
+            "typo_tolerance": {
+                "enabled": new_settings.typo_tolerance
+                    .as_ref()
+                    .set()
+                    .and_then(|s| s.enabled.as_ref().set())
+                    .copied(),
+                "disable_on_attributes": new_settings.typo_tolerance
+                    .as_ref()
+                    .set()
+                    .and_then(|s| s.disable_on_attributes.as_ref().set().map(|m| !m.is_empty())),
+                "disable_on_words": new_settings.typo_tolerance
+                    .as_ref()
+                    .set()
+                    .and_then(|s| s.disable_on_words.as_ref().set().map(|m| !m.is_empty())),
+                "min_word_size_for_one_typo": new_settings.typo_tolerance
+                    .as_ref()
+                    .set()
+                    .and_then(|s| s.min_word_size_for_typos
+                        .as_ref()
+                        .set()
+                        .map(|s| s.one_typo.set()))
+                    .flatten(),
+                "min_word_size_for_two_typos": new_settings.typo_tolerance
+                    .as_ref()
+                    .set()
+                    .and_then(|s| s.min_word_size_for_typos
+                        .as_ref()
+                        .set()
+                        .map(|s| s.two_typos.set()))
+                    .flatten(),
+            },
+            "faceting": {
+                "max_values_per_facet": new_settings.faceting
+                    .as_ref()
+                    .set()
+                    .and_then(|s| s.max_values_per_facet.as_ref().set()),
+                "sort_facet_values_by_star_count": new_settings.faceting
+                    .as_ref()
+                    .set()
+                    .and_then(|s| {
+                        s.sort_facet_values_by.as_ref().set().map(|s| s.iter().any(|(k, v)| k == "*" && v == &FacetValuesSort::Count))
+                    }),
+                "sort_facet_values_by_total": new_settings.faceting
+                    .as_ref()
+                    .set()
+                    .and_then(|s| s.sort_facet_values_by.as_ref().set().map(|s| s.len())),
+            },
+            "pagination": {
+                "max_total_hits": new_settings.pagination
+                    .as_ref()
+                    .set()
+                    .and_then(|s| s.max_total_hits.as_ref().set()),
+            },
+            "stop_words": {
+                "total": new_settings.stop_words.as_ref().set().map(|stop_words| stop_words.len()),
+            },
+            "synonyms": {
+                "total": new_settings.synonyms.as_ref().set().map(|synonyms| synonyms.len()),
+            },
+            "embedders": crate::routes::indexes::settings::embedder_analytics(new_settings.embedders.as_ref().set()),
+            "search_cutoff_ms": new_settings.search_cutoff_ms.as_ref().set(),
+            "locales": new_settings.localized_attributes.as_ref().set().map(|rules| rules.iter().flat_map(|rule| rule.locales.iter().cloned()).collect::<std::collections::BTreeSet<_>>()),
+        }),
+        Some(&req),
    );

    let allow_index_creation = index_scheduler.filters().allow_index_creation(&index_uid);
--- a/meilisearch/src/routes/indexes/settings_analytics.rs
+++ b/meilisearch/src/routes/indexes/settings_analytics.rs
@@ -1,621 +0,0 @@
-//! All the structures used to make the analytics on the settings works.
-//! The signatures of the `new` functions are not very rust idiomatic because they must match the types received
-//! through the sub-settings route directly without any manipulation.
-//! This is why we often use a `Option<&Vec<_>>` instead of a `Option<&[_]>`.
-
-use meilisearch_types::locales::{Locale, LocalizedAttributesRuleView};
-use meilisearch_types::milli::update::Setting;
-use meilisearch_types::milli::vector::settings::EmbeddingSettings;
-use meilisearch_types::settings::{
-    FacetingSettings, PaginationSettings, ProximityPrecisionView, TypoSettings,
-};
-use meilisearch_types::{facet_values_sort::FacetValuesSort, settings::RankingRuleView};
-use serde::Serialize;
-use std::collections::{BTreeMap, BTreeSet, HashSet};
-
-use crate::analytics::Aggregate;
-
-#[derive(Serialize, Default)]
-pub struct SettingsAnalytics {
-    pub ranking_rules: RankingRulesAnalytics,
-    pub searchable_attributes: SearchableAttributesAnalytics,
-    pub displayed_attributes: DisplayedAttributesAnalytics,
-    pub sortable_attributes: SortableAttributesAnalytics,
-    pub filterable_attributes: FilterableAttributesAnalytics,
-    pub distinct_attribute: DistinctAttributeAnalytics,
-    pub proximity_precision: ProximityPrecisionAnalytics,
-    pub typo_tolerance: TypoToleranceAnalytics,
-    pub faceting: FacetingAnalytics,
-    pub pagination: PaginationAnalytics,
-    pub stop_words: StopWordsAnalytics,
-    pub synonyms: SynonymsAnalytics,
-    pub embedders: EmbeddersAnalytics,
-    pub search_cutoff_ms: SearchCutoffMsAnalytics,
-    pub locales: LocalesAnalytics,
-    pub dictionary: DictionaryAnalytics,
-    pub separator_tokens: SeparatorTokensAnalytics,
-    pub non_separator_tokens: NonSeparatorTokensAnalytics,
-}
-
-impl Aggregate for SettingsAnalytics {
-    fn event_name(&self) -> &'static str {
-        "Settings Updated"
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self {
-            ranking_rules: RankingRulesAnalytics {
-                words_position: new
-                    .ranking_rules
-                    .words_position
-                    .or(self.ranking_rules.words_position),
-                typo_position: new.ranking_rules.typo_position.or(self.ranking_rules.typo_position),
-                proximity_position: new
-                    .ranking_rules
-                    .proximity_position
-                    .or(self.ranking_rules.proximity_position),
-                attribute_position: new
-                    .ranking_rules
-                    .attribute_position
-                    .or(self.ranking_rules.attribute_position),
-                sort_position: new.ranking_rules.sort_position.or(self.ranking_rules.sort_position),
-                exactness_position: new
-                    .ranking_rules
-                    .exactness_position
-                    .or(self.ranking_rules.exactness_position),
-                values: new.ranking_rules.values.or(self.ranking_rules.values),
-            },
-            searchable_attributes: SearchableAttributesAnalytics {
-                total: new.searchable_attributes.total.or(self.searchable_attributes.total),
-                with_wildcard: new
-                    .searchable_attributes
-                    .with_wildcard
-                    .or(self.searchable_attributes.with_wildcard),
-            },
-            displayed_attributes: DisplayedAttributesAnalytics {
-                total: new.displayed_attributes.total.or(self.displayed_attributes.total),
-                with_wildcard: new
-                    .displayed_attributes
-                    .with_wildcard
-                    .or(self.displayed_attributes.with_wildcard),
-            },
-            sortable_attributes: SortableAttributesAnalytics {
-                total: new.sortable_attributes.total.or(self.sortable_attributes.total),
-                has_geo: new.sortable_attributes.has_geo.or(self.sortable_attributes.has_geo),
-            },
-            filterable_attributes: FilterableAttributesAnalytics {
-                total: new.filterable_attributes.total.or(self.filterable_attributes.total),
-                has_geo: new.filterable_attributes.has_geo.or(self.filterable_attributes.has_geo),
-            },
-            distinct_attribute: DistinctAttributeAnalytics {
-                set: self.distinct_attribute.set | new.distinct_attribute.set,
-            },
-            proximity_precision: ProximityPrecisionAnalytics {
-                set: self.proximity_precision.set | new.proximity_precision.set,
-                value: new.proximity_precision.value.or(self.proximity_precision.value),
-            },
-            typo_tolerance: TypoToleranceAnalytics {
-                enabled: new.typo_tolerance.enabled.or(self.typo_tolerance.enabled),
-                disable_on_attributes: new
-                    .typo_tolerance
-                    .disable_on_attributes
-                    .or(self.typo_tolerance.disable_on_attributes),
-                disable_on_words: new
-                    .typo_tolerance
-                    .disable_on_words
-                    .or(self.typo_tolerance.disable_on_words),
-                min_word_size_for_one_typo: new
-                    .typo_tolerance
-                    .min_word_size_for_one_typo
-                    .or(self.typo_tolerance.min_word_size_for_one_typo),
-                min_word_size_for_two_typos: new
-                    .typo_tolerance
-                    .min_word_size_for_two_typos
-                    .or(self.typo_tolerance.min_word_size_for_two_typos),
-            },
-            faceting: FacetingAnalytics {
-                max_values_per_facet: new
-                    .faceting
-                    .max_values_per_facet
-                    .or(self.faceting.max_values_per_facet),
-                sort_facet_values_by_star_count: new
-                    .faceting
-                    .sort_facet_values_by_star_count
-                    .or(self.faceting.sort_facet_values_by_star_count),
-                sort_facet_values_by_total: new
-                    .faceting
-                    .sort_facet_values_by_total
-                    .or(self.faceting.sort_facet_values_by_total),
-            },
-            pagination: PaginationAnalytics {
-                max_total_hits: new.pagination.max_total_hits.or(self.pagination.max_total_hits),
-            },
-            stop_words: StopWordsAnalytics {
-                total: new.stop_words.total.or(self.stop_words.total),
-            },
-            synonyms: SynonymsAnalytics { total: new.synonyms.total.or(self.synonyms.total) },
-            embedders: EmbeddersAnalytics {
-                total: new.embedders.total.or(self.embedders.total),
-                sources: match (self.embedders.sources, new.embedders.sources) {
-                    (None, None) => None,
-                    (Some(sources), None) | (None, Some(sources)) => Some(sources),
-                    (Some(this), Some(other)) => Some(this.union(&other).cloned().collect()),
-                },
-                document_template_used: match (
-                    self.embedders.document_template_used,
-                    new.embedders.document_template_used,
-                ) {
-                    (None, None) => None,
-                    (Some(used), None) | (None, Some(used)) => Some(used),
-                    (Some(this), Some(other)) => Some(this | other),
-                },
-                document_template_max_bytes: match (
-                    self.embedders.document_template_max_bytes,
-                    new.embedders.document_template_max_bytes,
-                ) {
-                    (None, None) => None,
-                    (Some(bytes), None) | (None, Some(bytes)) => Some(bytes),
-                    (Some(this), Some(other)) => Some(this.max(other)),
-                },
-                binary_quantization_used: match (
-                    self.embedders.binary_quantization_used,
-                    new.embedders.binary_quantization_used,
-                ) {
-                    (None, None) => None,
-                    (Some(bq), None) | (None, Some(bq)) => Some(bq),
-                    (Some(this), Some(other)) => Some(this | other),
-                },
-            },
-            search_cutoff_ms: SearchCutoffMsAnalytics {
-                search_cutoff_ms: new
-                    .search_cutoff_ms
-                    .search_cutoff_ms
-                    .or(self.search_cutoff_ms.search_cutoff_ms),
-            },
-            locales: LocalesAnalytics { locales: new.locales.locales.or(self.locales.locales) },
-            dictionary: DictionaryAnalytics {
-                total: new.dictionary.total.or(self.dictionary.total),
-            },
-            separator_tokens: SeparatorTokensAnalytics {
-                total: new.non_separator_tokens.total.or(self.separator_tokens.total),
-            },
-            non_separator_tokens: NonSeparatorTokensAnalytics {
-                total: new.non_separator_tokens.total.or(self.non_separator_tokens.total),
-            },
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct RankingRulesAnalytics {
-    pub words_position: Option<usize>,
-    pub typo_position: Option<usize>,
-    pub proximity_position: Option<usize>,
-    pub attribute_position: Option<usize>,
-    pub sort_position: Option<usize>,
-    pub exactness_position: Option<usize>,
-    pub values: Option<String>,
-}
-
-impl RankingRulesAnalytics {
-    pub fn new(rr: Option<&Vec<RankingRuleView>>) -> Self {
-        RankingRulesAnalytics {
-            words_position: rr.as_ref().and_then(|rr| {
-                rr.iter()
-                    .position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Words))
-            }),
-            typo_position: rr.as_ref().and_then(|rr| {
-                rr.iter()
-                    .position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Typo))
-            }),
-            proximity_position: rr.as_ref().and_then(|rr| {
-                rr.iter().position(|s| {
-                    matches!(s, meilisearch_types::settings::RankingRuleView::Proximity)
-                })
-            }),
-            attribute_position: rr.as_ref().and_then(|rr| {
-                rr.iter().position(|s| {
-                    matches!(s, meilisearch_types::settings::RankingRuleView::Attribute)
-                })
-            }),
-            sort_position: rr.as_ref().and_then(|rr| {
-                rr.iter()
-                    .position(|s| matches!(s, meilisearch_types::settings::RankingRuleView::Sort))
-            }),
-            exactness_position: rr.as_ref().and_then(|rr| {
-                rr.iter().position(|s| {
-                    matches!(s, meilisearch_types::settings::RankingRuleView::Exactness)
-                })
-            }),
-            values: rr.as_ref().map(|rr| {
-                rr.iter()
-                    .filter(|s| {
-                        matches!(
-                            s,
-                            meilisearch_types::settings::RankingRuleView::Asc(_)
-                                | meilisearch_types::settings::RankingRuleView::Desc(_)
-                        )
-                    })
-                    .map(|x| x.to_string())
-                    .collect::<Vec<_>>()
-                    .join(", ")
-            }),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { ranking_rules: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct SearchableAttributesAnalytics {
-    pub total: Option<usize>,
-    pub with_wildcard: Option<bool>,
-}
-
-impl SearchableAttributesAnalytics {
-    pub fn new(setting: Option<&Vec<String>>) -> Self {
-        Self {
-            total: setting.as_ref().map(|searchable| searchable.len()),
-            with_wildcard: setting
-                .as_ref()
-                .map(|searchable| searchable.iter().any(|searchable| searchable == "*")),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { searchable_attributes: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct DisplayedAttributesAnalytics {
-    pub total: Option<usize>,
-    pub with_wildcard: Option<bool>,
-}
-
-impl DisplayedAttributesAnalytics {
-    pub fn new(displayed: Option<&Vec<String>>) -> Self {
-        Self {
-            total: displayed.as_ref().map(|displayed| displayed.len()),
-            with_wildcard: displayed
-                .as_ref()
-                .map(|displayed| displayed.iter().any(|displayed| displayed == "*")),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { displayed_attributes: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct SortableAttributesAnalytics {
-    pub total: Option<usize>,
-    pub has_geo: Option<bool>,
-}
-
-impl SortableAttributesAnalytics {
-    pub fn new(setting: Option<&BTreeSet<String>>) -> Self {
-        Self {
-            total: setting.as_ref().map(|sort| sort.len()),
-            has_geo: setting.as_ref().map(|sort| sort.contains("_geo")),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { sortable_attributes: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct FilterableAttributesAnalytics {
-    pub total: Option<usize>,
-    pub has_geo: Option<bool>,
-}
-
-impl FilterableAttributesAnalytics {
-    pub fn new(setting: Option<&BTreeSet<String>>) -> Self {
-        Self {
-            total: setting.as_ref().map(|filter| filter.len()),
-            has_geo: setting.as_ref().map(|filter| filter.contains("_geo")),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { filterable_attributes: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct DistinctAttributeAnalytics {
-    pub set: bool,
-}
-
-impl DistinctAttributeAnalytics {
-    pub fn new(distinct: Option<&String>) -> Self {
-        Self { set: distinct.is_some() }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { distinct_attribute: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct ProximityPrecisionAnalytics {
-    pub set: bool,
-    pub value: Option<ProximityPrecisionView>,
-}
-
-impl ProximityPrecisionAnalytics {
-    pub fn new(precision: Option<&ProximityPrecisionView>) -> Self {
-        Self { set: precision.is_some(), value: precision.cloned() }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { proximity_precision: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct TypoToleranceAnalytics {
-    pub enabled: Option<bool>,
-    pub disable_on_attributes: Option<bool>,
-    pub disable_on_words: Option<bool>,
-    pub min_word_size_for_one_typo: Option<u8>,
-    pub min_word_size_for_two_typos: Option<u8>,
-}
-
-impl TypoToleranceAnalytics {
-    pub fn new(setting: Option<&TypoSettings>) -> Self {
-        Self {
-            enabled: setting.as_ref().map(|s| !matches!(s.enabled, Setting::Set(false))),
-            disable_on_attributes: setting
-                .as_ref()
-                .and_then(|s| s.disable_on_attributes.as_ref().set().map(|m| !m.is_empty())),
-            disable_on_words: setting
-                .as_ref()
-                .and_then(|s| s.disable_on_words.as_ref().set().map(|m| !m.is_empty())),
-            min_word_size_for_one_typo: setting
-                .as_ref()
-                .and_then(|s| s.min_word_size_for_typos.as_ref().set().map(|s| s.one_typo.set()))
-                .flatten(),
-            min_word_size_for_two_typos: setting
-                .as_ref()
-                .and_then(|s| s.min_word_size_for_typos.as_ref().set().map(|s| s.two_typos.set()))
-                .flatten(),
-        }
-    }
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { typo_tolerance: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct FacetingAnalytics {
-    pub max_values_per_facet: Option<usize>,
-    pub sort_facet_values_by_star_count: Option<bool>,
-    pub sort_facet_values_by_total: Option<usize>,
-}
-
-impl FacetingAnalytics {
-    pub fn new(setting: Option<&FacetingSettings>) -> Self {
-        Self {
-            max_values_per_facet: setting.as_ref().and_then(|s| s.max_values_per_facet.set()),
-            sort_facet_values_by_star_count: setting.as_ref().and_then(|s| {
-                s.sort_facet_values_by
-                    .as_ref()
-                    .set()
-                    .map(|s| s.iter().any(|(k, v)| k == "*" && v == &FacetValuesSort::Count))
-            }),
-            sort_facet_values_by_total: setting
-                .as_ref()
-                .and_then(|s| s.sort_facet_values_by.as_ref().set().map(|s| s.len())),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { faceting: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct PaginationAnalytics {
-    pub max_total_hits: Option<usize>,
-}
-
-impl PaginationAnalytics {
-    pub fn new(setting: Option<&PaginationSettings>) -> Self {
-        Self { max_total_hits: setting.as_ref().and_then(|s| s.max_total_hits.set()) }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { pagination: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct StopWordsAnalytics {
-    pub total: Option<usize>,
-}
-
-impl StopWordsAnalytics {
-    pub fn new(stop_words: Option<&BTreeSet<String>>) -> Self {
-        Self { total: stop_words.as_ref().map(|stop_words| stop_words.len()) }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { stop_words: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct SynonymsAnalytics {
-    pub total: Option<usize>,
-}
-
-impl SynonymsAnalytics {
-    pub fn new(synonyms: Option<&BTreeMap<String, Vec<String>>>) -> Self {
-        Self { total: synonyms.as_ref().map(|synonyms| synonyms.len()) }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { synonyms: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct EmbeddersAnalytics {
-    // last
-    pub total: Option<usize>,
-    // Merge the sources
-    pub sources: Option<HashSet<String>>,
-    // |=
-    pub document_template_used: Option<bool>,
-    // max
-    pub document_template_max_bytes: Option<usize>,
-    // |=
-    pub binary_quantization_used: Option<bool>,
-}
-
-impl EmbeddersAnalytics {
-    pub fn new(setting: Option<&BTreeMap<String, Setting<EmbeddingSettings>>>) -> Self {
-        let mut sources = std::collections::HashSet::new();
-
-        if let Some(s) = &setting {
-            for source in s
-                .values()
-                .filter_map(|config| config.clone().set())
-                .filter_map(|config| config.source.set())
-            {
-                use meilisearch_types::milli::vector::settings::EmbedderSource;
-                match source {
-                    EmbedderSource::OpenAi => sources.insert("openAi".to_string()),
-                    EmbedderSource::HuggingFace => sources.insert("huggingFace".to_string()),
-                    EmbedderSource::UserProvided => sources.insert("userProvided".to_string()),
-                    EmbedderSource::Ollama => sources.insert("ollama".to_string()),
-                    EmbedderSource::Rest => sources.insert("rest".to_string()),
-                };
-            }
-        };
-
-        Self {
-            total: setting.as_ref().map(|s| s.len()),
-            sources: Some(sources),
-            document_template_used: setting.as_ref().map(|map| {
-                map.values()
-                    .filter_map(|config| config.clone().set())
-                    .any(|config| config.document_template.set().is_some())
-            }),
-            document_template_max_bytes: setting.as_ref().and_then(|map| {
-                map.values()
-                    .filter_map(|config| config.clone().set())
-                    .filter_map(|config| config.document_template_max_bytes.set())
-                    .max()
-            }),
-            binary_quantization_used: setting.as_ref().map(|map| {
-                map.values()
-                    .filter_map(|config| config.clone().set())
-                    .any(|config| config.binary_quantized.set().is_some())
-            }),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { embedders: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-#[serde(transparent)]
-pub struct SearchCutoffMsAnalytics {
-    pub search_cutoff_ms: Option<u64>,
-}
-
-impl SearchCutoffMsAnalytics {
-    pub fn new(setting: Option<&u64>) -> Self {
-        Self { search_cutoff_ms: setting.copied() }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { search_cutoff_ms: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-#[serde(transparent)]
-pub struct LocalesAnalytics {
-    pub locales: Option<BTreeSet<Locale>>,
-}
-
-impl LocalesAnalytics {
-    pub fn new(rules: Option<&Vec<LocalizedAttributesRuleView>>) -> Self {
-        LocalesAnalytics {
-            locales: rules.as_ref().map(|rules| {
-                rules
-                    .iter()
-                    .flat_map(|rule| rule.locales.iter().cloned())
-                    .collect::<std::collections::BTreeSet<_>>()
-            }),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { locales: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct DictionaryAnalytics {
-    pub total: Option<usize>,
-}
-
-impl DictionaryAnalytics {
-    pub fn new(dictionary: Option<&BTreeSet<String>>) -> Self {
-        Self { total: dictionary.as_ref().map(|dictionary| dictionary.len()) }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { dictionary: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct SeparatorTokensAnalytics {
-    pub total: Option<usize>,
-}
-
-impl SeparatorTokensAnalytics {
-    pub fn new(separator_tokens: Option<&BTreeSet<String>>) -> Self {
-        Self { total: separator_tokens.as_ref().map(|separator_tokens| separator_tokens.len()) }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { separator_tokens: self, ..Default::default() }
-    }
-}
-
-#[derive(Serialize, Default)]
-pub struct NonSeparatorTokensAnalytics {
-    pub total: Option<usize>,
-}
-
-impl NonSeparatorTokensAnalytics {
-    pub fn new(non_separator_tokens: Option<&BTreeSet<String>>) -> Self {
-        Self {
-            total: non_separator_tokens
-                .as_ref()
-                .map(|non_separator_tokens| non_separator_tokens.len()),
-        }
-    }
-
-    pub fn into_settings(self) -> SettingsAnalytics {
-        SettingsAnalytics { non_separator_tokens: self, ..Default::default() }
-    }
-}
--- a/meilisearch/src/routes/indexes/similar.rs
+++ b/meilisearch/src/routes/indexes/similar.rs
@@ -13,10 +13,9 @@ use serde_json::Value;
 use tracing::debug;

 use super::ActionPolicy;
-use crate::analytics::Analytics;
+use crate::analytics::{Analytics, SimilarAggregator};
 use crate::extractors::authentication::GuardedData;
 use crate::extractors::sequential_extractor::SeqHandler;
-use crate::routes::indexes::similar_analytics::{SimilarAggregator, SimilarGET, SimilarPOST};
 use crate::search::{
    add_search_rules, perform_similar, RankingScoreThresholdSimilar, RetrieveVectors, SearchKind,
    SimilarQuery, SimilarResult, DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET,
@@ -35,13 +34,13 @@ pub async fn similar_get(
    index_uid: web::Path<String>,
    params: AwebQueryParameter<SimilarQueryGet, DeserrQueryParamError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;

    let query = params.0.try_into()?;

-    let mut aggregate = SimilarAggregator::<SimilarGET>::from_query(&query);
+    let mut aggregate = SimilarAggregator::from_query(&query, &req);

    debug!(parameters = ?query, "Similar get");

@@ -50,7 +49,7 @@ pub async fn similar_get(
    if let Ok(similar) = &similar {
        aggregate.succeed(similar);
    }
-    analytics.publish(aggregate, &req);
+    analytics.get_similar(aggregate);

    let similar = similar?;

@@ -63,21 +62,21 @@ pub async fn similar_post(
    index_uid: web::Path<String>,
    params: AwebJson<SimilarQuery, DeserrJsonError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let index_uid = IndexUid::try_from(index_uid.into_inner())?;

    let query = params.into_inner();
    debug!(parameters = ?query, "Similar post");

-    let mut aggregate = SimilarAggregator::<SimilarPOST>::from_query(&query);
+    let mut aggregate = SimilarAggregator::from_query(&query, &req);

    let similar = similar(index_scheduler, index_uid, query).await;

    if let Ok(similar) = &similar {
        aggregate.succeed(similar);
    }
-    analytics.publish(aggregate, &req);
+    analytics.post_similar(aggregate);

    let similar = similar?;

--- a/meilisearch/src/routes/indexes/similar_analytics.rs
+++ b/meilisearch/src/routes/indexes/similar_analytics.rs
@@ -1,235 +0,0 @@
-use std::collections::{BinaryHeap, HashMap};
-
-use once_cell::sync::Lazy;
-use regex::Regex;
-use serde_json::{json, Value};
-
-use crate::{
-    aggregate_methods,
-    analytics::{Aggregate, AggregateMethod},
-    search::{SimilarQuery, SimilarResult},
-};
-
-aggregate_methods!(
-    SimilarPOST => "Similar POST",
-    SimilarGET => "Similar GET",
-);
-
-#[derive(Default)]
-pub struct SimilarAggregator<Method: AggregateMethod> {
-    // requests
-    total_received: usize,
-    total_succeeded: usize,
-    time_spent: BinaryHeap<usize>,
-
-    // filter
-    filter_with_geo_radius: bool,
-    filter_with_geo_bounding_box: bool,
-    // every time a request has a filter, this field must be incremented by the number of terms it contains
-    filter_sum_of_criteria_terms: usize,
-    // every time a request has a filter, this field must be incremented by one
-    filter_total_number_of_criteria: usize,
-    used_syntax: HashMap<String, usize>,
-
-    // Whether a non-default embedder was specified
-    retrieve_vectors: bool,
-
-    // pagination
-    max_limit: usize,
-    max_offset: usize,
-
-    // formatting
-    max_attributes_to_retrieve: usize,
-
-    // scoring
-    show_ranking_score: bool,
-    show_ranking_score_details: bool,
-    ranking_score_threshold: bool,
-
-    marker: std::marker::PhantomData<Method>,
-}
-
-impl<Method: AggregateMethod> SimilarAggregator<Method> {
-    #[allow(clippy::field_reassign_with_default)]
-    pub fn from_query(query: &SimilarQuery) -> Self {
-        let SimilarQuery {
-            id: _,
-            embedder: _,
-            offset,
-            limit,
-            attributes_to_retrieve: _,
-            retrieve_vectors,
-            show_ranking_score,
-            show_ranking_score_details,
-            filter,
-            ranking_score_threshold,
-        } = query;
-
-        let mut ret = Self::default();
-
-        ret.total_received = 1;
-
-        if let Some(ref filter) = filter {
-            static RE: Lazy<Regex> = Lazy::new(|| Regex::new("AND | OR").unwrap());
-            ret.filter_total_number_of_criteria = 1;
-
-            let syntax = match filter {
-                Value::String(_) => "string".to_string(),
-                Value::Array(values) => {
-                    if values.iter().map(|v| v.to_string()).any(|s| RE.is_match(&s)) {
-                        "mixed".to_string()
-                    } else {
-                        "array".to_string()
-                    }
-                }
-                _ => "none".to_string(),
-            };
-            // convert the string to a HashMap
-            ret.used_syntax.insert(syntax, 1);
-
-            let stringified_filters = filter.to_string();
-            ret.filter_with_geo_radius = stringified_filters.contains("_geoRadius(");
-            ret.filter_with_geo_bounding_box = stringified_filters.contains("_geoBoundingBox(");
-            ret.filter_sum_of_criteria_terms = RE.split(&stringified_filters).count();
-        }
-
-        ret.max_limit = *limit;
-        ret.max_offset = *offset;
-
-        ret.show_ranking_score = *show_ranking_score;
-        ret.show_ranking_score_details = *show_ranking_score_details;
-        ret.ranking_score_threshold = ranking_score_threshold.is_some();
-
-        ret.retrieve_vectors = *retrieve_vectors;
-
-        ret
-    }
-
-    pub fn succeed(&mut self, result: &SimilarResult) {
-        let SimilarResult { id: _, hits: _, processing_time_ms, hits_info: _ } = result;
-
-        self.total_succeeded = self.total_succeeded.saturating_add(1);
-
-        self.time_spent.push(*processing_time_ms as usize);
-    }
-}
-
-impl<Method: AggregateMethod> Aggregate for SimilarAggregator<Method> {
-    fn event_name(&self) -> &'static str {
-        Method::event_name()
-    }
-
-    /// Aggregate one [SimilarAggregator] into another.
-    fn aggregate(mut self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        let Self {
-            total_received,
-            total_succeeded,
-            mut time_spent,
-            filter_with_geo_radius,
-            filter_with_geo_bounding_box,
-            filter_sum_of_criteria_terms,
-            filter_total_number_of_criteria,
-            used_syntax,
-            max_limit,
-            max_offset,
-            max_attributes_to_retrieve,
-            show_ranking_score,
-            show_ranking_score_details,
-            ranking_score_threshold,
-            retrieve_vectors,
-            marker: _,
-        } = *new;
-
-        // request
-        self.total_received = self.total_received.saturating_add(total_received);
-        self.total_succeeded = self.total_succeeded.saturating_add(total_succeeded);
-        self.time_spent.append(&mut time_spent);
-
-        // filter
-        self.filter_with_geo_radius |= filter_with_geo_radius;
-        self.filter_with_geo_bounding_box |= filter_with_geo_bounding_box;
-        self.filter_sum_of_criteria_terms =
-            self.filter_sum_of_criteria_terms.saturating_add(filter_sum_of_criteria_terms);
-        self.filter_total_number_of_criteria =
-            self.filter_total_number_of_criteria.saturating_add(filter_total_number_of_criteria);
-        for (key, value) in used_syntax.into_iter() {
-            let used_syntax = self.used_syntax.entry(key).or_insert(0);
-            *used_syntax = used_syntax.saturating_add(value);
-        }
-
-        self.retrieve_vectors |= retrieve_vectors;
-
-        // pagination
-        self.max_limit = self.max_limit.max(max_limit);
-        self.max_offset = self.max_offset.max(max_offset);
-
-        // formatting
-        self.max_attributes_to_retrieve =
-            self.max_attributes_to_retrieve.max(max_attributes_to_retrieve);
-
-        // scoring
-        self.show_ranking_score |= show_ranking_score;
-        self.show_ranking_score_details |= show_ranking_score_details;
-        self.ranking_score_threshold |= ranking_score_threshold;
-
-        self
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        let Self {
-            total_received,
-            total_succeeded,
-            time_spent,
-            filter_with_geo_radius,
-            filter_with_geo_bounding_box,
-            filter_sum_of_criteria_terms,
-            filter_total_number_of_criteria,
-            used_syntax,
-            max_limit,
-            max_offset,
-            max_attributes_to_retrieve,
-            show_ranking_score,
-            show_ranking_score_details,
-            ranking_score_threshold,
-            retrieve_vectors,
-            marker: _,
-        } = *self;
-
-        // we get all the values in a sorted manner
-        let time_spent = time_spent.into_sorted_vec();
-        // the index of the 99th percentage of value
-        let percentile_99th = time_spent.len() * 99 / 100;
-        // We are only interested by the slowest value of the 99th fastest results
-        let time_spent = time_spent.get(percentile_99th);
-
-        json!({
-            "requests": {
-                "99th_response_time": time_spent.map(|t| format!("{:.2}", t)),
-                "total_succeeded": total_succeeded,
-                "total_failed": total_received.saturating_sub(total_succeeded), // just to be sure we never panics
-                "total_received": total_received,
-            },
-            "filter": {
-               "with_geoRadius": filter_with_geo_radius,
-               "with_geoBoundingBox": filter_with_geo_bounding_box,
-               "avg_criteria_number": format!("{:.2}", filter_sum_of_criteria_terms as f64 / filter_total_number_of_criteria as f64),
-               "most_used_syntax": used_syntax.iter().max_by_key(|(_, v)| *v).map(|(k, _)| json!(k)).unwrap_or_else(|| json!(null)),
-            },
-            "vector": {
-                "retrieve_vectors": retrieve_vectors,
-            },
-            "pagination": {
-               "max_limit": max_limit,
-               "max_offset": max_offset,
-            },
-            "formatting": {
-                "max_attributes_to_retrieve": max_attributes_to_retrieve,
-            },
-            "scoring": {
-                "show_ranking_score": show_ranking_score,
-                "show_ranking_score_details": show_ranking_score_details,
-                "ranking_score_threshold": ranking_score_threshold,
-            }
-        })
-    }
-}
--- a/meilisearch/src/routes/mod.rs
+++ b/meilisearch/src/routes/mod.rs
@@ -25,7 +25,6 @@ pub mod indexes;
 mod logs;
 mod metrics;
 mod multi_search;
-mod multi_search_analytics;
 mod snapshot;
 mod swap_indexes;
 pub mod tasks;
--- a/meilisearch/src/routes/multi_search.rs
+++ b/meilisearch/src/routes/multi_search.rs
@@ -9,7 +9,7 @@ use meilisearch_types::keys::actions;
 use serde::Serialize;
 use tracing::debug;

-use crate::analytics::Analytics;
+use crate::analytics::{Analytics, MultiSearchAggregator};
 use crate::error::MeilisearchHttpError;
 use crate::extractors::authentication::policies::ActionPolicy;
 use crate::extractors::authentication::{AuthenticationError, GuardedData};
@@ -21,8 +21,6 @@ use crate::search::{
 };
 use crate::search_queue::SearchQueue;

-use super::multi_search_analytics::MultiSearchAggregator;
-
 pub fn configure(cfg: &mut web::ServiceConfig) {
    cfg.service(web::resource("").route(web::post().to(SeqHandler(multi_search_with_post))));
 }
@@ -37,7 +35,7 @@ pub async fn multi_search_with_post(
    search_queue: Data<SearchQueue>,
    params: AwebJson<FederatedSearch, DeserrJsonError>,
    req: HttpRequest,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    // Since we don't want to process half of the search requests and then get a permit refused
    // we're going to get one permit for the whole duration of the multi-search request.
@@ -45,7 +43,7 @@ pub async fn multi_search_with_post(

    let federated_search = params.into_inner();

-    let mut multi_aggregate = MultiSearchAggregator::from_federated_search(&federated_search);
+    let mut multi_aggregate = MultiSearchAggregator::from_federated_search(&federated_search, &req);

    let FederatedSearch { mut queries, federation } = federated_search;

@@ -89,7 +87,7 @@ pub async fn multi_search_with_post(
                multi_aggregate.succeed();
            }

-            analytics.publish(multi_aggregate, &req);
+            analytics.post_multi_search(multi_aggregate);
            HttpResponse::Ok().json(search_result??)
        }
        None => {
@@ -151,7 +149,7 @@ pub async fn multi_search_with_post(
            if search_results.is_ok() {
                multi_aggregate.succeed();
            }
-            analytics.publish(multi_aggregate, &req);
+            analytics.post_multi_search(multi_aggregate);

            let search_results = search_results.map_err(|(mut err, query_index)| {
                // Add the query index that failed as context for the error message.
--- a/meilisearch/src/routes/multi_search_analytics.rs
+++ b/meilisearch/src/routes/multi_search_analytics.rs
@@ -1,170 +0,0 @@
-use std::collections::HashSet;
-
-use serde_json::json;
-
-use crate::{
-    analytics::Aggregate,
-    search::{FederatedSearch, SearchQueryWithIndex},
-};
-
-#[derive(Default)]
-pub struct MultiSearchAggregator {
-    // requests
-    total_received: usize,
-    total_succeeded: usize,
-
-    // sum of the number of distinct indexes in each single request, use with total_received to compute an avg
-    total_distinct_index_count: usize,
-    // number of queries with a single index, use with total_received to compute a proportion
-    total_single_index: usize,
-
-    // sum of the number of search queries in the requests, use with total_received to compute an average
-    total_search_count: usize,
-
-    // scoring
-    show_ranking_score: bool,
-    show_ranking_score_details: bool,
-
-    // federation
-    use_federation: bool,
-}
-
-impl MultiSearchAggregator {
-    pub fn from_federated_search(federated_search: &FederatedSearch) -> Self {
-        let use_federation = federated_search.federation.is_some();
-
-        let distinct_indexes: HashSet<_> = federated_search
-            .queries
-            .iter()
-            .map(|query| {
-                let query = &query;
-                // make sure we get a compilation error if a field gets added to / removed from SearchQueryWithIndex
-                let SearchQueryWithIndex {
-                    index_uid,
-                    federation_options: _,
-                    q: _,
-                    vector: _,
-                    offset: _,
-                    limit: _,
-                    page: _,
-                    hits_per_page: _,
-                    attributes_to_retrieve: _,
-                    retrieve_vectors: _,
-                    attributes_to_crop: _,
-                    crop_length: _,
-                    attributes_to_highlight: _,
-                    show_ranking_score: _,
-                    show_ranking_score_details: _,
-                    show_matches_position: _,
-                    filter: _,
-                    sort: _,
-                    distinct: _,
-                    facets: _,
-                    highlight_pre_tag: _,
-                    highlight_post_tag: _,
-                    crop_marker: _,
-                    matching_strategy: _,
-                    attributes_to_search_on: _,
-                    hybrid: _,
-                    ranking_score_threshold: _,
-                    locales: _,
-                } = query;
-
-                index_uid.as_str()
-            })
-            .collect();
-
-        let show_ranking_score =
-            federated_search.queries.iter().any(|query| query.show_ranking_score);
-        let show_ranking_score_details =
-            federated_search.queries.iter().any(|query| query.show_ranking_score_details);
-
-        Self {
-            total_received: 1,
-            total_succeeded: 0,
-            total_distinct_index_count: distinct_indexes.len(),
-            total_single_index: if distinct_indexes.len() == 1 { 1 } else { 0 },
-            total_search_count: federated_search.queries.len(),
-            show_ranking_score,
-            show_ranking_score_details,
-            use_federation,
-        }
-    }
-
-    pub fn succeed(&mut self) {
-        self.total_succeeded = self.total_succeeded.saturating_add(1);
-    }
-}
-
-impl Aggregate for MultiSearchAggregator {
-    fn event_name(&self) -> &'static str {
-        "Documents Searched by Multi-Search POST"
-    }
-
-    /// Aggregate one [MultiSearchAggregator] into another.
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        // write the aggregate in a way that will cause a compilation error if a field is added.
-
-        // get ownership of self, replacing it by a default value.
-        let this = *self;
-
-        let total_received = this.total_received.saturating_add(new.total_received);
-        let total_succeeded = this.total_succeeded.saturating_add(new.total_succeeded);
-        let total_distinct_index_count =
-            this.total_distinct_index_count.saturating_add(new.total_distinct_index_count);
-        let total_single_index = this.total_single_index.saturating_add(new.total_single_index);
-        let total_search_count = this.total_search_count.saturating_add(new.total_search_count);
-        let show_ranking_score = this.show_ranking_score || new.show_ranking_score;
-        let show_ranking_score_details =
-            this.show_ranking_score_details || new.show_ranking_score_details;
-        let use_federation = this.use_federation || new.use_federation;
-
-        Box::new(Self {
-            total_received,
-            total_succeeded,
-            total_distinct_index_count,
-            total_single_index,
-            total_search_count,
-            show_ranking_score,
-            show_ranking_score_details,
-            use_federation,
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        let Self {
-            total_received,
-            total_succeeded,
-            total_distinct_index_count,
-            total_single_index,
-            total_search_count,
-            show_ranking_score,
-            show_ranking_score_details,
-            use_federation,
-        } = *self;
-
-        json!({
-            "requests": {
-                "total_succeeded": total_succeeded,
-                "total_failed": total_received.saturating_sub(total_succeeded), // just to be sure we never panics
-                "total_received": total_received,
-            },
-            "indexes": {
-                "total_single_index": total_single_index,
-                "total_distinct_index_count": total_distinct_index_count,
-                "avg_distinct_index_count": (total_distinct_index_count as f64) / (total_received as f64), // not 0 else returned early
-            },
-            "searches": {
-                "total_search_count": total_search_count,
-                "avg_search_count": (total_search_count as f64) / (total_received as f64),
-            },
-            "scoring": {
-                "show_ranking_score": show_ranking_score,
-                "show_ranking_score_details": show_ranking_score_details,
-            },
-            "federation": {
-                "use_federation": use_federation,
-            }
-        })
-    }
-}
--- a/meilisearch/src/routes/snapshot.rs
+++ b/meilisearch/src/routes/snapshot.rs
@@ -3,6 +3,7 @@ use actix_web::{web, HttpRequest, HttpResponse};
 use index_scheduler::IndexScheduler;
 use meilisearch_types::error::ResponseError;
 use meilisearch_types::tasks::KindWithContent;
+use serde_json::json;
 use tracing::debug;

 use crate::analytics::Analytics;
@@ -16,15 +17,13 @@ pub fn configure(cfg: &mut web::ServiceConfig) {
    cfg.service(web::resource("").route(web::post().to(SeqHandler(create_snapshot))));
 }

-crate::empty_analytics!(SnapshotAnalytics, "Snapshot Created");
-
 pub async fn create_snapshot(
    index_scheduler: GuardedData<ActionPolicy<{ actions::SNAPSHOTS_CREATE }>, Data<IndexScheduler>>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
-    analytics.publish(SnapshotAnalytics::default(), &req);
+    analytics.publish("Snapshot Created".to_string(), json!({}), Some(&req));

    let task = KindWithContent::SnapshotCreation;
    let uid = get_task_id(&req, &opt)?;
--- a/meilisearch/src/routes/swap_indexes.rs
+++ b/meilisearch/src/routes/swap_indexes.rs
@@ -8,10 +8,10 @@ use meilisearch_types::error::deserr_codes::InvalidSwapIndexes;
 use meilisearch_types::error::ResponseError;
 use meilisearch_types::index_uid::IndexUid;
 use meilisearch_types::tasks::{IndexSwap, KindWithContent};
-use serde::Serialize;
+use serde_json::json;

 use super::{get_task_id, is_dry_run, SummarizedTaskView};
-use crate::analytics::{Aggregate, Analytics};
+use crate::analytics::Analytics;
 use crate::error::MeilisearchHttpError;
 use crate::extractors::authentication::policies::*;
 use crate::extractors::authentication::{AuthenticationError, GuardedData};
@@ -29,36 +29,21 @@ pub struct SwapIndexesPayload {
    indexes: Vec<IndexUid>,
 }

-#[derive(Serialize)]
-struct IndexSwappedAnalytics {
-    swap_operation_number: usize,
-}
-
-impl Aggregate for IndexSwappedAnalytics {
-    fn event_name(&self) -> &'static str {
-        "Indexes Swapped"
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self {
-            swap_operation_number: self.swap_operation_number.max(new.swap_operation_number),
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
-
 pub async fn swap_indexes(
    index_scheduler: GuardedData<ActionPolicy<{ actions::INDEXES_SWAP }>, Data<IndexScheduler>>,
    params: AwebJson<Vec<SwapIndexesPayload>, DeserrJsonError>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let params = params.into_inner();
-    analytics.publish(IndexSwappedAnalytics { swap_operation_number: params.len() }, &req);
+    analytics.publish(
+        "Indexes Swapped".to_string(),
+        json!({
+            "swap_operation_number": params.len(),
+        }),
+        Some(&req),
+    );
    let filters = index_scheduler.filters();

    let mut swaps = vec![];
--- a/meilisearch/src/routes/tasks.rs
+++ b/meilisearch/src/routes/tasks.rs
@@ -12,17 +12,18 @@ use meilisearch_types::star_or::{OptionStarOr, OptionStarOrList};
 use meilisearch_types::task_view::TaskView;
 use meilisearch_types::tasks::{Kind, KindWithContent, Status};
 use serde::Serialize;
+use serde_json::json;
 use time::format_description::well_known::Rfc3339;
 use time::macros::format_description;
 use time::{Date, Duration, OffsetDateTime, Time};
 use tokio::task;

 use super::{get_task_id, is_dry_run, SummarizedTaskView};
-use crate::analytics::{Aggregate, AggregateMethod, Analytics};
+use crate::analytics::Analytics;
 use crate::extractors::authentication::policies::*;
 use crate::extractors::authentication::GuardedData;
 use crate::extractors::sequential_extractor::SeqHandler;
-use crate::{aggregate_methods, Opt};
+use crate::Opt;

 const DEFAULT_LIMIT: u32 = 20;

@@ -157,69 +158,12 @@ impl TaskDeletionOrCancelationQuery {
    }
 }

-aggregate_methods!(
-    CancelTasks => "Tasks Canceled",
-    DeleteTasks => "Tasks Deleted",
-);
-
-#[derive(Serialize)]
-struct TaskFilterAnalytics<Method: AggregateMethod> {
-    filtered_by_uid: bool,
-    filtered_by_index_uid: bool,
-    filtered_by_type: bool,
-    filtered_by_status: bool,
-    filtered_by_canceled_by: bool,
-    filtered_by_before_enqueued_at: bool,
-    filtered_by_after_enqueued_at: bool,
-    filtered_by_before_started_at: bool,
-    filtered_by_after_started_at: bool,
-    filtered_by_before_finished_at: bool,
-    filtered_by_after_finished_at: bool,
-
-    #[serde(skip)]
-    marker: std::marker::PhantomData<Method>,
-}
-
-impl<Method: AggregateMethod + 'static> Aggregate for TaskFilterAnalytics<Method> {
-    fn event_name(&self) -> &'static str {
-        Method::event_name()
-    }
-
-    fn aggregate(self: Box<Self>, new: Box<Self>) -> Box<Self> {
-        Box::new(Self {
-            filtered_by_uid: self.filtered_by_uid | new.filtered_by_uid,
-            filtered_by_index_uid: self.filtered_by_index_uid | new.filtered_by_index_uid,
-            filtered_by_type: self.filtered_by_type | new.filtered_by_type,
-            filtered_by_status: self.filtered_by_status | new.filtered_by_status,
-            filtered_by_canceled_by: self.filtered_by_canceled_by | new.filtered_by_canceled_by,
-            filtered_by_before_enqueued_at: self.filtered_by_before_enqueued_at
-                | new.filtered_by_before_enqueued_at,
-            filtered_by_after_enqueued_at: self.filtered_by_after_enqueued_at
-                | new.filtered_by_after_enqueued_at,
-            filtered_by_before_started_at: self.filtered_by_before_started_at
-                | new.filtered_by_before_started_at,
-            filtered_by_after_started_at: self.filtered_by_after_started_at
-                | new.filtered_by_after_started_at,
-            filtered_by_before_finished_at: self.filtered_by_before_finished_at
-                | new.filtered_by_before_finished_at,
-            filtered_by_after_finished_at: self.filtered_by_after_finished_at
-                | new.filtered_by_after_finished_at,
-
-            marker: std::marker::PhantomData,
-        })
-    }
-
-    fn into_event(self: Box<Self>) -> serde_json::Value {
-        serde_json::to_value(*self).unwrap_or_default()
-    }
-}
-
 async fn cancel_tasks(
    index_scheduler: GuardedData<ActionPolicy<{ actions::TASKS_CANCEL }>, Data<IndexScheduler>>,
    params: AwebQueryParameter<TaskDeletionOrCancelationQuery, DeserrQueryParamError>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let params = params.into_inner();

@@ -228,22 +172,21 @@ async fn cancel_tasks(
    }

    analytics.publish(
-        TaskFilterAnalytics::<CancelTasks> {
-            filtered_by_uid: params.uids.is_some(),
-            filtered_by_index_uid: params.index_uids.is_some(),
-            filtered_by_type: params.types.is_some(),
-            filtered_by_status: params.statuses.is_some(),
-            filtered_by_canceled_by: params.canceled_by.is_some(),
-            filtered_by_before_enqueued_at: params.before_enqueued_at.is_some(),
-            filtered_by_after_enqueued_at: params.after_enqueued_at.is_some(),
-            filtered_by_before_started_at: params.before_started_at.is_some(),
-            filtered_by_after_started_at: params.after_started_at.is_some(),
-            filtered_by_before_finished_at: params.before_finished_at.is_some(),
-            filtered_by_after_finished_at: params.after_finished_at.is_some(),
-
-            marker: std::marker::PhantomData,
-        },
-        &req,
+        "Tasks Canceled".to_string(),
+        json!({
+            "filtered_by_uid": params.uids.is_some(),
+            "filtered_by_index_uid": params.index_uids.is_some(),
+            "filtered_by_type": params.types.is_some(),
+            "filtered_by_status": params.statuses.is_some(),
+            "filtered_by_canceled_by": params.canceled_by.is_some(),
+            "filtered_by_before_enqueued_at": params.before_enqueued_at.is_some(),
+            "filtered_by_after_enqueued_at": params.after_enqueued_at.is_some(),
+            "filtered_by_before_started_at": params.before_started_at.is_some(),
+            "filtered_by_after_started_at": params.after_started_at.is_some(),
+            "filtered_by_before_finished_at": params.before_finished_at.is_some(),
+            "filtered_by_after_finished_at": params.after_finished_at.is_some(),
+        }),
+        Some(&req),
    );

    let query = params.into_query();
@@ -271,7 +214,7 @@ async fn delete_tasks(
    params: AwebQueryParameter<TaskDeletionOrCancelationQuery, DeserrQueryParamError>,
    req: HttpRequest,
    opt: web::Data<Opt>,
-    analytics: web::Data<Analytics>,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
    let params = params.into_inner();

@@ -280,24 +223,22 @@ async fn delete_tasks(
    }

    analytics.publish(
-        TaskFilterAnalytics::<DeleteTasks> {
-            filtered_by_uid: params.uids.is_some(),
-            filtered_by_index_uid: params.index_uids.is_some(),
-            filtered_by_type: params.types.is_some(),
-            filtered_by_status: params.statuses.is_some(),
-            filtered_by_canceled_by: params.canceled_by.is_some(),
-            filtered_by_before_enqueued_at: params.before_enqueued_at.is_some(),
-            filtered_by_after_enqueued_at: params.after_enqueued_at.is_some(),
-            filtered_by_before_started_at: params.before_started_at.is_some(),
-            filtered_by_after_started_at: params.after_started_at.is_some(),
-            filtered_by_before_finished_at: params.before_finished_at.is_some(),
-            filtered_by_after_finished_at: params.after_finished_at.is_some(),
-
-            marker: std::marker::PhantomData,
-        },
-        &req,
+        "Tasks Deleted".to_string(),
+        json!({
+            "filtered_by_uid": params.uids.is_some(),
+            "filtered_by_index_uid": params.index_uids.is_some(),
+            "filtered_by_type": params.types.is_some(),
+            "filtered_by_status": params.statuses.is_some(),
+            "filtered_by_canceled_by": params.canceled_by.is_some(),
+            "filtered_by_before_enqueued_at": params.before_enqueued_at.is_some(),
+            "filtered_by_after_enqueued_at": params.after_enqueued_at.is_some(),
+            "filtered_by_before_started_at": params.before_started_at.is_some(),
+            "filtered_by_after_started_at": params.after_started_at.is_some(),
+            "filtered_by_before_finished_at": params.before_finished_at.is_some(),
+            "filtered_by_after_finished_at": params.after_finished_at.is_some(),
+        }),
+        Some(&req),
    );
-
    let query = params.into_query();

    let (tasks, _) = index_scheduler.get_task_ids_from_authorized_indexes(
--- a/meilisearch/src/search/mod.rs
+++ b/meilisearch/src/search/mod.rs
@@ -796,10 +796,8 @@ fn prepare_search<'t>(
                    let span = tracing::trace_span!(target: "search::vector", "embed_one");
                    let _entered = span.enter();

-                    let deadline = std::time::Instant::now() + std::time::Duration::from_secs(10);
-
                    embedder
-                        .embed_one(query.q.clone().unwrap(), Some(deadline))
+                        .embed_one(query.q.clone().unwrap())
                        .map_err(milli::vector::Error::from)
                        .map_err(milli::Error::from)?
                }
@@ -1197,13 +1195,8 @@ impl<'a> HitMaker<'a> {
        let vectors_is_hidden = match (&displayed_ids, vectors_fid) {
            // displayed_ids is a wildcard, so `_vectors` can be displayed regardless of its fid
            (None, _) => false,
-            // vectors has no fid, so check its explicit name
-            (Some(_), None) => {
-                // unwrap as otherwise we'd go to the first one
-                let displayed_names = index.displayed_fields(rtxn)?.unwrap();
-                !displayed_names
-                    .contains(&milli::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME)
-            }
+            // displayed_ids is a finite list, and `_vectors` cannot be part of it because it is not an existing field
+            (Some(_), None) => true,
            // displayed_ids is a finit list, so hide if `_vectors` is not part of it
            (Some(map), Some(vectors_fid)) => map.contains(&vectors_fid),
        };
@@ -1689,7 +1682,7 @@ fn add_non_formatted_ids_to_formatted_options(
 fn make_document(
    displayed_attributes: &BTreeSet<FieldId>,
    field_ids_map: &FieldsIdsMap,
-    obkv: obkv::KvReaderU16,
+    obkv: &obkv::KvReaderU16,
 ) -> Result<Document, MeilisearchHttpError> {
    let mut document = serde_json::Map::new();

--- a/meilisearch/tests/common/server.rs
+++ b/meilisearch/tests/common/server.rs
@@ -381,6 +381,7 @@ pub fn default_settings(dir: impl AsRef<Path>) -> Opt {
        db_path: dir.as_ref().join("db"),
        dump_dir: dir.as_ref().join("dumps"),
        env: "development".to_owned(),
+        #[cfg(feature = "analytics")]
        no_analytics: true,
        max_index_size: Byte::from_u64_with_unit(100, Unit::MiB).unwrap(),
        max_task_db_size: Byte::from_u64_with_unit(1, Unit::GiB).unwrap(),
--- a/meilisearch/tests/common/service.rs
+++ b/meilisearch/tests/common/service.rs
@@ -9,9 +9,8 @@ use actix_web::test;
 use actix_web::test::TestRequest;
 use actix_web::web::Data;
 use index_scheduler::IndexScheduler;
-use meilisearch::analytics::Analytics;
 use meilisearch::search_queue::SearchQueue;
-use meilisearch::{create_app, Opt, SubscriberForSecondLayer};
+use meilisearch::{analytics, create_app, Opt, SubscriberForSecondLayer};
 use meilisearch_auth::AuthController;
 use tracing::level_filters::LevelFilter;
 use tracing_subscriber::Layer;
@@ -142,7 +141,7 @@ impl Service {
            Data::new(search_queue),
            self.options.clone(),
            (route_layer_handle, stderr_layer_handle),
-            Data::new(Analytics::no_analytics()),
+            analytics::MockAnalytics::new(&self.options),
            true,
        ))
        .await
--- a/meilisearch/tests/logs/mod.rs
+++ b/meilisearch/tests/logs/mod.rs
@@ -7,9 +7,8 @@ use std::str::FromStr;
 use actix_web::http::header::ContentType;
 use actix_web::web::Data;
 use meili_snap::snapshot;
-use meilisearch::analytics::Analytics;
 use meilisearch::search_queue::SearchQueue;
-use meilisearch::{create_app, Opt, SubscriberForSecondLayer};
+use meilisearch::{analytics, create_app, Opt, SubscriberForSecondLayer};
 use tracing::level_filters::LevelFilter;
 use tracing_subscriber::layer::SubscriberExt;
 use tracing_subscriber::Layer;
@@ -55,7 +54,7 @@ async fn basic_test_log_stream_route() {
        Data::new(search_queue),
        server.service.options.clone(),
        (route_layer_handle, stderr_layer_handle),
-        Data::new(Analytics::no_analytics()),
+        analytics::MockAnalytics::new(&server.service.options),
        true,
    ))
    .await;
--- a/meilisearch/tests/search/hybrid.rs
+++ b/meilisearch/tests/search/hybrid.rs
@@ -568,57 +568,6 @@ async fn retrieve_vectors() {
    ]
    "###);

-    // use explicit `_vectors` in displayed attributes
-    let (response, code) = index
-        .update_settings(json!({ "displayedAttributes": ["id", "title", "desc", "_vectors"]} ))
-        .await;
-    assert_eq!(202, code, "{:?}", response);
-    index.wait_task(response.uid()).await;
-
-    let (response, code) = index
-        .search_post(
-            json!({"q": "Captain", "hybrid": {"embedder": "default", "semanticRatio": 0.2}, "retrieveVectors": true}),
-        )
-        .await;
-    snapshot!(code, @"200 OK");
-    insta::assert_json_snapshot!(response["hits"], {"[]._vectors.default.embeddings" => "[vectors]"},  @r###"
-    [
-      {
-        "title": "Captain Planet",
-        "desc": "He's not part of the Marvel Cinematic Universe",
-        "id": "2",
-        "_vectors": {
-          "default": {
-            "embeddings": "[vectors]",
-            "regenerate": true
-          }
-        }
-      },
-      {
-        "title": "Captain Marvel",
-        "desc": "a Shazam ersatz",
-        "id": "3",
-        "_vectors": {
-          "default": {
-            "embeddings": "[vectors]",
-            "regenerate": true
-          }
-        }
-      },
-      {
-        "title": "Shazam!",
-        "desc": "a Captain Marvel ersatz",
-        "id": "1",
-        "_vectors": {
-          "default": {
-            "embeddings": "[vectors]",
-            "regenerate": true
-          }
-        }
-      }
-    ]
-    "###);
-
    // remove `_vectors` from displayed attributes
    let (response, code) =
        index.update_settings(json!({ "displayedAttributes": ["id", "title", "desc"]} )).await;
--- a/meilisearch/tests/vector/openai.rs
+++ b/meilisearch/tests/vector/openai.rs
@@ -137,14 +137,13 @@ fn long_text() -> &'static str {
 }

 async fn create_mock_tokenized() -> (MockServer, Value) {
-    create_mock_with_template("{{doc.text}}", ModelDimensions::Large, false, false).await
+    create_mock_with_template("{{doc.text}}", ModelDimensions::Large, false).await
 }

 async fn create_mock_with_template(
    document_template: &str,
    model_dimensions: ModelDimensions,
    fallible: bool,
-    slow: bool,
 ) -> (MockServer, Value) {
    let mock_server = MockServer::start().await;
    const API_KEY: &str = "my-api-key";
@@ -155,11 +154,7 @@ async fn create_mock_with_template(
    Mock::given(method("POST"))
        .and(path("/"))
        .respond_with(move |req: &Request| {
-            // 0. wait for a long time
-            if slow {
-              std::thread::sleep(std::time::Duration::from_secs(1));
-            }
-            // 1. maybe return 500
+            // 0. maybe return 500
            if fallible {
             let attempt = attempt.fetch_add(1, Ordering::Relaxed);
             let failed = matches!(attempt % 4, 0 | 1 | 3);
@@ -172,7 +167,7 @@ async fn create_mock_with_template(
                }))
             }
            }
-            // 3. check API key
+            // 1. check API key
            match req.headers.get("Authorization") {
                Some(api_key) if api_key == API_KEY_BEARER => {
                    {}
@@ -207,7 +202,7 @@ async fn create_mock_with_template(
                    )
                }
            }
-            // 3. parse text inputs
+            // 2. parse text inputs
            let query: serde_json::Value = match req.body_json() {
                Ok(query) => query,
                Err(_error) => return ResponseTemplate::new(400).set_body_json(
@@ -228,7 +223,7 @@ async fn create_mock_with_template(
                panic!("Expected {model_dimensions:?}, got {query_model_dimensions:?}")
            }

-            // 4. for each text, find embedding in responses
+            // 3. for each text, find embedding in responses
            let serde_json::Value::Array(inputs) = &query["input"] else {
                panic!("Unexpected `input` value")
            };
@@ -288,7 +283,7 @@ async fn create_mock_with_template(
                "embedding": embedding,
            })).collect();

-            // 5. produce output from embeddings
+            // 4. produce output from embeddings
            ResponseTemplate::new(200).set_body_json(json!({
                "object": "list",
                "data": data,
@@ -322,27 +317,23 @@ const DOGGO_TEMPLATE: &str = r#"{%- if doc.gender == "F" -%}Une chienne nommée
        {%- endif %}, de race {{doc.breed}}."#;

 async fn create_mock() -> (MockServer, Value) {
-    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large, false, false).await
+    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large, false).await
 }

 async fn create_mock_dimensions() -> (MockServer, Value) {
-    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large512, false, false).await
+    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large512, false).await
 }

 async fn create_mock_small_embedding_model() -> (MockServer, Value) {
-    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Small, false, false).await
+    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Small, false).await
 }

 async fn create_mock_legacy_embedding_model() -> (MockServer, Value) {
-    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Ada, false, false).await
+    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Ada, false).await
 }

 async fn create_fallible_mock() -> (MockServer, Value) {
-    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large, true, false).await
-}
-
-async fn create_slow_mock() -> (MockServer, Value) {
-    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large, true, true).await
+    create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large, true).await
 }

 // basic test "it works"
@@ -1882,114 +1873,4 @@ async fn it_still_works() {
    ]
    "###);
 }
-
-// test with a server that responds 500 on 3 out of 4 calls
-#[actix_rt::test]
-async fn timeout() {
-    let (_mock, setting) = create_slow_mock().await;
-    let server = get_server_vector().await;
-    let index = server.index("doggo");
-
-    let (response, code) = index
-        .update_settings(json!({
-          "embedders": {
-              "default": setting,
-          },
-        }))
-        .await;
-    snapshot!(code, @"202 Accepted");
-    let task = server.wait_task(response.uid()).await;
-    snapshot!(task["status"], @r###""succeeded""###);
-    let documents = json!([
-      {"id": 0, "name": "kefir", "gender": "M", "birthyear": 2023, "breed": "Patou"},
-    ]);
-    let (value, code) = index.add_documents(documents, None).await;
-    snapshot!(code, @"202 Accepted");
-    let task = index.wait_task(value.uid()).await;
-    snapshot!(task, @r###"
-    {
-      "uid": "[uid]",
-      "indexUid": "doggo",
-      "status": "succeeded",
-      "type": "documentAdditionOrUpdate",
-      "canceledBy": null,
-      "details": {
-        "receivedDocuments": 1,
-        "indexedDocuments": 1
-      },
-      "error": null,
-      "duration": "[duration]",
-      "enqueuedAt": "[date]",
-      "startedAt": "[date]",
-      "finishedAt": "[date]"
-    }
-    "###);
-
-    let (documents, _code) = index
-        .get_all_documents(GetAllDocumentsOptions { retrieve_vectors: true, ..Default::default() })
-        .await;
-    snapshot!(json_string!(documents, {".results.*._vectors.default.embeddings" => "[vector]"}), @r###"
-    {
-      "results": [
-        {
-          "id": 0,
-          "name": "kefir",
-          "gender": "M",
-          "birthyear": 2023,
-          "breed": "Patou",
-          "_vectors": {
-            "default": {
-              "embeddings": "[vector]",
-              "regenerate": true
-            }
-          }
-        }
-      ],
-      "offset": 0,
-      "limit": 20,
-      "total": 1
-    }
-    "###);
-
-    let (response, code) = index
-        .search_post(json!({
-            "q": "grand chien de berger des montagnes",
-            "hybrid": {"semanticRatio": 0.99, "embedder": "default"}
-        }))
-        .await;
-    snapshot!(code, @"200 OK");
-    snapshot!(json_string!(response["semanticHitCount"]), @"0");
-    snapshot!(json_string!(response["hits"]), @"[]");
-
-    let (response, code) = index
-        .search_post(json!({
-            "q": "grand chien de berger des montagnes",
-            "hybrid": {"semanticRatio": 0.99, "embedder": "default"}
-        }))
-        .await;
-    snapshot!(code, @"200 OK");
-    snapshot!(json_string!(response["semanticHitCount"]), @"1");
-    snapshot!(json_string!(response["hits"]), @r###"
-    [
-      {
-        "id": 0,
-        "name": "kefir",
-        "gender": "M",
-        "birthyear": 2023,
-        "breed": "Patou"
-      }
-    ]
-    "###);
-
-    let (response, code) = index
-        .search_post(json!({
-            "q": "grand chien de berger des montagnes",
-            "hybrid": {"semanticRatio": 0.99, "embedder": "default"}
-        }))
-        .await;
-    snapshot!(code, @"200 OK");
-    snapshot!(json_string!(response["semanticHitCount"]), @"0");
-    snapshot!(json_string!(response["hits"]), @"[]");
-}
-
 // test with a server that wrongly responds 400
--- a/meilisearch/tests/vector/settings.rs
+++ b/meilisearch/tests/vector/settings.rs
@@ -4,53 +4,6 @@ use crate::common::{GetAllDocumentsOptions, Server};
 use crate::json;
 use crate::vector::generate_default_user_provided_documents;

-#[actix_rt::test]
-async fn field_unavailable_for_source() {
-    let server = Server::new().await;
-    let index = server.index("doggo");
-    let (value, code) = server.set_features(json!({"vectorStore": true})).await;
-    snapshot!(code, @"200 OK");
-    snapshot!(value, @r###"
-    {
-      "vectorStore": true,
-      "metrics": false,
-      "logsRoute": false,
-      "editDocumentsByFunction": false,
-      "containsFilter": false
-    }
-    "###);
-
-    let (response, code) = index
-        .update_settings(json!({
-          "embedders": { "manual": {"source": "userProvided", "documentTemplate": "{{doc.documentTemplate}}"}},
-        }))
-        .await;
-    snapshot!(code, @"400 Bad Request");
-    snapshot!(response, @r###"
-    {
-      "message": "`.embedders.manual`: Field `documentTemplate` unavailable for source `userProvided` (only available for sources: `huggingFace`, `openAi`, `ollama`, `rest`). Available fields: `source`, `dimensions`, `distribution`, `binaryQuantized`",
-      "code": "invalid_settings_embedders",
-      "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_settings_embedders"
-    }
-    "###);
-
-    let (response, code) = index
-        .update_settings(json!({
-          "embedders": { "default": {"source": "openAi", "revision": "42"}},
-        }))
-        .await;
-    snapshot!(code, @"400 Bad Request");
-    snapshot!(response, @r###"
-    {
-      "message": "`.embedders.default`: Field `revision` unavailable for source `openAi` (only available for sources: `huggingFace`). Available fields: `source`, `model`, `apiKey`, `documentTemplate`, `dimensions`, `distribution`, `url`, `binaryQuantized`",
-      "code": "invalid_settings_embedders",
-      "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_settings_embedders"
-    }
-    "###);
-}
-
 #[actix_rt::test]
 async fn update_embedder() {
    let server = Server::new().await;
--- a/meilitool/src/main.rs
+++ b/meilitool/src/main.rs
@@ -682,7 +682,7 @@ fn export_a_dump(
                        format!("While iterating on content file {:?}", content_file_uuid)
                    })? {
                        dump_content_file
-                            .push_document(&obkv_to_object(&doc, &documents_batch_index)?)?;
+                            .push_document(&obkv_to_object(doc, &documents_batch_index)?)?;
                    }
                    dump_content_file.flush()?;
                    count += 1;
--- a/milli/Cargo.toml
+++ b/milli/Cargo.toml
@@ -12,12 +12,14 @@ readme.workspace = true
 license.workspace = true

 [dependencies]
+big_s = "1.0.2"
 bimap = { version = "0.6.3", features = ["serde"] }
 bincode = "1.3.3"
 bstr = "1.9.1"
 bytemuck = { version = "1.16.1", features = ["extern_crate_alloc"] }
 byteorder = "1.5.0"
-charabia = { version = "0.9.1", default-features = false }
+# charabia = { version = "0.9.0", default-features = false }
+charabia = { git = "https://github.com/meilisearch/charabia", branch = "mutualize-char-normalizer", default-features = false }
 concat-arrays = "0.1.2"
 crossbeam-channel = "0.5.13"
 deserr = "0.6.2"
@@ -27,9 +29,9 @@ fst = "0.4.7"
 fxhash = "0.2.1"
 geoutils = "0.5.1"
 grenad = { version = "0.4.7", default-features = false, features = [
-    "rayon",
-    "tempfile",
-] }
+    "rayon", # TODO Should we keep this feature
+    "tempfile"
+], git = "https://github.com/meilisearch/grenad", branch = "various-improvements" }
 heed = { version = "0.20.3", default-features = false, features = [
    "serde-json",
    "serde-bincode",
@@ -40,14 +42,14 @@ json-depth-checker = { path = "../json-depth-checker" }
 levenshtein_automata = { version = "0.2.1", features = ["fst_automaton"] }
 memchr = "2.5.0"
 memmap2 = "0.9.4"
-obkv = "0.2.2"
+obkv = { git = "https://github.com/kerollmops/obkv", branch = "unsized-kvreader" }
 once_cell = "1.19.0"
 ordered-float = "4.2.1"
 rayon = "1.10.0"
 roaring = { version = "0.10.6", features = ["serde"] }
 rstar = { version = "0.12.0", features = ["serde"] }
 serde = { version = "1.0.204", features = ["derive"] }
-serde_json = { version = "1.0.120", features = ["preserve_order"] }
+serde_json = { version = "1.0.120", features = ["preserve_order", "raw_value"] }
 slice-group-by = "0.3.1"
 smallstr = { version = "0.3.0", features = ["serde"] }
 smallvec = "1.13.2"
@@ -79,17 +81,17 @@ hf-hub = { git = "https://github.com/dureuill/hf-hub.git", branch = "rust_tls",
 ] }
 tiktoken-rs = "0.5.9"
 liquid = "0.26.6"
-rhai = { git = "https://github.com/rhaiscript/rhai", rev = "ef3df63121d27aacd838f366f2b83fd65f20a1e4", features = ["serde", "no_module", "no_custom_syntax", "no_time", "sync"] }
-arroy = "0.5.0"
+rhai = { version = "1.19.0", features = ["serde", "no_module", "no_custom_syntax", "no_time", "sync"] }
+arroy = { git = "https://github.com/meilisearch/arroy/", rev = "2386594dfb009ce08821a925ccc89fb8e30bf73d" }
 rand = "0.8.5"
 tracing = "0.1.40"
 ureq = { version = "2.10.0", features = ["json"] }
 url = "2.5.2"
 rayon-par-bridge = "0.1.0"
+hashbrown = "0.14.5"

 [dev-dependencies]
 mimalloc = { version = "0.1.43", default-features = false }
-big_s = "1.0.2"
 insta = "1.39.0"
 maplit = "1.0.2"
 md5 = "0.7.0"
@@ -98,7 +100,16 @@ rand = { version = "0.8.5", features = ["small_rng"] }

 [features]
 all-tokenizations = [
-    "charabia/default",
+    "charabia/chinese",
+    "charabia/hebrew",
+    "charabia/japanese",
+    "charabia/thai",
+    "charabia/korean",
+    "charabia/greek",
+    "charabia/khmer",
+    "charabia/vietnamese",
+    "charabia/swedish-recomposition",
+    "charabia/german-segmentation",
 ]

 # Use POSIX semaphores instead of SysV semaphores in LMDB
@@ -137,8 +148,5 @@ german = ["charabia/german-segmentation"]
 # force swedish character recomposition
 swedish-recomposition = ["charabia/swedish-recomposition"]

-# allow turkish specialized tokenization
-turkish = ["charabia/turkish"]
-
 # allow CUDA support, see <https://github.com/meilisearch/meilisearch/issues/4306>
 cuda = ["candle-core/cuda"]
--- a/milli/src/documents/builder.rs
+++ b/milli/src/documents/builder.rs
@@ -292,7 +292,7 @@ mod test {
            .unwrap()
            .into_cursor_and_fields_index();
        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -321,7 +321,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -348,7 +348,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -375,7 +375,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -402,7 +402,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -429,7 +429,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -456,7 +456,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -483,7 +483,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -510,7 +510,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
@@ -555,7 +555,7 @@ mod test {
            .into_cursor_and_fields_index();

        let doc = cursor.next_document().unwrap().unwrap();
-        let val = obkv_to_object(&doc, &index).map(Value::from).unwrap();
+        let val = obkv_to_object(doc, &index).map(Value::from).unwrap();

        assert_eq!(
            val,
--- a/milli/src/documents/enriched.rs
+++ b/milli/src/documents/enriched.rs
@@ -69,7 +69,7 @@ impl<R: io::Read + io::Seek> EnrichedDocumentsBatchReader<R> {

 #[derive(Debug, Clone)]
 pub struct EnrichedDocument<'a> {
-    pub document: KvReader<'a, FieldId>,
+    pub document: &'a KvReader<FieldId>,
    pub document_id: DocumentId,
 }

--- a/milli/src/documents/mod.rs
+++ b/milli/src/documents/mod.rs
@@ -27,7 +27,7 @@ use crate::{FieldId, Object, Result};
 const DOCUMENTS_BATCH_INDEX_KEY: [u8; 8] = u64::MAX.to_be_bytes();

 /// Helper function to convert an obkv reader into a JSON object.
-pub fn obkv_to_object(obkv: &KvReader<'_, FieldId>, index: &DocumentsBatchIndex) -> Result<Object> {
+pub fn obkv_to_object(obkv: &KvReader<FieldId>, index: &DocumentsBatchIndex) -> Result<Object> {
    obkv.iter()
        .map(|(field_id, value)| {
            let field_name = index
@@ -76,7 +76,7 @@ impl DocumentsBatchIndex {
        self.0.get_by_right(name).cloned()
    }

-    pub fn recreate_json(&self, document: &obkv::KvReaderU16<'_>) -> Result<Object> {
+    pub fn recreate_json(&self, document: &obkv::KvReaderU16) -> Result<Object> {
        let mut map = Object::new();

        for (k, v) in document.iter() {
--- a/milli/src/documents/primary_key.rs
+++ b/milli/src/documents/primary_key.rs
@@ -1,8 +1,10 @@
+use std::borrow::Cow;
 use std::iter;
 use std::result::Result as StdResult;

-use serde_json::Value;
+use serde_json::{from_str, Value};

+use crate::update::new::{CowStr, TopLevelMap};
 use crate::{FieldId, InternalError, Object, Result, UserError};

 /// The symbol used to define levels in a nested primary key.
@@ -52,7 +54,7 @@ impl<'a> PrimaryKey<'a> {

    pub fn document_id(
        &self,
-        document: &obkv::KvReader<'_, FieldId>,
+        document: &obkv::KvReader<FieldId>,
        fields: &impl FieldIdMapper,
    ) -> Result<StdResult<String, DocumentIdExtractionError>> {
        match self {
@@ -100,6 +102,45 @@ impl<'a> PrimaryKey<'a> {
        }
    }

+    /// Returns the document ID based on the primary and
+    /// search for it recursively in zero-copy-deserialized documents.
+    pub fn document_id_from_top_level_map<'p>(
+        &self,
+        document: &TopLevelMap<'p>,
+    ) -> Result<StdResult<CowStr<'p>, DocumentIdExtractionError>> {
+        fn get_docid<'p>(
+            document: &TopLevelMap<'p>,
+            primary_key: &[&str],
+        ) -> Result<StdResult<CowStr<'p>, DocumentIdExtractionError>> {
+            match primary_key {
+                [] => unreachable!("arrrgh"), // would None be ok?
+                [primary_key] => match document.0.get(*primary_key) {
+                    Some(value) => match from_str::<u64>(value.get()) {
+                        Ok(value) => Ok(Ok(CowStr(Cow::Owned(value.to_string())))),
+                        Err(_) => match from_str(value.get()) {
+                            Ok(document_id) => Ok(Ok(document_id)),
+                            Err(e) => Ok(Err(DocumentIdExtractionError::InvalidDocumentId(
+                                UserError::SerdeJson(e),
+                            ))),
+                        },
+                    },
+                    None => Ok(Err(DocumentIdExtractionError::MissingDocumentId)),
+                },
+                [head, tail @ ..] => match document.0.get(*head) {
+                    Some(value) => {
+                        let document = from_str(value.get()).map_err(InternalError::SerdeJson)?;
+                        get_docid(&document, tail)
+                    }
+                    None => Ok(Err(DocumentIdExtractionError::MissingDocumentId)),
+                },
+            }
+        }
+
+        /// TODO do not allocate a vec everytime here
+        let primary_key: Vec<_> = self.name().split(PRIMARY_KEY_SPLIT_SYMBOL).collect();
+        get_docid(document, &primary_key)
+    }
+
    /// Returns an `Iterator` that gives all the possible fields names the primary key
    /// can have depending of the first level name and depth of the objects.
    pub fn possible_level_names(&self) -> impl Iterator<Item = (&str, &str)> + '_ {
--- a/milli/src/documents/reader.rs
+++ b/milli/src/documents/reader.rs
@@ -72,15 +72,24 @@ impl<R> DocumentsBatchCursor<R> {
 }

 impl<R: io::Read + io::Seek> DocumentsBatchCursor<R> {
+    /// Returns a single document from the database.
+    pub fn get(
+        &mut self,
+        offset: u32,
+    ) -> Result<Option<&KvReader<FieldId>>, DocumentsBatchCursorError> {
+        match self.cursor.move_on_key_equal_to(offset.to_be_bytes())? {
+            Some((key, value)) if key != DOCUMENTS_BATCH_INDEX_KEY => Ok(Some(value.into())),
+            _otherwise => Ok(None),
+        }
+    }
+
    /// Returns the next document, starting from the first one. Subsequent calls to
    /// `next_document` advance the document reader until all the documents have been read.
    pub fn next_document(
        &mut self,
-    ) -> Result<Option<KvReader<'_, FieldId>>, DocumentsBatchCursorError> {
+    ) -> Result<Option<&KvReader<FieldId>>, DocumentsBatchCursorError> {
        match self.cursor.move_on_next()? {
-            Some((key, value)) if key != DOCUMENTS_BATCH_INDEX_KEY => {
-                Ok(Some(KvReader::new(value)))
-            }
+            Some((key, value)) if key != DOCUMENTS_BATCH_INDEX_KEY => Ok(Some(value.into())),
            _otherwise => Ok(None),
        }
    }
--- a/milli/src/error.rs
+++ b/milli/src/error.rs
@@ -297,7 +297,6 @@ impl From<arroy::Error> for Error {
            arroy::Error::InvalidVecDimension { expected, received } => {
                Error::UserError(UserError::InvalidVectorDimensions { expected, found: received })
            }
-            arroy::Error::BuildCancelled => Error::InternalError(InternalError::AbortedIndexation),
            arroy::Error::DatabaseFull
            | arroy::Error::InvalidItemAppend
            | arroy::Error::UnmatchingDistance { .. }
--- a/milli/src/fields_ids_map.rs
+++ b/milli/src/fields_ids_map.rs
@@ -4,6 +4,9 @@ use serde::{Deserialize, Serialize};

 use crate::FieldId;

+mod global;
+pub use global::GlobalFieldsIdsMap;
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct FieldsIdsMap {
    names_ids: BTreeMap<String, FieldId>,
--- a/milli/src/fields_ids_map/global.rs
+++ b/milli/src/fields_ids_map/global.rs
@@ -0,0 +1,86 @@
+use std::collections::BTreeMap;
+use std::sync::RwLock;
+
+use crate::{FieldId, FieldsIdsMap};
+
+/// A fields ids map that can be globally updated to add fields
+#[derive(Debug, Clone)]
+pub struct GlobalFieldsIdsMap<'indexing> {
+    global: &'indexing RwLock<FieldsIdsMap>,
+    local: LocalFieldsIdsMap,
+}
+
+#[derive(Debug, Clone)]
+struct LocalFieldsIdsMap {
+    names_ids: BTreeMap<String, FieldId>,
+    ids_names: BTreeMap<FieldId, String>,
+}
+
+impl LocalFieldsIdsMap {
+    fn new(global: &RwLock<FieldsIdsMap>) -> Self {
+        let global = global.read().unwrap();
+        Self { names_ids: global.names_ids.clone(), ids_names: global.ids_names.clone() }
+    }
+
+    fn insert(&mut self, name: &str, field_id: FieldId) {
+        self.names_ids.insert(name.to_owned(), field_id);
+        self.ids_names.insert(field_id, name.to_owned());
+    }
+
+    fn name(&self, id: FieldId) -> Option<&str> {
+        self.ids_names.get(&id).map(String::as_str)
+    }
+
+    fn id(&self, name: &str) -> Option<FieldId> {
+        self.names_ids.get(name).copied()
+    }
+}
+
+impl<'indexing> GlobalFieldsIdsMap<'indexing> {
+    pub fn new(global: &'indexing RwLock<FieldsIdsMap>) -> Self {
+        Self { local: LocalFieldsIdsMap::new(global), global }
+    }
+
+    /// Returns the field id related to a field name, it will create a new field id if the
+    /// name is not already known. Returns `None` if the maximum field id as been reached.
+    pub fn id_or_insert(&mut self, name: &str) -> Option<FieldId> {
+        if let Some(field_id) = self.local.id(name) {
+            return Some(field_id);
+        }
+
+        {
+            // optimistically lookup the global map
+            let global = self.global.read().unwrap();
+
+            if let Some(field_id) = global.id(name) {
+                self.local.insert(name, field_id);
+                return Some(field_id);
+            }
+        }
+
+        {
+            let mut global = self.global.write().unwrap();
+
+            if let Some(field_id) = global.id(name) {
+                self.local.insert(name, field_id);
+                return Some(field_id);
+            }
+
+            let field_id = global.insert(name)?;
+            self.local.insert(name, field_id);
+            Some(field_id)
+        }
+    }
+
+    /// Get the name of a field based on its id.
+    pub fn name(&mut self, id: FieldId) -> Option<&str> {
+        if self.local.name(id).is_none() {
+            let global = self.global.read().unwrap();
+
+            let name = global.name(id)?;
+            self.local.insert(name, id);
+        }
+
+        self.local.name(id)
+    }
+}
--- a/milli/src/heed_codec/obkv_codec.rs
+++ b/milli/src/heed_codec/obkv_codec.rs
@@ -6,10 +6,10 @@ use obkv::{KvReaderU16, KvWriterU16};
 pub struct ObkvCodec;

 impl<'a> heed::BytesDecode<'a> for ObkvCodec {
-    type DItem = KvReaderU16<'a>;
+    type DItem = &'a KvReaderU16;

    fn bytes_decode(bytes: &'a [u8]) -> Result<Self::DItem, BoxedError> {
-        Ok(KvReaderU16::new(bytes))
+        Ok(KvReaderU16::from_slice(bytes))
    }
 }

--- a/milli/src/heed_codec/roaring_bitmap/cbo_roaring_bitmap_codec.rs
+++ b/milli/src/heed_codec/roaring_bitmap/cbo_roaring_bitmap_codec.rs
@@ -122,7 +122,7 @@ impl CboRoaringBitmapCodec {

    /// Merges a DelAdd delta into a CboRoaringBitmap.
    pub fn merge_deladd_into<'a>(
-        deladd: KvReaderDelAdd<'_>,
+        deladd: &KvReaderDelAdd,
        previous: &[u8],
        buffer: &'a mut Vec<u8>,
    ) -> io::Result<Option<&'a [u8]>> {
--- a/milli/src/index.rs
+++ b/milli/src/index.rs
@@ -1251,12 +1251,20 @@ impl Index {

    /* documents */

+    /// Returns a document by using the document id.
+    pub fn document<'t>(&self, rtxn: &'t RoTxn, id: DocumentId) -> Result<&'t obkv::KvReaderU16> {
+        self.documents
+            .get(rtxn, &id)?
+            .ok_or(UserError::UnknownInternalDocumentId { document_id: id })
+            .map_err(Into::into)
+    }
+
    /// Returns an iterator over the requested documents. The next item will be an error if a document is missing.
    pub fn iter_documents<'a, 't: 'a>(
        &'a self,
        rtxn: &'t RoTxn<'t>,
        ids: impl IntoIterator<Item = DocumentId> + 'a,
-    ) -> Result<impl Iterator<Item = Result<(DocumentId, obkv::KvReaderU16<'t>)>> + 'a> {
+    ) -> Result<impl Iterator<Item = Result<(DocumentId, &'t obkv::KvReaderU16)>> + 'a> {
        Ok(ids.into_iter().map(move |id| {
            let kv = self
                .documents
@@ -1271,7 +1279,7 @@ impl Index {
        &self,
        rtxn: &'t RoTxn<'t>,
        ids: impl IntoIterator<Item = DocumentId>,
-    ) -> Result<Vec<(DocumentId, obkv::KvReaderU16<'t>)>> {
+    ) -> Result<Vec<(DocumentId, &'t obkv::KvReaderU16)>> {
        self.iter_documents(rtxn, ids)?.collect()
    }

@@ -1279,7 +1287,7 @@ impl Index {
    pub fn all_documents<'a, 't: 'a>(
        &'a self,
        rtxn: &'t RoTxn<'t>,
-    ) -> Result<impl Iterator<Item = Result<(DocumentId, obkv::KvReaderU16<'t>)>> + 'a> {
+    ) -> Result<impl Iterator<Item = Result<(DocumentId, &'t obkv::KvReaderU16)>> + 'a> {
        self.iter_documents(rtxn, self.documents_ids(rtxn)?)
    }

@@ -1303,7 +1311,7 @@ impl Index {
        })?;
        Ok(self.iter_documents(rtxn, ids)?.map(move |entry| -> Result<_> {
            let (_docid, obkv) = entry?;
-            match primary_key.document_id(&obkv, &fields)? {
+            match primary_key.document_id(obkv, &fields)? {
                Ok(document_id) => Ok(document_id),
                Err(_) => Err(InternalError::DocumentsError(
                    crate::documents::Error::InvalidDocumentFormat,
@@ -1610,6 +1618,24 @@ impl Index {
            .unwrap_or_default())
    }

+    pub fn arroy_readers<'a>(
+        &'a self,
+        rtxn: &'a RoTxn<'a>,
+        embedder_id: u8,
+        quantized: bool,
+    ) -> impl Iterator<Item = Result<ArroyWrapper>> + 'a {
+        crate::vector::arroy_db_range_for_embedder(embedder_id).map_while(move |k| {
+            let reader = ArroyWrapper::new(self.vector_arroy, k, quantized);
+            // Here we don't care about the dimensions, but we want to know if we can read
+            // in the database or if its metadata are missing because there is no document with that many vectors.
+            match reader.dimensions(rtxn) {
+                Ok(_) => Some(Ok(reader)),
+                Err(arroy::Error::MissingMetadata(_)) => None,
+                Err(e) => Some(Err(e.into())),
+            }
+        })
+    }
+
    pub(crate) fn put_search_cutoff(&self, wtxn: &mut RwTxn<'_>, cutoff: u64) -> heed::Result<()> {
        self.main.remap_types::<Str, BEU64>().put(wtxn, main_key::SEARCH_CUTOFF, &cutoff)
    }
@@ -1631,9 +1657,14 @@ impl Index {
        let embedding_configs = self.embedding_configs(rtxn)?;
        for config in embedding_configs {
            let embedder_id = self.embedder_category_id.get(rtxn, &config.name)?.unwrap();
-            let reader =
-                ArroyWrapper::new(self.vector_arroy, embedder_id, config.config.quantized());
-            let embeddings = reader.item_vectors(rtxn, docid)?;
+            let embeddings = self
+                .arroy_readers(rtxn, embedder_id, config.config.quantized())
+                .map_while(|reader| {
+                    reader
+                        .and_then(|r| r.item_vector(rtxn, docid).map_err(|e| e.into()))
+                        .transpose()
+                })
+                .collect::<Result<Vec<_>>>()?;
            res.insert(config.name.to_owned(), embeddings);
        }
        Ok(res)
--- a/milli/src/lib.rs
+++ b/milli/src/lib.rs
@@ -55,7 +55,7 @@ pub use self::error::{
 };
 pub use self::external_documents_ids::ExternalDocumentsIds;
 pub use self::fieldids_weights_map::FieldidsWeightsMap;
-pub use self::fields_ids_map::FieldsIdsMap;
+pub use self::fields_ids_map::{FieldsIdsMap, GlobalFieldsIdsMap};
 pub use self::heed_codec::{
    BEU16StrCodec, BEU32StrCodec, BoRoaringBitmapCodec, BoRoaringBitmapLenCodec,
    CboRoaringBitmapCodec, CboRoaringBitmapLenCodec, FieldIdWordCountCodec, ObkvCodec,
@@ -214,7 +214,7 @@ pub fn bucketed_position(relative: u16) -> u16 {
 pub fn obkv_to_json(
    displayed_fields: &[FieldId],
    fields_ids_map: &FieldsIdsMap,
-    obkv: obkv::KvReaderU16<'_>,
+    obkv: &obkv::KvReaderU16,
 ) -> Result<Object> {
    displayed_fields
        .iter()
@@ -232,10 +232,7 @@ pub fn obkv_to_json(
 }

 /// Transform every field of a raw obkv store into a JSON Object.
-pub fn all_obkv_to_json(
-    obkv: obkv::KvReaderU16<'_>,
-    fields_ids_map: &FieldsIdsMap,
-) -> Result<Object> {
+pub fn all_obkv_to_json(obkv: &obkv::KvReaderU16, fields_ids_map: &FieldsIdsMap) -> Result<Object> {
    let all_keys = obkv.iter().map(|(k, _v)| k).collect::<Vec<_>>();
    obkv_to_json(all_keys.as_slice(), fields_ids_map, obkv)
 }
@@ -434,7 +431,7 @@ mod tests {
        writer.insert(id1, b"1234").unwrap();
        writer.insert(id2, b"4321").unwrap();
        let contents = writer.into_inner().unwrap();
-        let obkv = obkv::KvReaderU16::new(&contents);
+        let obkv = obkv::KvReaderU16::from_slice(&contents);

        let expected = json!({
            "field1": 1234,
--- a/milli/src/prompt/document.rs
+++ b/milli/src/prompt/document.rs
@@ -30,13 +30,13 @@ impl ParsedValue {

 impl<'a> Document<'a> {
    pub fn new(
-        data: obkv::KvReaderU16<'a>,
+        data: &'a obkv::KvReaderU16,
        side: DelAdd,
        inverted_field_map: &'a FieldsIdsMap,
    ) -> Self {
        let mut out_data = BTreeMap::new();
        for (fid, raw) in data {
-            let obkv = KvReaderDelAdd::new(raw);
+            let obkv = KvReaderDelAdd::from_slice(raw);
            let Some(raw) = obkv.get(side) else {
                continue;
            };
--- a/milli/src/prompt/mod.rs
+++ b/milli/src/prompt/mod.rs
@@ -111,7 +111,7 @@ impl Prompt {

    pub fn render(
        &self,
-        document: obkv::KvReaderU16<'_>,
+        document: &obkv::KvReaderU16,
        side: DelAdd,
        field_id_map: &FieldsIdsMapWithMetadata,
    ) -> Result<String, RenderPromptError> {
--- a/milli/src/search/hybrid.rs
+++ b/milli/src/search/hybrid.rs
@@ -201,9 +201,7 @@ impl<'a> Search<'a> {
                let span = tracing::trace_span!(target: "search::hybrid", "embed_one");
                let _entered = span.enter();

-                let deadline = std::time::Instant::now() + std::time::Duration::from_secs(3);
-
-                match embedder.embed_one(query, Some(deadline)) {
+                match embedder.embed_one(query) {
                    Ok(embedding) => embedding,
                    Err(error) => {
                        tracing::error!(error=%error, "Embedding failed");
--- a/milli/src/search/new/db_cache.rs
+++ b/milli/src/search/new/db_cache.rs
@@ -3,6 +3,7 @@ use std::collections::hash_map::Entry;
 use std::hash::Hash;

 use fxhash::FxHashMap;
+use grenad::MergeFunction;
 use heed::types::Bytes;
 use heed::{BytesEncode, Database, RoTxn};
 use roaring::RoaringBitmap;
@@ -11,7 +12,7 @@ use super::interner::Interned;
 use super::Word;
 use crate::heed_codec::{BytesDecodeOwned, StrBEU16Codec};
 use crate::proximity::ProximityPrecision;
-use crate::update::{merge_cbo_roaring_bitmaps, MergeFn};
+use crate::update::MergeCboRoaringBitmaps;
 use crate::{
    CboRoaringBitmapCodec, CboRoaringBitmapLenCodec, Result, SearchContext, U8StrStrCodec,
 };
@@ -110,19 +111,21 @@ impl<'ctx> DatabaseCache<'ctx> {
            .map_err(Into::into)
    }

-    fn get_value_from_keys<'v, K1, KC>(
+    fn get_value_from_keys<'v, K1, KC, MF>(
        txn: &'ctx RoTxn<'_>,
        cache_key: K1,
        db_keys: &'v [KC::EItem],
        cache: &mut FxHashMap<K1, Option<Cow<'ctx, [u8]>>>,
        db: Database<KC, Bytes>,
        universe: Option<&RoaringBitmap>,
-        merger: MergeFn,
+        merger: MF,
    ) -> Result<Option<RoaringBitmap>>
    where
        K1: Copy + Eq + Hash,
        KC: BytesEncode<'v>,
        KC::EItem: Sized,
+        MF: MergeFunction,
+        crate::Error: From<MF::Error>,
    {
        if let Entry::Vacant(entry) = cache.entry(cache_key) {
            let bitmap_ptr: Option<Cow<'ctx, [u8]>> = match db_keys {
@@ -138,7 +141,7 @@ impl<'ctx> DatabaseCache<'ctx> {
                    if bitmaps.is_empty() {
                        None
                    } else {
-                        Some(merger(&[], &bitmaps[..])?)
+                        Some(merger.merge(&[], &bitmaps[..])?)
                    }
                }
            };
@@ -213,17 +216,17 @@ impl<'ctx> SearchContext<'ctx> {
                let keys: Vec<_> =
                    restricted_fids.tolerant.iter().map(|(fid, _)| (interned, *fid)).collect();

-                DatabaseCache::get_value_from_keys::<_, _>(
+                DatabaseCache::get_value_from_keys(
                    self.txn,
                    word,
                    &keys[..],
                    &mut self.db_cache.word_docids,
                    self.index.word_fid_docids.remap_data_type::<Bytes>(),
                    universe,
-                    merge_cbo_roaring_bitmaps,
+                    MergeCboRoaringBitmaps,
                )
            }
-            None => DatabaseCache::get_value::<_, _>(
+            None => DatabaseCache::get_value(
                self.txn,
                word,
                self.word_interner.get(word).as_str(),
@@ -245,17 +248,17 @@ impl<'ctx> SearchContext<'ctx> {
                let keys: Vec<_> =
                    restricted_fids.exact.iter().map(|(fid, _)| (interned, *fid)).collect();

-                DatabaseCache::get_value_from_keys::<_, _>(
+                DatabaseCache::get_value_from_keys(
                    self.txn,
                    word,
                    &keys[..],
                    &mut self.db_cache.exact_word_docids,
                    self.index.word_fid_docids.remap_data_type::<Bytes>(),
                    universe,
-                    merge_cbo_roaring_bitmaps,
+                    MergeCboRoaringBitmaps,
                )
            }
-            None => DatabaseCache::get_value::<_, _>(
+            None => DatabaseCache::get_value(
                self.txn,
                word,
                self.word_interner.get(word).as_str(),
@@ -302,17 +305,17 @@ impl<'ctx> SearchContext<'ctx> {
                let keys: Vec<_> =
                    restricted_fids.tolerant.iter().map(|(fid, _)| (interned, *fid)).collect();

-                DatabaseCache::get_value_from_keys::<_, _>(
+                DatabaseCache::get_value_from_keys(
                    self.txn,
                    prefix,
                    &keys[..],
                    &mut self.db_cache.word_prefix_docids,
                    self.index.word_prefix_fid_docids.remap_data_type::<Bytes>(),
                    universe,
-                    merge_cbo_roaring_bitmaps,
+                    MergeCboRoaringBitmaps,
                )
            }
-            None => DatabaseCache::get_value::<_, _>(
+            None => DatabaseCache::get_value(
                self.txn,
                prefix,
                self.word_interner.get(prefix).as_str(),
@@ -334,17 +337,17 @@ impl<'ctx> SearchContext<'ctx> {
                let keys: Vec<_> =
                    restricted_fids.exact.iter().map(|(fid, _)| (interned, *fid)).collect();

-                DatabaseCache::get_value_from_keys::<_, _>(
+                DatabaseCache::get_value_from_keys(
                    self.txn,
                    prefix,
                    &keys[..],
                    &mut self.db_cache.exact_word_prefix_docids,
                    self.index.word_prefix_fid_docids.remap_data_type::<Bytes>(),
                    universe,
-                    merge_cbo_roaring_bitmaps,
+                    MergeCboRoaringBitmaps,
                )
            }
-            None => DatabaseCache::get_value::<_, _>(
+            None => DatabaseCache::get_value(
                self.txn,
                prefix,
                self.word_interner.get(prefix).as_str(),
@@ -405,7 +408,7 @@ impl<'ctx> SearchContext<'ctx> {

                Ok(docids)
            }
-            ProximityPrecision::ByWord => DatabaseCache::get_value::<_, _>(
+            ProximityPrecision::ByWord => DatabaseCache::get_value(
                self.txn,
                (proximity, word1, word2),
                &(
@@ -538,7 +541,7 @@ impl<'ctx> SearchContext<'ctx> {
            return Ok(None);
        }

-        DatabaseCache::get_value::<_, _>(
+        DatabaseCache::get_value(
            self.txn,
            (word, fid),
            &(self.word_interner.get(word).as_str(), fid),
@@ -559,7 +562,7 @@ impl<'ctx> SearchContext<'ctx> {
            return Ok(None);
        }

-        DatabaseCache::get_value::<_, _>(
+        DatabaseCache::get_value(
            self.txn,
            (word_prefix, fid),
            &(self.word_interner.get(word_prefix).as_str(), fid),
@@ -629,7 +632,7 @@ impl<'ctx> SearchContext<'ctx> {
        word: Interned<String>,
        position: u16,
    ) -> Result<Option<RoaringBitmap>> {
-        DatabaseCache::get_value::<_, _>(
+        DatabaseCache::get_value(
            self.txn,
            (word, position),
            &(self.word_interner.get(word).as_str(), position),
@@ -645,7 +648,7 @@ impl<'ctx> SearchContext<'ctx> {
        word_prefix: Interned<String>,
        position: u16,
    ) -> Result<Option<RoaringBitmap>> {
-        DatabaseCache::get_value::<_, _>(
+        DatabaseCache::get_value(
            self.txn,
            (word_prefix, position),
            &(self.word_interner.get(word_prefix).as_str(), position),
--- a/milli/src/search/new/mod.rs
+++ b/milli/src/search/new/mod.rs
@@ -553,7 +553,7 @@ fn resolve_sort_criteria<'ctx, Query: RankingRuleQueryTrait>(
    Ok(())
 }

-#[tracing::instrument(level = "debug", skip_all, target = "search::universe")]
+#[tracing::instrument(level = "trace", skip_all, target = "search::universe")]
 pub fn filtered_universe(
    index: &Index,
    txn: &RoTxn<'_>,
--- a/milli/src/search/new/vector_sort.rs
+++ b/milli/src/search/new/vector_sort.rs
@@ -1,10 +1,11 @@
 use std::iter::FromIterator;

+use ordered_float::OrderedFloat;
 use roaring::RoaringBitmap;

 use super::ranking_rules::{RankingRule, RankingRuleOutput, RankingRuleQueryTrait};
 use crate::score_details::{self, ScoreDetails};
-use crate::vector::{ArroyWrapper, DistributionShift, Embedder};
+use crate::vector::{DistributionShift, Embedder};
 use crate::{DocumentId, Result, SearchContext, SearchLogger};

 pub struct VectorSort<Q: RankingRuleQueryTrait> {
@@ -52,9 +53,14 @@ impl<Q: RankingRuleQueryTrait> VectorSort<Q> {
        vector_candidates: &RoaringBitmap,
    ) -> Result<()> {
        let target = &self.target;
+        let mut results = Vec::new();

-        let reader = ArroyWrapper::new(ctx.index.vector_arroy, self.embedder_index, self.quantized);
-        let results = reader.nns_by_vector(ctx.txn, target, self.limit, Some(vector_candidates))?;
+        for reader in ctx.index.arroy_readers(ctx.txn, self.embedder_index, self.quantized) {
+            let nns_by_vector =
+                reader?.nns_by_vector(ctx.txn, target, self.limit, Some(vector_candidates))?;
+            results.extend(nns_by_vector.into_iter());
+        }
+        results.sort_unstable_by_key(|(_, distance)| OrderedFloat(*distance));
        self.cached_sorted_docids = results.into_iter();

        Ok(())
@@ -83,7 +89,7 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
    }

    #[allow(clippy::only_used_in_recursion)]
-    #[tracing::instrument(level = "debug", skip_all, target = "search::vector_sort")]
+    #[tracing::instrument(level = "trace", skip_all, target = "search::vector_sort")]
    fn next_bucket(
        &mut self,
        ctx: &mut SearchContext<'ctx>,
--- a/milli/src/search/similar.rs
+++ b/milli/src/search/similar.rs
@@ -1,9 +1,10 @@
 use std::sync::Arc;

+use ordered_float::OrderedFloat;
 use roaring::RoaringBitmap;

 use crate::score_details::{self, ScoreDetails};
-use crate::vector::{ArroyWrapper, Embedder};
+use crate::vector::Embedder;
 use crate::{filtered_universe, DocumentId, Filter, Index, Result, SearchResult};

 pub struct Similar<'a> {
@@ -70,13 +71,23 @@ impl<'a> Similar<'a> {
                .get(self.rtxn, &self.embedder_name)?
                .ok_or_else(|| crate::UserError::InvalidEmbedder(self.embedder_name.to_owned()))?;

-        let reader = ArroyWrapper::new(self.index.vector_arroy, embedder_index, self.quantized);
-        let results = reader.nns_by_item(
-            self.rtxn,
-            self.id,
-            self.limit + self.offset + 1,
-            Some(&universe),
-        )?;
+        let mut results = Vec::new();
+
+        for reader in self.index.arroy_readers(self.rtxn, embedder_index, self.quantized) {
+            let nns_by_item = reader?.nns_by_item(
+                self.rtxn,
+                self.id,
+                self.limit + self.offset + 1,
+                Some(&universe),
+            )?;
+            if let Some(mut nns_by_item) = nns_by_item {
+                results.append(&mut nns_by_item);
+            } else {
+                break;
+            }
+        }
+
+        results.sort_unstable_by_key(|(_, distance)| OrderedFloat(*distance));

        let mut documents_ids = Vec::with_capacity(self.limit);
        let mut document_scores = Vec::with_capacity(self.limit);
--- a/milli/src/update/available_documents_ids.rs
+++ b/milli/src/update/available_documents_ids.rs
@@ -3,12 +3,12 @@ use std::ops::RangeInclusive;

 use roaring::bitmap::{IntoIter, RoaringBitmap};

-pub struct AvailableDocumentsIds {
+pub struct AvailableIds {
    iter: Chain<IntoIter, RangeInclusive<u32>>,
 }

-impl AvailableDocumentsIds {
-    pub fn from_documents_ids(docids: &RoaringBitmap) -> AvailableDocumentsIds {
+impl AvailableIds {
+    pub fn new(docids: &RoaringBitmap) -> AvailableIds {
        match docids.max() {
            Some(last_id) => {
                let mut available = RoaringBitmap::from_iter(0..last_id);
@@ -20,17 +20,17 @@ impl AvailableDocumentsIds {
                    None => 1..=0, // empty range iterator
                };

-                AvailableDocumentsIds { iter: available.into_iter().chain(iter) }
+                AvailableIds { iter: available.into_iter().chain(iter) }
            }
            None => {
                let empty = RoaringBitmap::new().into_iter();
-                AvailableDocumentsIds { iter: empty.chain(0..=u32::MAX) }
+                AvailableIds { iter: empty.chain(0..=u32::MAX) }
            }
        }
    }
 }

-impl Iterator for AvailableDocumentsIds {
+impl Iterator for AvailableIds {
    type Item = u32;

    fn next(&mut self) -> Option<Self::Item> {
@@ -45,7 +45,7 @@ mod tests {
    #[test]
    fn empty() {
        let base = RoaringBitmap::new();
-        let left = AvailableDocumentsIds::from_documents_ids(&base);
+        let left = AvailableIds::new(&base);
        let right = 0..=u32::MAX;
        left.zip(right).take(500).for_each(|(l, r)| assert_eq!(l, r));
    }
@@ -58,7 +58,7 @@ mod tests {
        base.insert(100);
        base.insert(405);

-        let left = AvailableDocumentsIds::from_documents_ids(&base);
+        let left = AvailableIds::new(&base);
        let right = (0..=u32::MAX).filter(|&n| n != 0 && n != 10 && n != 100 && n != 405);
        left.zip(right).take(500).for_each(|(l, r)| assert_eq!(l, r));
    }
--- a/milli/src/update/concurrent_available_ids.rs
+++ b/milli/src/update/concurrent_available_ids.rs
@@ -0,0 +1,59 @@
+use std::sync::atomic::{AtomicBool, AtomicU32, AtomicU64, Ordering};
+
+use roaring::RoaringBitmap;
+
+/// A concurrent ID generate that will never return the same ID twice.
+#[derive(Debug)]
+pub struct ConcurrentAvailableIds {
+    /// The current tree node ID we should use if there is no other IDs available.
+    current: AtomicU32,
+    /// The total number of tree node IDs used.
+    used: AtomicU64,
+
+    /// A list of IDs to exhaust before picking IDs from `current`.
+    available: RoaringBitmap,
+    /// The current Nth ID to select in the bitmap.
+    select_in_bitmap: AtomicU32,
+    /// Tells if you should look in the roaring bitmap or if all the IDs are already exhausted.
+    look_into_bitmap: AtomicBool,
+}
+
+impl ConcurrentAvailableIds {
+    /// Creates an ID generator returning unique IDs, avoiding the specified used IDs.
+    pub fn new(used: RoaringBitmap) -> ConcurrentAvailableIds {
+        let last_id = used.max().map_or(0, |id| id + 1);
+        let used_ids = used.len();
+        let available = RoaringBitmap::from_sorted_iter(0..last_id).unwrap() - used;
+
+        ConcurrentAvailableIds {
+            current: AtomicU32::new(last_id),
+            used: AtomicU64::new(used_ids),
+            select_in_bitmap: AtomicU32::new(0),
+            look_into_bitmap: AtomicBool::new(!available.is_empty()),
+            available,
+        }
+    }
+
+    /// Returns a new unique ID and increase the count of IDs used.
+    pub fn next(&self) -> Option<u32> {
+        if self.used.fetch_add(1, Ordering::Relaxed) > u32::MAX as u64 {
+            None
+        } else if self.look_into_bitmap.load(Ordering::Relaxed) {
+            let current = self.select_in_bitmap.fetch_add(1, Ordering::Relaxed);
+            match self.available.select(current) {
+                Some(id) => Some(id),
+                None => {
+                    self.look_into_bitmap.store(false, Ordering::Relaxed);
+                    Some(self.current.fetch_add(1, Ordering::Relaxed))
+                }
+            }
+        } else {
+            Some(self.current.fetch_add(1, Ordering::Relaxed))
+        }
+    }
+
+    /// Returns the number of used ids in total.
+    pub fn used(&self) -> u64 {
+        self.used.load(Ordering::Relaxed)
+    }
+}
--- a/milli/src/update/del_add.rs
+++ b/milli/src/update/del_add.rs
@@ -1,7 +1,7 @@
 use obkv::Key;

 pub type KvWriterDelAdd<W> = obkv::KvWriter<W, DelAdd>;
-pub type KvReaderDelAdd<'a> = obkv::KvReader<'a, DelAdd>;
+pub type KvReaderDelAdd = obkv::KvReader<DelAdd>;

 /// DelAdd defines the new value to add in the database and old value to delete from the database.
 ///
@@ -36,7 +36,7 @@ impl Key for DelAdd {
 /// Addition: put all the values under DelAdd::Addition,
 /// DeletionAndAddition: put all the values under DelAdd::Deletion and DelAdd::Addition,
 pub fn into_del_add_obkv<K: obkv::Key + PartialOrd>(
-    reader: obkv::KvReader<'_, K>,
+    reader: &obkv::KvReader<K>,
    operation: DelAddOperation,
    buffer: &mut Vec<u8>,
 ) -> Result<(), std::io::Error> {
@@ -46,7 +46,7 @@ pub fn into_del_add_obkv<K: obkv::Key + PartialOrd>(
 /// Akin to the [into_del_add_obkv] function but lets you
 /// conditionally define the `DelAdd` variant based on the obkv key.
 pub fn into_del_add_obkv_conditional_operation<K, F>(
-    reader: obkv::KvReader<'_, K>,
+    reader: &obkv::KvReader<K>,
    buffer: &mut Vec<u8>,
    operation: F,
 ) -> std::io::Result<()>
@@ -86,8 +86,8 @@ pub enum DelAddOperation {
 /// putting each deletion obkv's keys under an DelAdd::Deletion
 /// and putting each addition obkv's keys under an DelAdd::Addition
 pub fn del_add_from_two_obkvs<K: obkv::Key + PartialOrd + Ord>(
-    deletion: &obkv::KvReader<'_, K>,
-    addition: &obkv::KvReader<'_, K>,
+    deletion: &obkv::KvReader<K>,
+    addition: &obkv::KvReader<K>,
    buffer: &mut Vec<u8>,
 ) -> Result<(), std::io::Error> {
    use itertools::merge_join_by;
@@ -121,7 +121,7 @@ pub fn del_add_from_two_obkvs<K: obkv::Key + PartialOrd + Ord>(
    writer.finish()
 }

-pub fn is_noop_del_add_obkv(del_add: KvReaderDelAdd<'_>) -> bool {
+pub fn is_noop_del_add_obkv(del_add: &KvReaderDelAdd) -> bool {
    del_add.get(DelAdd::Deletion) == del_add.get(DelAdd::Addition)
 }

@@ -136,5 +136,5 @@ pub fn deladd_serialize_add_side<'a>(
    obkv: &'a [u8],
    _buffer: &mut Vec<u8>,
 ) -> crate::Result<&'a [u8]> {
-    Ok(KvReaderDelAdd::new(obkv).get(DelAdd::Addition).unwrap_or_default())
+    Ok(KvReaderDelAdd::from_slice(obkv).get(DelAdd::Addition).unwrap_or_default())
 }
--- a/milli/src/update/facet/bulk.rs
+++ b/milli/src/update/facet/bulk.rs
@@ -14,7 +14,7 @@ use crate::heed_codec::facet::{
 use crate::heed_codec::BytesRefCodec;
 use crate::update::del_add::{DelAdd, KvReaderDelAdd};
 use crate::update::index_documents::{create_writer, valid_lmdb_key, writer_into_reader};
-use crate::update::MergeFn;
+use crate::update::MergeDeladdCboRoaringBitmaps;
 use crate::{CboRoaringBitmapCodec, CboRoaringBitmapLenCodec, FieldId, Index, Result};

 /// Algorithm to insert elememts into the `facet_id_(string/f64)_docids` databases
@@ -29,7 +29,7 @@ pub struct FacetsUpdateBulk<'i> {
    facet_type: FacetType,
    field_ids: Vec<FieldId>,
    // None if level 0 does not need to be updated
-    delta_data: Option<Merger<BufReader<File>, MergeFn>>,
+    delta_data: Option<Merger<BufReader<File>, MergeDeladdCboRoaringBitmaps>>,
 }

 impl<'i> FacetsUpdateBulk<'i> {
@@ -37,7 +37,7 @@ impl<'i> FacetsUpdateBulk<'i> {
        index: &'i Index,
        field_ids: Vec<FieldId>,
        facet_type: FacetType,
-        delta_data: Merger<BufReader<File>, MergeFn>,
+        delta_data: Merger<BufReader<File>, MergeDeladdCboRoaringBitmaps>,
        group_size: u8,
        min_level_size: u8,
    ) -> FacetsUpdateBulk<'i> {
@@ -90,7 +90,7 @@ impl<'i> FacetsUpdateBulk<'i> {
 /// Implementation of `FacetsUpdateBulk` that is independent of milli's `Index` type
 pub(crate) struct FacetsUpdateBulkInner<R: std::io::Read + std::io::Seek> {
    pub db: heed::Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
-    pub delta_data: Option<Merger<R, MergeFn>>,
+    pub delta_data: Option<Merger<R, MergeDeladdCboRoaringBitmaps>>,
    pub group_size: u8,
    pub min_level_size: u8,
 }
@@ -135,7 +135,7 @@ impl<R: std::io::Read + std::io::Seek> FacetsUpdateBulkInner<R> {
                if !valid_lmdb_key(key) {
                    continue;
                }
-                let value = KvReaderDelAdd::new(value);
+                let value = KvReaderDelAdd::from_slice(value);

                // DB is empty, it is safe to ignore Del operations
                let Some(value) = value.get(DelAdd::Addition) else {
@@ -161,7 +161,7 @@ impl<R: std::io::Read + std::io::Seek> FacetsUpdateBulkInner<R> {
                    continue;
                }

-                let value = KvReaderDelAdd::new(value);
+                let value = KvReaderDelAdd::from_slice(value);

                // the value is a CboRoaringBitmap, but I still need to prepend the
                // group size for level 0 (= 1) to it
--- a/milli/src/update/facet/incremental.rs
+++ b/milli/src/update/facet/incremental.rs
@@ -15,7 +15,7 @@ use crate::heed_codec::BytesRefCodec;
 use crate::search::facet::get_highest_level;
 use crate::update::del_add::DelAdd;
 use crate::update::index_documents::valid_lmdb_key;
-use crate::update::MergeFn;
+use crate::update::MergeDeladdCboRoaringBitmaps;
 use crate::{CboRoaringBitmapCodec, Index, Result};

 /// Enum used as a return value for the facet incremental indexing.
@@ -57,14 +57,14 @@ enum ModificationResult {
 /// `facet_id_(string/f64)_docids` databases.
 pub struct FacetsUpdateIncremental {
    inner: FacetsUpdateIncrementalInner,
-    delta_data: Merger<BufReader<File>, MergeFn>,
+    delta_data: Merger<BufReader<File>, MergeDeladdCboRoaringBitmaps>,
 }

 impl FacetsUpdateIncremental {
    pub fn new(
        index: &Index,
        facet_type: FacetType,
-        delta_data: Merger<BufReader<File>, MergeFn>,
+        delta_data: Merger<BufReader<File>, MergeDeladdCboRoaringBitmaps>,
        group_size: u8,
        min_level_size: u8,
        max_group_size: u8,
@@ -109,7 +109,7 @@ impl FacetsUpdateIncremental {
            }
            current_field_id = Some(key.field_id);

-            let value = KvReader::new(value);
+            let value = KvReader::from_slice(value);
            let docids_to_delete = value
                .get(DelAdd::Deletion)
                .map(CboRoaringBitmapCodec::bytes_decode)
--- a/milli/src/update/facet/mod.rs
+++ b/milli/src/update/facet/mod.rs
@@ -86,12 +86,11 @@ use time::OffsetDateTime;
 use tracing::debug;

 use self::incremental::FacetsUpdateIncremental;
-use super::FacetsUpdateBulk;
+use super::{FacetsUpdateBulk, MergeDeladdBtreesetString, MergeDeladdCboRoaringBitmaps};
 use crate::facet::FacetType;
 use crate::heed_codec::facet::{FacetGroupKey, FacetGroupKeyCodec, FacetGroupValueCodec};
 use crate::heed_codec::BytesRefCodec;
 use crate::update::del_add::{DelAdd, KvReaderDelAdd};
-use crate::update::MergeFn;
 use crate::{try_split_array_at, FieldId, Index, Result};

 pub mod bulk;
@@ -105,8 +104,8 @@ pub struct FacetsUpdate<'i> {
    index: &'i Index,
    database: heed::Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
    facet_type: FacetType,
-    delta_data: Merger<BufReader<File>, MergeFn>,
-    normalized_delta_data: Option<Merger<BufReader<File>, MergeFn>>,
+    delta_data: Merger<BufReader<File>, MergeDeladdCboRoaringBitmaps>,
+    normalized_delta_data: Option<Merger<BufReader<File>, MergeDeladdBtreesetString>>,
    group_size: u8,
    max_group_size: u8,
    min_level_size: u8,
@@ -116,8 +115,8 @@ impl<'i> FacetsUpdate<'i> {
    pub fn new(
        index: &'i Index,
        facet_type: FacetType,
-        delta_data: Merger<BufReader<File>, MergeFn>,
-        normalized_delta_data: Option<Merger<BufReader<File>, MergeFn>>,
+        delta_data: Merger<BufReader<File>, MergeDeladdCboRoaringBitmaps>,
+        normalized_delta_data: Option<Merger<BufReader<File>, MergeDeladdBtreesetString>>,
        data_size: u64,
    ) -> Self {
        let database = match facet_type {
@@ -182,12 +181,12 @@ impl<'i> FacetsUpdate<'i> {

 fn index_facet_search(
    wtxn: &mut heed::RwTxn<'_>,
-    normalized_delta_data: Merger<BufReader<File>, MergeFn>,
+    normalized_delta_data: Merger<BufReader<File>, MergeDeladdBtreesetString>,
    index: &Index,
 ) -> Result<()> {
    let mut iter = normalized_delta_data.into_stream_merger_iter()?;
    while let Some((key_bytes, delta_bytes)) = iter.next()? {
-        let deladd_reader = KvReaderDelAdd::new(delta_bytes);
+        let deladd_reader = KvReaderDelAdd::from_slice(delta_bytes);

        let database_set = index
            .facet_id_normalized_string_strings
@@ -298,8 +297,8 @@ pub(crate) mod test_helpers {
    use crate::search::facet::get_highest_level;
    use crate::snapshot_tests::display_bitmap;
    use crate::update::del_add::{DelAdd, KvWriterDelAdd};
-    use crate::update::index_documents::merge_deladd_cbo_roaring_bitmaps;
-    use crate::update::{FacetsUpdateIncrementalInner, MergeFn};
+    use crate::update::index_documents::MergeDeladdCboRoaringBitmaps;
+    use crate::update::FacetsUpdateIncrementalInner;
    use crate::CboRoaringBitmapCodec;

    /// Utility function to generate a string whose position in a lexicographically
@@ -484,7 +483,7 @@ pub(crate) mod test_helpers {
            }
            writer.finish().unwrap();
            let reader = grenad::Reader::new(std::io::Cursor::new(new_data)).unwrap();
-            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            builder.push(reader.into_cursor().unwrap());
            let merger = builder.build();

--- a/milli/src/update/index_documents/enrich.rs
+++ b/milli/src/update/index_documents/enrich.rs
@@ -47,7 +47,7 @@ pub fn enrich_documents_batch<R: Read + Seek>(
                return match cursor.next_document()? {
                    Some(first_document) => Ok(Err(UserError::MissingDocumentId {
                        primary_key: primary_key.to_string(),
-                        document: obkv_to_object(&first_document, &documents_batch_index)?,
+                        document: obkv_to_object(first_document, &documents_batch_index)?,
                    })),
                    None => unreachable!("Called with reader.is_empty()"),
                };
@@ -106,7 +106,7 @@ pub fn enrich_documents_batch<R: Read + Seek>(
    let mut count = 0;
    while let Some(document) = cursor.next_document()? {
        let document_id = match fetch_or_generate_document_id(
-            &document,
+            document,
            &documents_batch_index,
            primary_key,
            autogenerate_docids,
@@ -145,7 +145,7 @@ pub fn enrich_documents_batch<R: Read + Seek>(
 #[tracing::instrument(level = "trace", skip(uuid_buffer, documents_batch_index, document)
 target = "indexing::documents")]
 fn fetch_or_generate_document_id(
-    document: &obkv::KvReader<'_, FieldId>,
+    document: &obkv::KvReader<FieldId>,
    documents_batch_index: &DocumentsBatchIndex,
    primary_key: PrimaryKey<'_>,
    autogenerate_docids: bool,
--- a/milli/src/update/index_documents/extract/extract_docid_word_positions.rs
+++ b/milli/src/update/index_documents/extract/extract_docid_word_positions.rs
@@ -8,7 +8,7 @@ use obkv::{KvReader, KvWriterU16};
 use roaring::RoaringBitmap;
 use serde_json::Value;

-use super::helpers::{create_sorter, keep_latest_obkv, sorter_into_reader, GrenadParameters};
+use super::helpers::{create_sorter, sorter_into_reader, GrenadParameters, KeepLatestObkv};
 use crate::error::{InternalError, SerializationError};
 use crate::update::del_add::{del_add_from_two_obkvs, DelAdd, KvReaderDelAdd};
 use crate::update::settings::{InnerIndexSettings, InnerIndexSettingsDiff};
@@ -35,7 +35,7 @@ pub fn extract_docid_word_positions<R: io::Read + io::Seek>(
    let mut documents_ids = RoaringBitmap::new();
    let mut docid_word_positions_sorter = create_sorter(
        grenad::SortAlgorithm::Stable,
-        keep_latest_obkv,
+        KeepLatestObkv,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -80,10 +80,10 @@ pub fn extract_docid_word_positions<R: io::Read + io::Seek>(
            .try_into()
            .map(u32::from_be_bytes)
            .map_err(|_| SerializationError::InvalidNumberSerialization)?;
-        let obkv = KvReader::<FieldId>::new(value);
+        let obkv = KvReader::<FieldId>::from_slice(value);

        // if the searchable fields didn't change, skip the searchable indexing for this document.
-        if !force_reindexing && !searchable_fields_changed(&obkv, settings_diff) {
+        if !force_reindexing && !searchable_fields_changed(obkv, settings_diff) {
            continue;
        }

@@ -98,7 +98,7 @@ pub fn extract_docid_word_positions<R: io::Read + io::Seek>(
            || {
                // deletions
                tokens_from_document(
-                    &obkv,
+                    obkv,
                    &settings_diff.old,
                    &del_tokenizer,
                    max_positions_per_attributes,
@@ -109,7 +109,7 @@ pub fn extract_docid_word_positions<R: io::Read + io::Seek>(
            || {
                // additions
                tokens_from_document(
-                    &obkv,
+                    obkv,
                    &settings_diff.new,
                    &add_tokenizer,
                    max_positions_per_attributes,
@@ -126,13 +126,13 @@ pub fn extract_docid_word_positions<R: io::Read + io::Seek>(
        // transforming two KV<FieldId, KV<u16, String>> into one KV<FieldId, KV<DelAdd, KV<u16, String>>>
        value_buffer.clear();
        del_add_from_two_obkvs(
-            &KvReader::<FieldId>::new(del_obkv),
-            &KvReader::<FieldId>::new(add_obkv),
+            KvReader::<FieldId>::from_slice(del_obkv),
+            KvReader::<FieldId>::from_slice(add_obkv),
            &mut value_buffer,
        )?;

        // write each KV<DelAdd, KV<u16, String>> into the sorter, field by field.
-        let obkv = KvReader::<FieldId>::new(&value_buffer);
+        let obkv = KvReader::<FieldId>::from_slice(&value_buffer);
        for (field_id, value) in obkv.iter() {
            key_buffer.truncate(mem::size_of::<u32>());
            key_buffer.extend_from_slice(&field_id.to_be_bytes());
@@ -146,13 +146,13 @@ pub fn extract_docid_word_positions<R: io::Read + io::Seek>(

 /// Check if any searchable fields of a document changed.
 fn searchable_fields_changed(
-    obkv: &KvReader<'_, FieldId>,
+    obkv: &KvReader<FieldId>,
    settings_diff: &InnerIndexSettingsDiff,
 ) -> bool {
    let searchable_fields = &settings_diff.new.searchable_fields_ids;
    for (field_id, field_bytes) in obkv.iter() {
        if searchable_fields.contains(&field_id) {
-            let del_add = KvReaderDelAdd::new(field_bytes);
+            let del_add = KvReaderDelAdd::from_slice(field_bytes);
            match (del_add.get(DelAdd::Deletion), del_add.get(DelAdd::Addition)) {
                // if both fields are None, check the next field.
                (None, None) => (),
@@ -189,7 +189,7 @@ fn tokenizer_builder<'a>(

 /// Extract words mapped with their positions of a document.
 fn tokens_from_document<'a>(
-    obkv: &KvReader<'a, FieldId>,
+    obkv: &'a KvReader<FieldId>,
    settings: &InnerIndexSettings,
    tokenizer: &Tokenizer<'_>,
    max_positions_per_attributes: u32,
@@ -202,7 +202,7 @@ fn tokens_from_document<'a>(
        // if field is searchable.
        if settings.searchable_fields_ids.contains(&field_id) {
            // extract deletion or addition only.
-            if let Some(field_bytes) = KvReaderDelAdd::new(field_bytes).get(del_add) {
+            if let Some(field_bytes) = KvReaderDelAdd::from_slice(field_bytes).get(del_add) {
                // parse json.
                let value =
                    serde_json::from_slice(field_bytes).map_err(InternalError::SerdeJson)?;
--- a/milli/src/update/index_documents/extract/extract_facet_number_docids.rs
+++ b/milli/src/update/index_documents/extract/extract_facet_number_docids.rs
@@ -4,7 +4,7 @@ use std::io::{self, BufReader};
 use heed::{BytesDecode, BytesEncode};

 use super::helpers::{
-    create_sorter, merge_deladd_cbo_roaring_bitmaps, sorter_into_reader, GrenadParameters,
+    create_sorter, sorter_into_reader, GrenadParameters, MergeDeladdCboRoaringBitmaps,
 };
 use crate::heed_codec::facet::{
    FacetGroupKey, FacetGroupKeyCodec, FieldDocIdFacetF64Codec, OrderedF64Codec,
@@ -27,7 +27,7 @@ pub fn extract_facet_number_docids<R: io::Read + io::Seek>(

    let mut facet_number_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Unstable,
-        merge_deladd_cbo_roaring_bitmaps,
+        MergeDeladdCboRoaringBitmaps,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -45,7 +45,7 @@ pub fn extract_facet_number_docids<R: io::Read + io::Seek>(

        buffer.clear();
        let mut obkv = KvWriterDelAdd::new(&mut buffer);
-        for (deladd_key, _) in KvReaderDelAdd::new(deladd_obkv_bytes).iter() {
+        for (deladd_key, _) in KvReaderDelAdd::from_slice(deladd_obkv_bytes).iter() {
            obkv.insert(deladd_key, document_id.to_ne_bytes())?;
        }
        obkv.finish()?;
--- a/milli/src/update/index_documents/extract/extract_facet_string_docids.rs
+++ b/milli/src/update/index_documents/extract/extract_facet_string_docids.rs
@@ -15,7 +15,7 @@ use crate::heed_codec::{BEU16StrCodec, StrRefCodec};
 use crate::localized_attributes_rules::LocalizedFieldIds;
 use crate::update::del_add::{DelAdd, KvReaderDelAdd, KvWriterDelAdd};
 use crate::update::index_documents::helpers::{
-    merge_deladd_btreeset_string, merge_deladd_cbo_roaring_bitmaps,
+    MergeDeladdBtreesetString, MergeDeladdCboRoaringBitmaps,
 };
 use crate::update::settings::InnerIndexSettingsDiff;
 use crate::{FieldId, Result, MAX_FACET_VALUE_LENGTH};
@@ -56,7 +56,7 @@ fn extract_facet_string_docids_document_update<R: io::Read + io::Seek>(

    let mut facet_string_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Stable,
-        merge_deladd_cbo_roaring_bitmaps,
+        MergeDeladdCboRoaringBitmaps,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -65,7 +65,7 @@ fn extract_facet_string_docids_document_update<R: io::Read + io::Seek>(

    let mut normalized_facet_string_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Stable,
-        merge_deladd_btreeset_string,
+        MergeDeladdBtreesetString,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -75,7 +75,7 @@ fn extract_facet_string_docids_document_update<R: io::Read + io::Seek>(
    let mut buffer = Vec::new();
    let mut cursor = docid_fid_facet_string.into_cursor()?;
    while let Some((key, deladd_original_value_bytes)) = cursor.move_on_next()? {
-        let deladd_reader = KvReaderDelAdd::new(deladd_original_value_bytes);
+        let deladd_reader = KvReaderDelAdd::from_slice(deladd_original_value_bytes);

        let is_same_value = deladd_reader.get(DelAdd::Deletion).is_some()
            && deladd_reader.get(DelAdd::Addition).is_some();
@@ -144,7 +144,7 @@ fn extract_facet_string_docids_settings<R: io::Read + io::Seek>(

    let mut facet_string_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Stable,
-        merge_deladd_cbo_roaring_bitmaps,
+        MergeDeladdCboRoaringBitmaps,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -153,7 +153,7 @@ fn extract_facet_string_docids_settings<R: io::Read + io::Seek>(

    let mut normalized_facet_string_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Stable,
-        merge_deladd_btreeset_string,
+        MergeDeladdBtreesetString,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -163,7 +163,7 @@ fn extract_facet_string_docids_settings<R: io::Read + io::Seek>(
    let mut buffer = Vec::new();
    let mut cursor = docid_fid_facet_string.into_cursor()?;
    while let Some((key, deladd_original_value_bytes)) = cursor.move_on_next()? {
-        let deladd_reader = KvReaderDelAdd::new(deladd_original_value_bytes);
+        let deladd_reader = KvReaderDelAdd::from_slice(deladd_original_value_bytes);

        let is_same_value = deladd_reader.get(DelAdd::Deletion).is_some()
            && deladd_reader.get(DelAdd::Addition).is_some();
--- a/milli/src/update/index_documents/extract/extract_fid_docid_facet_values.rs
+++ b/milli/src/update/index_documents/extract/extract_fid_docid_facet_values.rs
@@ -1,10 +1,8 @@
-use std::borrow::Cow;
 use std::collections::{BTreeMap, BTreeSet};
 use std::convert::TryInto;
 use std::fs::File;
 use std::io::{self, BufReader};
 use std::mem::size_of;
-use std::result::Result as StdResult;

 use bytemuck::bytes_of;
 use grenad::Sorter;
@@ -15,13 +13,13 @@ use roaring::RoaringBitmap;
 use serde_json::{from_slice, Value};
 use FilterableValues::{Empty, Null, Values};

-use super::helpers::{create_sorter, keep_first, sorter_into_reader, GrenadParameters};
+use super::helpers::{create_sorter, sorter_into_reader, GrenadParameters, KeepFirst};
 use crate::error::InternalError;
 use crate::facet::value_encoding::f64_into_bytes;
 use crate::update::del_add::{DelAdd, KvReaderDelAdd, KvWriterDelAdd};
 use crate::update::index_documents::{create_writer, writer_into_reader};
 use crate::update::settings::InnerIndexSettingsDiff;
-use crate::{CboRoaringBitmapCodec, DocumentId, Error, FieldId, Result, MAX_FACET_VALUE_LENGTH};
+use crate::{CboRoaringBitmapCodec, DocumentId, FieldId, Result, MAX_FACET_VALUE_LENGTH};

 /// The length of the elements that are always in the buffer when inserting new values.
 const TRUNCATE_SIZE: usize = size_of::<FieldId>() + size_of::<DocumentId>();
@@ -50,7 +48,7 @@ pub fn extract_fid_docid_facet_values<R: io::Read + io::Seek>(

    let mut fid_docid_facet_numbers_sorter = create_sorter(
        grenad::SortAlgorithm::Stable,
-        keep_first,
+        KeepFirst,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -59,7 +57,7 @@ pub fn extract_fid_docid_facet_values<R: io::Read + io::Seek>(

    let mut fid_docid_facet_strings_sorter = create_sorter(
        grenad::SortAlgorithm::Stable,
-        keep_first,
+        KeepFirst,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -83,10 +81,10 @@ pub fn extract_fid_docid_facet_values<R: io::Read + io::Seek>(
    if !settings_diff.settings_update_only || old_faceted_fids != new_faceted_fids {
        let mut cursor = obkv_documents.into_cursor()?;
        while let Some((docid_bytes, value)) = cursor.move_on_next()? {
-            let obkv = obkv::KvReader::new(value);
+            let obkv = obkv::KvReader::from_slice(value);
            let get_document_json_value = move |field_id, side| {
                obkv.get(field_id)
-                    .map(KvReaderDelAdd::new)
+                    .map(KvReaderDelAdd::from_slice)
                    .and_then(|kv| kv.get(side))
                    .map(from_slice)
                    .transpose()
@@ -330,15 +328,12 @@ fn truncate_str(s: &str) -> &str {

 /// Computes the diff between both Del and Add numbers and
 /// only inserts the parts that differ in the sorter.
-fn insert_numbers_diff<MF>(
-    fid_docid_facet_numbers_sorter: &mut Sorter<MF>,
+fn insert_numbers_diff(
+    fid_docid_facet_numbers_sorter: &mut Sorter<KeepFirst>,
    key_buffer: &mut Vec<u8>,
    mut del_numbers: Vec<f64>,
    mut add_numbers: Vec<f64>,
-) -> Result<()>
-where
-    MF: for<'a> Fn(&[u8], &[Cow<'a, [u8]>]) -> StdResult<Cow<'a, [u8]>, Error>,
-{
+) -> Result<()> {
    // We sort and dedup the float numbers
    del_numbers.sort_unstable_by_key(|f| OrderedFloat(*f));
    add_numbers.sort_unstable_by_key(|f| OrderedFloat(*f));
@@ -390,15 +385,12 @@ where

 /// Computes the diff between both Del and Add strings and
 /// only inserts the parts that differ in the sorter.
-fn insert_strings_diff<MF>(
-    fid_docid_facet_strings_sorter: &mut Sorter<MF>,
+fn insert_strings_diff(
+    fid_docid_facet_strings_sorter: &mut Sorter<KeepFirst>,
    key_buffer: &mut Vec<u8>,
    mut del_strings: Vec<(String, String)>,
    mut add_strings: Vec<(String, String)>,
-) -> Result<()>
-where
-    MF: for<'a> Fn(&[u8], &[Cow<'a, [u8]>]) -> StdResult<Cow<'a, [u8]>, Error>,
-{
+) -> Result<()> {
    // We sort and dedup the normalized and original strings
    del_strings.sort_unstable();
    add_strings.sort_unstable();
--- a/milli/src/update/index_documents/extract/extract_fid_word_count_docids.rs
+++ b/milli/src/update/index_documents/extract/extract_fid_word_count_docids.rs
@@ -4,8 +4,8 @@ use std::io::{self, BufReader};
 use obkv::KvReaderU16;

 use super::helpers::{
-    create_sorter, merge_deladd_cbo_roaring_bitmaps, sorter_into_reader, try_split_array_at,
-    GrenadParameters,
+    create_sorter, sorter_into_reader, try_split_array_at, GrenadParameters,
+    MergeDeladdCboRoaringBitmaps,
 };
 use crate::error::SerializationError;
 use crate::index::db_name::DOCID_WORD_POSITIONS;
@@ -30,7 +30,7 @@ pub fn extract_fid_word_count_docids<R: io::Read + io::Seek>(

    let mut fid_word_count_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Unstable,
-        merge_deladd_cbo_roaring_bitmaps,
+        MergeDeladdCboRoaringBitmaps,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -45,19 +45,23 @@ pub fn extract_fid_word_count_docids<R: io::Read + io::Seek>(
            .ok_or(SerializationError::Decoding { db_name: Some(DOCID_WORD_POSITIONS) })?;
        let document_id = u32::from_be_bytes(document_id_bytes);

-        let del_add_reader = KvReaderDelAdd::new(value);
+        let del_add_reader = KvReaderDelAdd::from_slice(value);
        let deletion = del_add_reader
            // get deleted words
            .get(DelAdd::Deletion)
            // count deleted words
-            .map(|deletion| KvReaderU16::new(deletion).iter().take(MAX_COUNTED_WORDS + 1).count())
+            .map(|deletion| {
+                KvReaderU16::from_slice(deletion).iter().take(MAX_COUNTED_WORDS + 1).count()
+            })
            // keep the count if under or equal to MAX_COUNTED_WORDS
            .filter(|&word_count| word_count <= MAX_COUNTED_WORDS);
        let addition = del_add_reader
            // get added words
            .get(DelAdd::Addition)
            // count added words
-            .map(|addition| KvReaderU16::new(addition).iter().take(MAX_COUNTED_WORDS + 1).count())
+            .map(|addition| {
+                KvReaderU16::from_slice(addition).iter().take(MAX_COUNTED_WORDS + 1).count()
+            })
            // keep the count if under or equal to MAX_COUNTED_WORDS
            .filter(|&word_count| word_count <= MAX_COUNTED_WORDS);

--- a/milli/src/update/index_documents/extract/extract_geo_points.rs
+++ b/milli/src/update/index_documents/extract/extract_geo_points.rs
@@ -29,22 +29,20 @@ pub fn extract_geo_points<R: io::Read + io::Seek>(

    let mut cursor = obkv_documents.into_cursor()?;
    while let Some((docid_bytes, value)) = cursor.move_on_next()? {
-        let obkv = obkv::KvReader::new(value);
+        let obkv = obkv::KvReader::from_slice(value);
        // since we only need the primary key when we throw an error
        // we create this getter to lazily get it when needed
        let document_id = || -> Value {
-            let reader = KvReaderDelAdd::new(obkv.get(primary_key_id).unwrap());
+            let reader = KvReaderDelAdd::from_slice(obkv.get(primary_key_id).unwrap());
            let document_id =
                reader.get(DelAdd::Deletion).or(reader.get(DelAdd::Addition)).unwrap();
            serde_json::from_slice(document_id).unwrap()
        };

        // extract old version
-        let del_lat_lng =
-            extract_lat_lng(&obkv, &settings_diff.old, DelAdd::Deletion, document_id)?;
+        let del_lat_lng = extract_lat_lng(obkv, &settings_diff.old, DelAdd::Deletion, document_id)?;
        // extract new version
-        let add_lat_lng =
-            extract_lat_lng(&obkv, &settings_diff.new, DelAdd::Addition, document_id)?;
+        let add_lat_lng = extract_lat_lng(obkv, &settings_diff.new, DelAdd::Addition, document_id)?;

        if del_lat_lng != add_lat_lng {
            let mut obkv = KvWriterDelAdd::memory();
@@ -68,15 +66,17 @@ pub fn extract_geo_points<R: io::Read + io::Seek>(

 /// Extract the finite floats lat and lng from two bytes slices.
 fn extract_lat_lng(
-    document: &obkv::KvReader<'_, FieldId>,
+    document: &obkv::KvReader<FieldId>,
    settings: &InnerIndexSettings,
    deladd: DelAdd,
    document_id: impl Fn() -> Value,
 ) -> Result<Option<[f64; 2]>> {
    match settings.geo_fields_ids {
        Some((lat_fid, lng_fid)) => {
-            let lat = document.get(lat_fid).map(KvReaderDelAdd::new).and_then(|r| r.get(deladd));
-            let lng = document.get(lng_fid).map(KvReaderDelAdd::new).and_then(|r| r.get(deladd));
+            let lat =
+                document.get(lat_fid).map(KvReaderDelAdd::from_slice).and_then(|r| r.get(deladd));
+            let lng =
+                document.get(lng_fid).map(KvReaderDelAdd::from_slice).and_then(|r| r.get(deladd));
            let (lat, lng) = match (lat, lng) {
                (Some(lat), Some(lng)) => (lat, lng),
                (Some(_), None) => {
--- a/milli/src/update/index_documents/extract/extract_vector_points.rs
+++ b/milli/src/update/index_documents/extract/extract_vector_points.rs
@@ -313,7 +313,7 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
        debug_assert!(from_utf8(external_id_bytes).is_ok());
        let docid = DocumentId::from_be_bytes(docid_bytes);

-        let obkv = obkv::KvReader::new(value);
+        let obkv = obkv::KvReader::from_slice(value);
        key_buffer.clear();
        key_buffer.extend_from_slice(docid_bytes.as_slice());

@@ -481,7 +481,7 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
 #[allow(clippy::too_many_arguments)] // feel free to find efficient way to factor arguments
 fn extract_vector_document_diff(
    docid: DocumentId,
-    obkv: obkv::KvReader<'_, FieldId>,
+    obkv: &obkv::KvReader<FieldId>,
    prompt: &Prompt,
    (add_to_user_provided, remove_from_user_provided): (&mut RoaringBitmap, &mut RoaringBitmap),
    (old, new): (VectorState, VectorState),
@@ -526,7 +526,7 @@ fn extract_vector_document_diff(
            // Do we keep this document?
            let document_is_kept = obkv
                .iter()
-                .map(|(_, deladd)| KvReaderDelAdd::new(deladd))
+                .map(|(_, deladd)| KvReaderDelAdd::from_slice(deladd))
                .any(|deladd| deladd.get(DelAdd::Addition).is_some());

            if document_is_kept {
@@ -562,7 +562,7 @@ fn extract_vector_document_diff(
            // Do we keep this document?
            let document_is_kept = obkv
                .iter()
-                .map(|(_, deladd)| KvReaderDelAdd::new(deladd))
+                .map(|(_, deladd)| KvReaderDelAdd::from_slice(deladd))
                .any(|deladd| deladd.get(DelAdd::Addition).is_some());
            if document_is_kept {
                if embedder_is_manual {
@@ -588,7 +588,7 @@ fn extract_vector_document_diff(
            // Do we keep this document?
            let document_is_kept = obkv
                .iter()
-                .map(|(_, deladd)| KvReaderDelAdd::new(deladd))
+                .map(|(_, deladd)| KvReaderDelAdd::from_slice(deladd))
                .any(|deladd| deladd.get(DelAdd::Addition).is_some());
            if document_is_kept {
                // if the new version of documents has the vectors in the DB,
@@ -606,7 +606,7 @@ fn extract_vector_document_diff(
 }

 fn regenerate_if_prompt_changed(
-    obkv: obkv::KvReader<'_, FieldId>,
+    obkv: &obkv::KvReader<FieldId>,
    (old_prompt, new_prompt): (&Prompt, &Prompt),
    (old_fields_ids_map, new_fields_ids_map): (
        &FieldsIdsMapWithMetadata,
@@ -624,7 +624,7 @@ fn regenerate_if_prompt_changed(
 }

 fn regenerate_prompt(
-    obkv: obkv::KvReader<'_, FieldId>,
+    obkv: &obkv::KvReader<FieldId>,
    prompt: &Prompt,
    new_fields_ids_map: &FieldsIdsMapWithMetadata,
 ) -> Result<VectorStateDelta> {
--- a/milli/src/update/index_documents/extract/extract_word_docids.rs
+++ b/milli/src/update/index_documents/extract/extract_word_docids.rs
@@ -7,8 +7,8 @@ use obkv::KvReaderU16;
 use roaring::RoaringBitmap;

 use super::helpers::{
-    create_sorter, create_writer, merge_deladd_cbo_roaring_bitmaps, try_split_array_at,
-    writer_into_reader, GrenadParameters,
+    create_sorter, create_writer, try_split_array_at, writer_into_reader, GrenadParameters,
+    MergeDeladdCboRoaringBitmaps,
 };
 use crate::error::SerializationError;
 use crate::heed_codec::StrBEU16Codec;
@@ -16,7 +16,6 @@ use crate::index::db_name::DOCID_WORD_POSITIONS;
 use crate::update::del_add::{is_noop_del_add_obkv, DelAdd, KvReaderDelAdd, KvWriterDelAdd};
 use crate::update::index_documents::helpers::sorter_into_reader;
 use crate::update::settings::InnerIndexSettingsDiff;
-use crate::update::MergeFn;
 use crate::{CboRoaringBitmapCodec, DocumentId, FieldId, Result};

 /// Extracts the word and the documents ids where this word appear.
@@ -40,7 +39,7 @@ pub fn extract_word_docids<R: io::Read + io::Seek>(

    let mut word_fid_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Unstable,
-        merge_deladd_cbo_roaring_bitmaps,
+        MergeDeladdCboRoaringBitmaps,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -58,17 +57,17 @@ pub fn extract_word_docids<R: io::Read + io::Seek>(
        let document_id = u32::from_be_bytes(document_id_bytes);
        let fid = u16::from_be_bytes(fid_bytes);

-        let del_add_reader = KvReaderDelAdd::new(value);
+        let del_add_reader = KvReaderDelAdd::from_slice(value);
        // extract all unique words to remove.
        if let Some(deletion) = del_add_reader.get(DelAdd::Deletion) {
-            for (_pos, word) in KvReaderU16::new(deletion).iter() {
+            for (_pos, word) in KvReaderU16::from_slice(deletion).iter() {
                del_words.insert(word.to_vec());
            }
        }

        // extract all unique additional words.
        if let Some(addition) = del_add_reader.get(DelAdd::Addition) {
-            for (_pos, word) in KvReaderU16::new(addition).iter() {
+            for (_pos, word) in KvReaderU16::from_slice(addition).iter() {
                add_words.insert(word.to_vec());
            }
        }
@@ -94,7 +93,7 @@ pub fn extract_word_docids<R: io::Read + io::Seek>(

    let mut word_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Unstable,
-        merge_deladd_cbo_roaring_bitmaps,
+        MergeDeladdCboRoaringBitmaps,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -103,7 +102,7 @@ pub fn extract_word_docids<R: io::Read + io::Seek>(

    let mut exact_word_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Unstable,
-        merge_deladd_cbo_roaring_bitmaps,
+        MergeDeladdCboRoaringBitmaps,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -115,7 +114,7 @@ pub fn extract_word_docids<R: io::Read + io::Seek>(
    // NOTE: replacing sorters by bitmap merging is less efficient, so, use sorters.
    while let Some((key, value)) = iter.next()? {
        // only keep the value if their is a change to apply in the DB.
-        if !is_noop_del_add_obkv(KvReaderDelAdd::new(value)) {
+        if !is_noop_del_add_obkv(KvReaderDelAdd::from_slice(value)) {
            word_fid_docids_writer.insert(key, value)?;
        }

@@ -123,7 +122,7 @@ pub fn extract_word_docids<R: io::Read + io::Seek>(
            .map_err(|_| SerializationError::Decoding { db_name: Some(DOCID_WORD_POSITIONS) })?;

        // merge all deletions
-        let obkv = KvReaderDelAdd::new(value);
+        let obkv = KvReaderDelAdd::from_slice(value);
        if let Some(value) = obkv.get(DelAdd::Deletion) {
            let delete_from_exact = settings_diff.old.exact_attributes.contains(&fid);
            buffer.clear();
@@ -163,7 +162,7 @@ fn words_into_sorter(
    key_buffer: &mut Vec<u8>,
    del_words: &BTreeSet<Vec<u8>>,
    add_words: &BTreeSet<Vec<u8>>,
-    word_fid_docids_sorter: &mut grenad::Sorter<MergeFn>,
+    word_fid_docids_sorter: &mut grenad::Sorter<MergeDeladdCboRoaringBitmaps>,
 ) -> Result<()> {
    use itertools::merge_join_by;
    use itertools::EitherOrBoth::{Both, Left, Right};
--- a/milli/src/update/index_documents/extract/extract_word_pair_proximity_docids.rs
+++ b/milli/src/update/index_documents/extract/extract_word_pair_proximity_docids.rs
@@ -6,8 +6,8 @@ use std::{cmp, io};
 use obkv::KvReaderU16;

 use super::helpers::{
-    create_sorter, create_writer, merge_deladd_cbo_roaring_bitmaps, try_split_array_at,
-    writer_into_reader, GrenadParameters, MergeFn,
+    create_sorter, create_writer, try_split_array_at, writer_into_reader, GrenadParameters,
+    MergeDeladdCboRoaringBitmaps,
 };
 use crate::error::SerializationError;
 use crate::index::db_name::DOCID_WORD_POSITIONS;
@@ -44,7 +44,7 @@ pub fn extract_word_pair_proximity_docids<R: io::Read + io::Seek>(
        .map(|_| {
            create_sorter(
                grenad::SortAlgorithm::Unstable,
-                merge_deladd_cbo_roaring_bitmaps,
+                MergeDeladdCboRoaringBitmaps,
                indexer.chunk_compression_type,
                indexer.chunk_compression_level,
                indexer.max_nb_chunks,
@@ -92,8 +92,8 @@ pub fn extract_word_pair_proximity_docids<R: io::Read + io::Seek>(
                }

                // deletions
-                if let Some(deletion) = KvReaderDelAdd::new(value).get(DelAdd::Deletion) {
-                    for (position, word) in KvReaderU16::new(deletion).iter() {
+                if let Some(deletion) = KvReaderDelAdd::from_slice(value).get(DelAdd::Deletion) {
+                    for (position, word) in KvReaderU16::from_slice(deletion).iter() {
                        // drain the proximity window until the head word is considered close to the word we are inserting.
                        while del_word_positions.front().map_or(false, |(_w, p)| {
                            index_proximity(*p as u32, position as u32) >= MAX_DISTANCE
@@ -125,8 +125,8 @@ pub fn extract_word_pair_proximity_docids<R: io::Read + io::Seek>(
                }

                // additions
-                if let Some(addition) = KvReaderDelAdd::new(value).get(DelAdd::Addition) {
-                    for (position, word) in KvReaderU16::new(addition).iter() {
+                if let Some(addition) = KvReaderDelAdd::from_slice(value).get(DelAdd::Addition) {
+                    for (position, word) in KvReaderU16::from_slice(addition).iter() {
                        // drain the proximity window until the head word is considered close to the word we are inserting.
                        while add_word_positions.front().map_or(false, |(_w, p)| {
                            index_proximity(*p as u32, position as u32) >= MAX_DISTANCE
@@ -197,7 +197,7 @@ fn document_word_positions_into_sorter(
    document_id: DocumentId,
    del_word_pair_proximity: &BTreeMap<(String, String), u8>,
    add_word_pair_proximity: &BTreeMap<(String, String), u8>,
-    word_pair_proximity_docids_sorters: &mut [grenad::Sorter<MergeFn>],
+    word_pair_proximity_docids_sorters: &mut [grenad::Sorter<MergeDeladdCboRoaringBitmaps>],
 ) -> Result<()> {
    use itertools::merge_join_by;
    use itertools::EitherOrBoth::{Both, Left, Right};
--- a/milli/src/update/index_documents/extract/extract_word_position_docids.rs
+++ b/milli/src/update/index_documents/extract/extract_word_position_docids.rs
@@ -5,14 +5,13 @@ use std::io::{self, BufReader};
 use obkv::KvReaderU16;

 use super::helpers::{
-    create_sorter, merge_deladd_cbo_roaring_bitmaps, sorter_into_reader, try_split_array_at,
-    GrenadParameters,
+    create_sorter, sorter_into_reader, try_split_array_at, GrenadParameters,
+    MergeDeladdCboRoaringBitmaps,
 };
 use crate::error::SerializationError;
 use crate::index::db_name::DOCID_WORD_POSITIONS;
 use crate::update::del_add::{DelAdd, KvReaderDelAdd, KvWriterDelAdd};
 use crate::update::settings::InnerIndexSettingsDiff;
-use crate::update::MergeFn;
 use crate::{bucketed_position, DocumentId, Result};

 /// Extracts the word positions and the documents ids where this word appear.
@@ -29,7 +28,7 @@ pub fn extract_word_position_docids<R: io::Read + io::Seek>(

    let mut word_position_docids_sorter = create_sorter(
        grenad::SortAlgorithm::Unstable,
-        merge_deladd_cbo_roaring_bitmaps,
+        MergeDeladdCboRoaringBitmaps,
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
        indexer.max_nb_chunks,
@@ -60,10 +59,10 @@ pub fn extract_word_position_docids<R: io::Read + io::Seek>(

        current_document_id = Some(document_id);

-        let del_add_reader = KvReaderDelAdd::new(value);
+        let del_add_reader = KvReaderDelAdd::from_slice(value);
        // extract all unique words to remove.
        if let Some(deletion) = del_add_reader.get(DelAdd::Deletion) {
-            for (position, word_bytes) in KvReaderU16::new(deletion).iter() {
+            for (position, word_bytes) in KvReaderU16::from_slice(deletion).iter() {
                let position = bucketed_position(position);
                del_word_positions.insert((position, word_bytes.to_vec()));
            }
@@ -71,7 +70,7 @@ pub fn extract_word_position_docids<R: io::Read + io::Seek>(

        // extract all unique additional words.
        if let Some(addition) = del_add_reader.get(DelAdd::Addition) {
-            for (position, word_bytes) in KvReaderU16::new(addition).iter() {
+            for (position, word_bytes) in KvReaderU16::from_slice(addition).iter() {
                let position = bucketed_position(position);
                add_word_positions.insert((position, word_bytes.to_vec()));
            }
@@ -100,7 +99,7 @@ fn words_position_into_sorter(
    key_buffer: &mut Vec<u8>,
    del_word_positions: &BTreeSet<(u16, Vec<u8>)>,
    add_word_positions: &BTreeSet<(u16, Vec<u8>)>,
-    word_position_docids_sorter: &mut grenad::Sorter<MergeFn>,
+    word_position_docids_sorter: &mut grenad::Sorter<MergeDeladdCboRoaringBitmaps>,
 ) -> Result<()> {
    use itertools::merge_join_by;
    use itertools::EitherOrBoth::{Both, Left, Right};
--- a/milli/src/update/index_documents/helpers/grenad_helpers.rs
+++ b/milli/src/update/index_documents/helpers/grenad_helpers.rs
@@ -1,11 +1,10 @@
-use std::borrow::Cow;
 use std::fs::File;
 use std::io::{self, BufReader, BufWriter, Seek};

-use grenad::{CompressionType, Sorter};
+use grenad::{CompressionType, MergeFunction, Sorter};
 use heed::types::Bytes;

-use super::{ClonableMmap, MergeFn};
+use super::ClonableMmap;
 use crate::update::index_documents::valid_lmdb_key;
 use crate::Result;

@@ -31,14 +30,14 @@ pub fn create_writer<R: io::Write>(
 /// A helper function that creates a grenad sorter
 /// with the given parameters. The max memory is
 /// clamped to something reasonable.
-pub fn create_sorter(
+pub fn create_sorter<MF: MergeFunction>(
    sort_algorithm: grenad::SortAlgorithm,
-    merge: MergeFn,
+    merge: MF,
    chunk_compression_type: grenad::CompressionType,
    chunk_compression_level: Option<u32>,
    max_nb_chunks: Option<usize>,
    max_memory: Option<usize>,
-) -> grenad::Sorter<MergeFn> {
+) -> grenad::Sorter<MF> {
    let mut builder = grenad::Sorter::builder(merge);
    builder.chunk_compression_type(chunk_compression_type);
    if let Some(level) = chunk_compression_level {
@@ -57,10 +56,14 @@ pub fn create_sorter(
 }

 #[tracing::instrument(level = "trace", skip_all, target = "indexing::grenad")]
-pub fn sorter_into_reader(
-    sorter: grenad::Sorter<MergeFn>,
+pub fn sorter_into_reader<MF>(
+    sorter: grenad::Sorter<MF>,
    indexer: GrenadParameters,
-) -> Result<grenad::Reader<BufReader<File>>> {
+) -> Result<grenad::Reader<BufReader<File>>>
+where
+    MF: MergeFunction,
+    crate::Error: From<MF::Error>,
+{
    let mut writer = create_writer(
        indexer.chunk_compression_type,
        indexer.chunk_compression_level,
@@ -169,8 +172,8 @@ pub fn grenad_obkv_into_chunks<R: io::Read + io::Seek>(
 /// Write provided sorter in database using serialize_value function.
 /// merge_values function is used if an entry already exist in the database.
 #[tracing::instrument(level = "trace", skip_all, target = "indexing::grenad")]
-pub fn write_sorter_into_database<K, V, FS, FM>(
-    sorter: Sorter<MergeFn>,
+pub fn write_sorter_into_database<K, V, FS, FM, MF>(
+    sorter: Sorter<MF>,
    database: &heed::Database<K, V>,
    wtxn: &mut heed::RwTxn<'_>,
    index_is_empty: bool,
@@ -180,6 +183,8 @@ pub fn write_sorter_into_database<K, V, FS, FM>(
 where
    FS: for<'a> Fn(&'a [u8], &'a mut Vec<u8>) -> Result<&'a [u8]>,
    FM: for<'a> Fn(&[u8], &[u8], &'a mut Vec<u8>) -> Result<Option<&'a [u8]>>,
+    MF: MergeFunction,
+    crate::Error: From<MF::Error>,
 {
    let mut buffer = Vec::new();
    let database = database.remap_types::<Bytes, Bytes>();
@@ -207,8 +212,3 @@ where

    Ok(())
 }
-
-/// Used when trying to merge readers, but you don't actually care about the values.
-pub fn merge_ignore_values<'a>(_key: &[u8], _values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
-    Ok(Cow::Owned(Vec::new()))
-}
--- a/milli/src/update/index_documents/helpers/merge_functions.rs
+++ b/milli/src/update/index_documents/helpers/merge_functions.rs
@@ -3,6 +3,8 @@ use std::collections::BTreeSet;
 use std::io;
 use std::result::Result as StdResult;

+use either::Either;
+use grenad::MergeFunction;
 use roaring::RoaringBitmap;

 use crate::heed_codec::CboRoaringBitmapCodec;
@@ -10,7 +12,8 @@ use crate::update::del_add::{DelAdd, KvReaderDelAdd, KvWriterDelAdd};
 use crate::update::index_documents::transform::Operation;
 use crate::Result;

-pub type MergeFn = for<'a> fn(&[u8], &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>>;
+pub type EitherObkvMerge =
+    Either<ObkvsKeepLastAdditionMergeDeletions, ObkvsMergeAdditionsAndDeletions>;

 pub fn serialize_roaring_bitmap(bitmap: &RoaringBitmap, buffer: &mut Vec<u8>) -> io::Result<()> {
    buffer.clear();
@@ -18,35 +21,53 @@ pub fn serialize_roaring_bitmap(bitmap: &RoaringBitmap, buffer: &mut Vec<u8>) ->
    bitmap.serialize_into(buffer)
 }

-pub fn merge_roaring_bitmaps<'a>(_key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
-    if values.len() == 1 {
-        Ok(values[0].clone())
-    } else {
-        let merged = values
-            .iter()
-            .map(AsRef::as_ref)
-            .map(RoaringBitmap::deserialize_from)
-            .map(StdResult::unwrap)
-            .reduce(|a, b| a | b)
-            .unwrap();
-        let mut buffer = Vec::new();
-        serialize_roaring_bitmap(&merged, &mut buffer)?;
-        Ok(Cow::Owned(buffer))
+pub struct MergeRoaringBitmaps;
+
+impl MergeFunction for MergeRoaringBitmaps {
+    type Error = crate::Error;
+
+    fn merge<'a>(&self, _key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
+        if values.len() == 1 {
+            Ok(values[0].clone())
+        } else {
+            let merged = values
+                .iter()
+                .map(AsRef::as_ref)
+                .map(RoaringBitmap::deserialize_from)
+                .map(StdResult::unwrap)
+                .reduce(|a, b| a | b)
+                .unwrap();
+            let mut buffer = Vec::new();
+            serialize_roaring_bitmap(&merged, &mut buffer)?;
+            Ok(Cow::Owned(buffer))
+        }
    }
 }

-pub fn keep_first<'a>(_key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
-    Ok(values[0].clone())
+pub struct KeepFirst;
+
+impl MergeFunction for KeepFirst {
+    type Error = crate::Error;
+
+    fn merge<'a>(&self, _key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
+        Ok(values[0].clone())
+    }
 }

 /// Only the last value associated with an id is kept.
-pub fn keep_latest_obkv<'a>(_key: &[u8], obkvs: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
-    Ok(obkvs.last().unwrap().clone())
+pub struct KeepLatestObkv;
+
+impl MergeFunction for KeepLatestObkv {
+    type Error = crate::Error;
+
+    fn merge<'a>(&self, _key: &[u8], obkvs: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
+        Ok(obkvs.last().unwrap().clone())
+    }
 }

 pub fn merge_two_del_add_obkvs(
-    base: obkv::KvReaderU16<'_>,
-    update: obkv::KvReaderU16<'_>,
+    base: &obkv::KvReaderU16,
+    update: &obkv::KvReaderU16,
    merge_additions: bool,
    buffer: &mut Vec<u8>,
 ) {
@@ -66,7 +87,7 @@ pub fn merge_two_del_add_obkvs(
                    // If merge_additions is false, recreate an obkv keeping the deletions only.
                    value_buffer.clear();
                    let mut value_writer = KvWriterDelAdd::new(&mut value_buffer);
-                    let base_reader = KvReaderDelAdd::new(v);
+                    let base_reader = KvReaderDelAdd::from_slice(v);

                    if let Some(deletion) = base_reader.get(DelAdd::Deletion) {
                        value_writer.insert(DelAdd::Deletion, deletion).unwrap();
@@ -80,8 +101,8 @@ pub fn merge_two_del_add_obkvs(
                // merge deletions and additions.
                value_buffer.clear();
                let mut value_writer = KvWriterDelAdd::new(&mut value_buffer);
-                let base_reader = KvReaderDelAdd::new(base);
-                let update_reader = KvReaderDelAdd::new(update);
+                let base_reader = KvReaderDelAdd::from_slice(base);
+                let update_reader = KvReaderDelAdd::from_slice(update);

                // keep newest deletion.
                if let Some(deletion) = update_reader
@@ -131,8 +152,8 @@ fn inner_merge_del_add_obkvs<'a>(
            break;
        }

-        let newest = obkv::KvReader::new(&acc);
-        let oldest = obkv::KvReader::new(&current[1..]);
+        let newest = obkv::KvReader::from_slice(&acc);
+        let oldest = obkv::KvReader::from_slice(&current[1..]);
        merge_two_del_add_obkvs(oldest, newest, merge_additions, &mut buffer);

        // we want the result of the merge into our accumulator.
@@ -145,65 +166,79 @@ fn inner_merge_del_add_obkvs<'a>(
 }

 /// Merge all the obkvs from the newest to the oldest.
-pub fn obkvs_merge_additions_and_deletions<'a>(
-    _key: &[u8],
-    obkvs: &[Cow<'a, [u8]>],
-) -> Result<Cow<'a, [u8]>> {
-    inner_merge_del_add_obkvs(obkvs, true)
+#[derive(Copy, Clone)]
+pub struct ObkvsMergeAdditionsAndDeletions;
+
+impl MergeFunction for ObkvsMergeAdditionsAndDeletions {
+    type Error = crate::Error;
+
+    fn merge<'a>(&self, _key: &[u8], obkvs: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
+        inner_merge_del_add_obkvs(obkvs, true)
+    }
 }

 /// Merge all the obkvs deletions from the newest to the oldest and keep only the newest additions.
-pub fn obkvs_keep_last_addition_merge_deletions<'a>(
-    _key: &[u8],
-    obkvs: &[Cow<'a, [u8]>],
-) -> Result<Cow<'a, [u8]>> {
-    inner_merge_del_add_obkvs(obkvs, false)
+#[derive(Copy, Clone)]
+pub struct ObkvsKeepLastAdditionMergeDeletions;
+
+impl MergeFunction for ObkvsKeepLastAdditionMergeDeletions {
+    type Error = crate::Error;
+
+    fn merge<'a>(&self, _key: &[u8], obkvs: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
+        inner_merge_del_add_obkvs(obkvs, false)
+    }
 }

 /// Do a union of all the CboRoaringBitmaps in the values.
-pub fn merge_cbo_roaring_bitmaps<'a>(
-    _key: &[u8],
-    values: &[Cow<'a, [u8]>],
-) -> Result<Cow<'a, [u8]>> {
-    if values.len() == 1 {
-        Ok(values[0].clone())
-    } else {
-        let mut vec = Vec::new();
-        CboRoaringBitmapCodec::merge_into(values, &mut vec)?;
-        Ok(Cow::from(vec))
+pub struct MergeCboRoaringBitmaps;
+
+impl MergeFunction for MergeCboRoaringBitmaps {
+    type Error = crate::Error;
+
+    fn merge<'a>(&self, _key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
+        if values.len() == 1 {
+            Ok(values[0].clone())
+        } else {
+            let mut vec = Vec::new();
+            CboRoaringBitmapCodec::merge_into(values, &mut vec)?;
+            Ok(Cow::from(vec))
+        }
    }
 }

 /// Do a union of CboRoaringBitmaps on both sides of a DelAdd obkv
 /// separately and outputs a new DelAdd with both unions.
-pub fn merge_deladd_cbo_roaring_bitmaps<'a>(
-    _key: &[u8],
-    values: &[Cow<'a, [u8]>],
-) -> Result<Cow<'a, [u8]>> {
-    if values.len() == 1 {
-        Ok(values[0].clone())
-    } else {
-        // Retrieve the bitmaps from both sides
-        let mut del_bitmaps_bytes = Vec::new();
-        let mut add_bitmaps_bytes = Vec::new();
-        for value in values {
-            let obkv = KvReaderDelAdd::new(value);
-            if let Some(bitmap_bytes) = obkv.get(DelAdd::Deletion) {
-                del_bitmaps_bytes.push(bitmap_bytes);
-            }
-            if let Some(bitmap_bytes) = obkv.get(DelAdd::Addition) {
-                add_bitmaps_bytes.push(bitmap_bytes);
-            }
-        }
+pub struct MergeDeladdCboRoaringBitmaps;

-        let mut output_deladd_obkv = KvWriterDelAdd::memory();
-        let mut buffer = Vec::new();
-        CboRoaringBitmapCodec::merge_into(del_bitmaps_bytes, &mut buffer)?;
-        output_deladd_obkv.insert(DelAdd::Deletion, &buffer)?;
-        buffer.clear();
-        CboRoaringBitmapCodec::merge_into(add_bitmaps_bytes, &mut buffer)?;
-        output_deladd_obkv.insert(DelAdd::Addition, &buffer)?;
-        output_deladd_obkv.into_inner().map(Cow::from).map_err(Into::into)
+impl MergeFunction for MergeDeladdCboRoaringBitmaps {
+    type Error = crate::Error;
+
+    fn merge<'a>(&self, _key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
+        if values.len() == 1 {
+            Ok(values[0].clone())
+        } else {
+            // Retrieve the bitmaps from both sides
+            let mut del_bitmaps_bytes = Vec::new();
+            let mut add_bitmaps_bytes = Vec::new();
+            for value in values {
+                let obkv = KvReaderDelAdd::from_slice(value);
+                if let Some(bitmap_bytes) = obkv.get(DelAdd::Deletion) {
+                    del_bitmaps_bytes.push(bitmap_bytes);
+                }
+                if let Some(bitmap_bytes) = obkv.get(DelAdd::Addition) {
+                    add_bitmaps_bytes.push(bitmap_bytes);
+                }
+            }
+
+            let mut output_deladd_obkv = KvWriterDelAdd::memory();
+            let mut buffer = Vec::new();
+            CboRoaringBitmapCodec::merge_into(del_bitmaps_bytes, &mut buffer)?;
+            output_deladd_obkv.insert(DelAdd::Deletion, &buffer)?;
+            buffer.clear();
+            CboRoaringBitmapCodec::merge_into(add_bitmaps_bytes, &mut buffer)?;
+            output_deladd_obkv.insert(DelAdd::Addition, &buffer)?;
+            output_deladd_obkv.into_inner().map(Cow::from).map_err(Into::into)
+        }
    }
 }

@@ -217,7 +252,7 @@ pub fn merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap<'a>(
    buffer: &'a mut Vec<u8>,
 ) -> Result<Option<&'a [u8]>> {
    Ok(CboRoaringBitmapCodec::merge_deladd_into(
-        KvReaderDelAdd::new(deladd_obkv),
+        KvReaderDelAdd::from_slice(deladd_obkv),
        previous,
        buffer,
    )?)
@@ -225,37 +260,55 @@ pub fn merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap<'a>(

 /// Do a union of BtreeSet on both sides of a DelAdd obkv
 /// separately and outputs a new DelAdd with both unions.
-pub fn merge_deladd_btreeset_string<'a>(
-    _key: &[u8],
-    values: &[Cow<'a, [u8]>],
-) -> Result<Cow<'a, [u8]>> {
-    if values.len() == 1 {
-        Ok(values[0].clone())
-    } else {
-        // Retrieve the bitmaps from both sides
-        let mut del_set = BTreeSet::new();
-        let mut add_set = BTreeSet::new();
-        for value in values {
-            let obkv = KvReaderDelAdd::new(value);
-            if let Some(bytes) = obkv.get(DelAdd::Deletion) {
-                let set = serde_json::from_slice::<BTreeSet<String>>(bytes).unwrap();
-                for value in set {
-                    del_set.insert(value);
-                }
-            }
-            if let Some(bytes) = obkv.get(DelAdd::Addition) {
-                let set = serde_json::from_slice::<BTreeSet<String>>(bytes).unwrap();
-                for value in set {
-                    add_set.insert(value);
-                }
-            }
-        }
+pub struct MergeDeladdBtreesetString;

-        let mut output_deladd_obkv = KvWriterDelAdd::memory();
-        let del = serde_json::to_vec(&del_set).unwrap();
-        output_deladd_obkv.insert(DelAdd::Deletion, &del)?;
-        let add = serde_json::to_vec(&add_set).unwrap();
-        output_deladd_obkv.insert(DelAdd::Addition, &add)?;
-        output_deladd_obkv.into_inner().map(Cow::from).map_err(Into::into)
+impl MergeFunction for MergeDeladdBtreesetString {
+    type Error = crate::Error;
+
+    fn merge<'a>(&self, _key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Cow<'a, [u8]>> {
+        if values.len() == 1 {
+            Ok(values[0].clone())
+        } else {
+            // Retrieve the bitmaps from both sides
+            let mut del_set = BTreeSet::new();
+            let mut add_set = BTreeSet::new();
+            for value in values {
+                let obkv = KvReaderDelAdd::from_slice(value);
+                if let Some(bytes) = obkv.get(DelAdd::Deletion) {
+                    let set = serde_json::from_slice::<BTreeSet<String>>(bytes).unwrap();
+                    for value in set {
+                        del_set.insert(value);
+                    }
+                }
+                if let Some(bytes) = obkv.get(DelAdd::Addition) {
+                    let set = serde_json::from_slice::<BTreeSet<String>>(bytes).unwrap();
+                    for value in set {
+                        add_set.insert(value);
+                    }
+                }
+            }
+
+            let mut output_deladd_obkv = KvWriterDelAdd::memory();
+            let del = serde_json::to_vec(&del_set).unwrap();
+            output_deladd_obkv.insert(DelAdd::Deletion, &del)?;
+            let add = serde_json::to_vec(&add_set).unwrap();
+            output_deladd_obkv.insert(DelAdd::Addition, &add)?;
+            output_deladd_obkv.into_inner().map(Cow::from).map_err(Into::into)
+        }
+    }
+}
+
+/// Used when trying to merge readers, but you don't actually care about the values.
+pub struct MergeIgnoreValues;
+
+impl MergeFunction for MergeIgnoreValues {
+    type Error = crate::Error;
+
+    fn merge<'a>(
+        &self,
+        _key: &[u8],
+        _values: &[Cow<'a, [u8]>],
+    ) -> std::result::Result<Cow<'a, [u8]>, Self::Error> {
+        Ok(Cow::Owned(Vec::new()))
    }
 }
--- a/milli/src/update/index_documents/helpers/mod.rs
+++ b/milli/src/update/index_documents/helpers/mod.rs
@@ -7,17 +7,8 @@ use std::convert::{TryFrom, TryInto};

 pub use clonable_mmap::{ClonableMmap, CursorClonableMmap};
 use fst::{IntoStreamer, Streamer};
-pub use grenad_helpers::{
-    as_cloneable_grenad, create_sorter, create_writer, grenad_obkv_into_chunks,
-    merge_ignore_values, sorter_into_reader, write_sorter_into_database, writer_into_reader,
-    GrenadParameters,
-};
-pub use merge_functions::{
-    keep_first, keep_latest_obkv, merge_cbo_roaring_bitmaps, merge_deladd_btreeset_string,
-    merge_deladd_cbo_roaring_bitmaps, merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap,
-    merge_roaring_bitmaps, obkvs_keep_last_addition_merge_deletions,
-    obkvs_merge_additions_and_deletions, MergeFn,
-};
+pub use grenad_helpers::*;
+pub use merge_functions::*;

 use crate::MAX_WORD_LENGTH;

--- a/milli/src/update/index_documents/mod.rs
+++ b/milli/src/update/index_documents/mod.rs
@@ -27,13 +27,7 @@ use typed_chunk::{write_typed_chunk_into_index, ChunkAccumulator, TypedChunk};

 use self::enrich::enrich_documents_batch;
 pub use self::enrich::{extract_finite_float_from_value, DocumentId};
-pub use self::helpers::{
-    as_cloneable_grenad, create_sorter, create_writer, fst_stream_into_hashset,
-    fst_stream_into_vec, merge_cbo_roaring_bitmaps, merge_deladd_cbo_roaring_bitmaps,
-    merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap, merge_roaring_bitmaps,
-    valid_lmdb_key, write_sorter_into_database, writer_into_reader, MergeFn,
-};
-use self::helpers::{grenad_obkv_into_chunks, GrenadParameters};
+pub use self::helpers::*;
 pub use self::transform::{Transform, TransformOutput};
 use crate::documents::{obkv_to_object, DocumentsBatchBuilder, DocumentsBatchReader};
 use crate::error::{Error, InternalError, UserError};
@@ -605,7 +599,7 @@ where
                                let cloneable_chunk =
                                    unsafe { as_cloneable_grenad(&word_docids_reader)? };
                                let word_docids = word_docids.get_or_insert_with(|| {
-                                    MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn)
+                                    MergerBuilder::new(MergeDeladdCboRoaringBitmaps)
                                });
                                word_docids.push(cloneable_chunk.into_cursor()?);
                                let cloneable_chunk =
@@ -613,14 +607,14 @@ where
                                let exact_word_docids =
                                    exact_word_docids.get_or_insert_with(|| {
                                        MergerBuilder::new(
-                                            merge_deladd_cbo_roaring_bitmaps as MergeFn,
+                                            MergeDeladdCboRoaringBitmaps,
                                        )
                                    });
                                exact_word_docids.push(cloneable_chunk.into_cursor()?);
                                let cloneable_chunk =
                                    unsafe { as_cloneable_grenad(&word_fid_docids_reader)? };
                                let word_fid_docids = word_fid_docids.get_or_insert_with(|| {
-                                    MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn)
+                                    MergerBuilder::new(MergeDeladdCboRoaringBitmaps)
                                });
                                word_fid_docids.push(cloneable_chunk.into_cursor()?);
                                TypedChunk::WordDocids {
@@ -634,7 +628,7 @@ where
                                let word_position_docids =
                                    word_position_docids.get_or_insert_with(|| {
                                        MergerBuilder::new(
-                                            merge_deladd_cbo_roaring_bitmaps as MergeFn,
+                                            MergeDeladdCboRoaringBitmaps,
                                        )
                                    });
                                word_position_docids.push(cloneable_chunk.into_cursor()?);
@@ -689,8 +683,9 @@ where
                        key: None,
                    },
                )?;
+                let first_id = crate::vector::arroy_db_range_for_embedder(index).next().unwrap();
                let reader =
-                    ArroyWrapper::new(self.index.vector_arroy, index, action.was_quantized);
+                    ArroyWrapper::new(self.index.vector_arroy, first_id, action.was_quantized);
                let dim = reader.dimensions(self.wtxn)?;
                dimension.insert(name.to_string(), dim);
            }
@@ -699,7 +694,6 @@ where
        for (embedder_name, dimension) in dimension {
            let wtxn = &mut *self.wtxn;
            let vector_arroy = self.index.vector_arroy;
-            let cancel = &self.should_abort;

            let embedder_index = self.index.embedder_category_id.get(wtxn, &embedder_name)?.ok_or(
                InternalError::DatabaseMissingEntry { db_name: "embedder_category_id", key: None },
@@ -713,8 +707,17 @@ where
            let is_quantizing = embedder_config.map_or(false, |action| action.is_being_quantized);

            pool.install(|| {
-                let mut writer = ArroyWrapper::new(vector_arroy, embedder_index, was_quantized);
-                writer.build_and_quantize(wtxn, &mut rng, dimension, is_quantizing, cancel)?;
+                for k in crate::vector::arroy_db_range_for_embedder(embedder_index) {
+                    let mut writer = ArroyWrapper::new(vector_arroy, k, was_quantized);
+                    if is_quantizing {
+                        writer.quantize(wtxn, k, dimension)?;
+                    }
+                    if writer.need_build(wtxn, dimension)? {
+                        writer.build(wtxn, &mut rng, dimension)?;
+                    } else if writer.is_empty(wtxn, dimension)? {
+                        break;
+                    }
+                }
                Result::Ok(())
            })
            .map_err(InternalError::from)??;
@@ -738,10 +741,10 @@ where
    )]
    pub fn execute_prefix_databases(
        self,
-        word_docids: Option<Merger<CursorClonableMmap, MergeFn>>,
-        exact_word_docids: Option<Merger<CursorClonableMmap, MergeFn>>,
-        word_position_docids: Option<Merger<CursorClonableMmap, MergeFn>>,
-        word_fid_docids: Option<Merger<CursorClonableMmap, MergeFn>>,
+        word_docids: Option<Merger<CursorClonableMmap, MergeDeladdCboRoaringBitmaps>>,
+        exact_word_docids: Option<Merger<CursorClonableMmap, MergeDeladdCboRoaringBitmaps>>,
+        word_position_docids: Option<Merger<CursorClonableMmap, MergeDeladdCboRoaringBitmaps>>,
+        word_fid_docids: Option<Merger<CursorClonableMmap, MergeDeladdCboRoaringBitmaps>>,
    ) -> Result<()>
    where
        FP: Fn(UpdateIndexingStep) + Sync,
@@ -921,7 +924,7 @@ where
 )]
 fn execute_word_prefix_docids(
    txn: &mut heed::RwTxn<'_>,
-    merger: Merger<CursorClonableMmap, MergeFn>,
+    merger: Merger<CursorClonableMmap, MergeDeladdCboRoaringBitmaps>,
    word_docids_db: Database<Str, CboRoaringBitmapCodec>,
    word_prefix_docids_db: Database<Str, CboRoaringBitmapCodec>,
    indexer_config: &IndexerConfig,
--- a/milli/src/update/index_documents/parallel.rs
+++ b/milli/src/update/index_documents/parallel.rs
@@ -31,14 +31,14 @@ impl<'t> ImmutableObkvs<'t> {
    }

    /// Returns the OBKVs identified by the given ID.
-    pub fn obkv(&self, docid: DocumentId) -> heed::Result<Option<KvReaderU16<'t>>> {
+    pub fn obkv(&self, docid: DocumentId) -> heed::Result<Option<&'t KvReaderU16>> {
        match self
            .ids
            .rank(docid)
            .checked_sub(1)
            .and_then(|offset| self.slices.get(offset as usize))
        {
-            Some(bytes) => Ok(Some(KvReaderU16::new(bytes))),
+            Some(&bytes) => Ok(Some(bytes.into())),
            None => Ok(None),
        }
    }
--- a/milli/src/update/index_documents/transform.rs
+++ b/milli/src/update/index_documents/transform.rs
@@ -5,6 +5,7 @@ use std::collections::{BTreeMap, HashMap, HashSet};
 use std::fs::File;
 use std::io::{Read, Seek};

+use either::Either;
 use fxhash::FxHashMap;
 use itertools::Itertools;
 use obkv::{KvReader, KvReaderU16, KvWriter};
@@ -13,10 +14,10 @@ use serde_json::Value;
 use smartstring::SmartString;

 use super::helpers::{
-    create_sorter, create_writer, keep_first, obkvs_keep_last_addition_merge_deletions,
-    obkvs_merge_additions_and_deletions, sorter_into_reader, MergeFn,
+    create_sorter, create_writer, sorter_into_reader, EitherObkvMerge,
+    ObkvsKeepLastAdditionMergeDeletions, ObkvsMergeAdditionsAndDeletions,
 };
-use super::{IndexDocumentsMethod, IndexerConfig};
+use super::{IndexDocumentsMethod, IndexerConfig, KeepFirst};
 use crate::documents::{DocumentsBatchIndex, EnrichedDocument, EnrichedDocumentsBatchReader};
 use crate::error::{Error, InternalError, UserError};
 use crate::index::{db_name, main_key};
@@ -26,7 +27,7 @@ use crate::update::del_add::{
 };
 use crate::update::index_documents::GrenadParameters;
 use crate::update::settings::{InnerIndexSettings, InnerIndexSettingsDiff};
-use crate::update::{AvailableDocumentsIds, UpdateIndexingStep};
+use crate::update::{AvailableIds, UpdateIndexingStep};
 use crate::vector::parsed_vectors::{ExplicitVectors, VectorOrArrayOfVectors};
 use crate::vector::settings::WriteBackToDocuments;
 use crate::vector::ArroyWrapper;
@@ -55,13 +56,13 @@ pub struct Transform<'a, 'i> {

    indexer_settings: &'a IndexerConfig,
    pub index_documents_method: IndexDocumentsMethod,
-    available_documents_ids: AvailableDocumentsIds,
+    available_documents_ids: AvailableIds,

    // Both grenad follows the same format:
    // key | value
    // u32 | 1 byte for the Operation byte, the rest is the obkv of the document stored
-    original_sorter: grenad::Sorter<MergeFn>,
-    flattened_sorter: grenad::Sorter<MergeFn>,
+    original_sorter: grenad::Sorter<EitherObkvMerge>,
+    flattened_sorter: grenad::Sorter<EitherObkvMerge>,

    replaced_documents_ids: RoaringBitmap,
    new_documents_ids: RoaringBitmap,
@@ -109,17 +110,19 @@ impl<'a, 'i> Transform<'a, 'i> {
        index_documents_method: IndexDocumentsMethod,
        _autogenerate_docids: bool,
    ) -> Result<Self> {
+        use IndexDocumentsMethod::{ReplaceDocuments, UpdateDocuments};
+
        // We must choose the appropriate merge function for when two or more documents
        // with the same user id must be merged or fully replaced in the same batch.
        let merge_function = match index_documents_method {
-            IndexDocumentsMethod::ReplaceDocuments => obkvs_keep_last_addition_merge_deletions,
-            IndexDocumentsMethod::UpdateDocuments => obkvs_merge_additions_and_deletions,
+            ReplaceDocuments => Either::Left(ObkvsKeepLastAdditionMergeDeletions),
+            UpdateDocuments => Either::Right(ObkvsMergeAdditionsAndDeletions),
        };

        // We initialize the sorter with the user indexing settings.
        let original_sorter = create_sorter(
            grenad::SortAlgorithm::Stable,
-            merge_function,
+            merge_function.clone(),
            indexer_settings.chunk_compression_type,
            indexer_settings.chunk_compression_level,
            indexer_settings.max_nb_chunks,
@@ -141,7 +144,7 @@ impl<'a, 'i> Transform<'a, 'i> {
            index,
            fields_ids_map: index.fields_ids_map(wtxn)?,
            indexer_settings,
-            available_documents_ids: AvailableDocumentsIds::from_documents_ids(&documents_ids),
+            available_documents_ids: AvailableIds::new(&documents_ids),
            original_sorter,
            flattened_sorter,
            index_documents_method,
@@ -279,21 +282,21 @@ impl<'a, 'i> Transform<'a, 'i> {
                    document_sorter_value_buffer.clear();
                    document_sorter_value_buffer.push(Operation::Addition as u8);
                    into_del_add_obkv(
-                        KvReaderU16::new(base_obkv),
+                        KvReaderU16::from_slice(base_obkv),
                        deladd_operation,
                        &mut document_sorter_value_buffer,
                    )?;
                    self.original_sorter
                        .insert(&document_sorter_key_buffer, &document_sorter_value_buffer)?;
-                    let base_obkv = KvReader::new(base_obkv);
+                    let base_obkv = KvReader::from_slice(base_obkv);
                    if let Some(flattened_obkv) =
-                        Self::flatten_from_fields_ids_map(&base_obkv, &mut self.fields_ids_map)?
+                        Self::flatten_from_fields_ids_map(base_obkv, &mut self.fields_ids_map)?
                    {
                        // we recreate our buffer with the flattened documents
                        document_sorter_value_buffer.clear();
                        document_sorter_value_buffer.push(Operation::Addition as u8);
                        into_del_add_obkv(
-                            KvReaderU16::new(&flattened_obkv),
+                            KvReaderU16::from_slice(&flattened_obkv),
                            deladd_operation,
                            &mut document_sorter_value_buffer,
                        )?;
@@ -312,7 +315,7 @@ impl<'a, 'i> Transform<'a, 'i> {
                document_sorter_value_buffer.clear();
                document_sorter_value_buffer.push(Operation::Addition as u8);
                into_del_add_obkv(
-                    KvReaderU16::new(&obkv_buffer),
+                    KvReaderU16::from_slice(&obkv_buffer),
                    DelAddOperation::Addition,
                    &mut document_sorter_value_buffer,
                )?;
@@ -320,14 +323,14 @@ impl<'a, 'i> Transform<'a, 'i> {
                self.original_sorter
                    .insert(&document_sorter_key_buffer, &document_sorter_value_buffer)?;

-                let flattened_obkv = KvReader::new(&obkv_buffer);
+                let flattened_obkv = KvReader::from_slice(&obkv_buffer);
                if let Some(obkv) =
-                    Self::flatten_from_fields_ids_map(&flattened_obkv, &mut self.fields_ids_map)?
+                    Self::flatten_from_fields_ids_map(flattened_obkv, &mut self.fields_ids_map)?
                {
                    document_sorter_value_buffer.clear();
                    document_sorter_value_buffer.push(Operation::Addition as u8);
                    into_del_add_obkv(
-                        KvReaderU16::new(&obkv),
+                        KvReaderU16::from_slice(&obkv),
                        DelAddOperation::Addition,
                        &mut document_sorter_value_buffer,
                    )?
@@ -520,22 +523,22 @@ impl<'a, 'i> Transform<'a, 'i> {
        document_sorter_value_buffer.clear();
        document_sorter_value_buffer.push(Operation::Deletion as u8);
        into_del_add_obkv(
-            KvReaderU16::new(base_obkv),
+            KvReaderU16::from_slice(base_obkv),
            DelAddOperation::Deletion,
            document_sorter_value_buffer,
        )?;
        self.original_sorter.insert(&document_sorter_key_buffer, &document_sorter_value_buffer)?;

        // flatten it and push it as to delete in the flattened_sorter
-        let flattened_obkv = KvReader::new(base_obkv);
+        let flattened_obkv = KvReader::from_slice(base_obkv);
        if let Some(obkv) =
-            Self::flatten_from_fields_ids_map(&flattened_obkv, &mut self.fields_ids_map)?
+            Self::flatten_from_fields_ids_map(flattened_obkv, &mut self.fields_ids_map)?
        {
            // we recreate our buffer with the flattened documents
            document_sorter_value_buffer.clear();
            document_sorter_value_buffer.push(Operation::Deletion as u8);
            into_del_add_obkv(
-                KvReaderU16::new(&obkv),
+                KvReaderU16::from_slice(&obkv),
                DelAddOperation::Deletion,
                document_sorter_value_buffer,
            )?;
@@ -553,7 +556,7 @@ impl<'a, 'i> Transform<'a, 'i> {
        target = "indexing::transform"
    )]
    fn flatten_from_fields_ids_map(
-        obkv: &KvReader<'_, FieldId>,
+        obkv: &KvReader<FieldId>,
        fields_ids_map: &mut FieldsIdsMap,
    ) -> Result<Option<Vec<u8>>> {
        if obkv
@@ -721,10 +724,10 @@ impl<'a, 'i> Transform<'a, 'i> {
                total_documents: self.documents_count,
            });

-            for (key, value) in KvReader::new(val) {
-                let reader = KvReaderDelAdd::new(value);
+            for (key, value) in KvReader::from_slice(val) {
+                let reader = KvReaderDelAdd::from_slice(value);
                match (reader.get(DelAdd::Deletion), reader.get(DelAdd::Addition)) {
-                    (None, None) => {}
+                    (None, None) => (),
                    (None, Some(_)) => {
                        // New field
                        let name = self.fields_ids_map.name(key).ok_or(
@@ -838,7 +841,7 @@ impl<'a, 'i> Transform<'a, 'i> {
    /// then fill the provided buffers with delta documents using KvWritterDelAdd.
    #[allow(clippy::too_many_arguments)] // need the vectors + fid, feel free to create a struct xo xo
    fn rebind_existing_document(
-        old_obkv: KvReader<'_, FieldId>,
+        old_obkv: &KvReader<FieldId>,
        settings_diff: &InnerIndexSettingsDiff,
        modified_faceted_fields: &HashSet<String>,
        mut injected_vectors: serde_json::Map<String, serde_json::Value>,
@@ -926,7 +929,7 @@ impl<'a, 'i> Transform<'a, 'i> {
        }

        let data = obkv_writer.into_inner()?;
-        let obkv = KvReader::<FieldId>::new(&data);
+        let obkv = KvReader::<FieldId>::from_slice(&data);

        if let Some(original_obkv_buffer) = original_obkv_buffer {
            original_obkv_buffer.clear();
@@ -936,8 +939,8 @@ impl<'a, 'i> Transform<'a, 'i> {
        if let Some(flattened_obkv_buffer) = flattened_obkv_buffer {
            // take the non-flattened version if flatten_from_fields_ids_map returns None.
            let mut fields_ids_map = settings_diff.new.fields_ids_map.clone();
-            let flattened = Self::flatten_from_fields_ids_map(&obkv, &mut fields_ids_map)?;
-            let flattened = flattened.as_deref().map_or(obkv, KvReader::new);
+            let flattened = Self::flatten_from_fields_ids_map(obkv, &mut fields_ids_map)?;
+            let flattened = flattened.as_deref().map_or(obkv, KvReader::from_slice);

            flattened_obkv_buffer.clear();
            into_del_add_obkv_conditional_operation(flattened, flattened_obkv_buffer, |id| {
@@ -980,7 +983,7 @@ impl<'a, 'i> Transform<'a, 'i> {
        let mut original_sorter = if settings_diff.reindex_vectors() {
            Some(create_sorter(
                grenad::SortAlgorithm::Stable,
-                keep_first,
+                KeepFirst,
                self.indexer_settings.chunk_compression_type,
                self.indexer_settings.chunk_compression_level,
                self.indexer_settings.max_nb_chunks,
@@ -990,24 +993,27 @@ impl<'a, 'i> Transform<'a, 'i> {
            None
        };

-        let readers: BTreeMap<&str, (ArroyWrapper, &RoaringBitmap)> = settings_diff
+        let readers: Result<BTreeMap<&str, (Vec<ArroyWrapper>, &RoaringBitmap)>> = settings_diff
            .embedding_config_updates
            .iter()
            .filter_map(|(name, action)| {
                if let Some(WriteBackToDocuments { embedder_id, user_provided }) =
                    action.write_back()
                {
-                    let reader = ArroyWrapper::new(
-                        self.index.vector_arroy,
-                        *embedder_id,
-                        action.was_quantized,
-                    );
-                    Some((name.as_str(), (reader, user_provided)))
+                    let readers: Result<Vec<_>> = self
+                        .index
+                        .arroy_readers(wtxn, *embedder_id, action.was_quantized)
+                        .collect();
+                    match readers {
+                        Ok(readers) => Some(Ok((name.as_str(), (readers, user_provided)))),
+                        Err(error) => Some(Err(error)),
+                    }
                } else {
                    None
                }
            })
            .collect();
+        let readers = readers?;

        let old_vectors_fid = settings_diff
            .old
@@ -1019,7 +1025,7 @@ impl<'a, 'i> Transform<'a, 'i> {
            if settings_diff.reindex_searchable() || settings_diff.reindex_facets() {
                Some(create_sorter(
                    grenad::SortAlgorithm::Stable,
-                    keep_first,
+                    KeepFirst,
                    self.indexer_settings.chunk_compression_type,
                    self.indexer_settings.chunk_compression_level,
                    self.indexer_settings.max_nb_chunks,
@@ -1045,24 +1051,34 @@ impl<'a, 'i> Transform<'a, 'i> {
                    arroy::Error,
                > = readers
                    .iter()
-                    .filter_map(|(name, (reader, user_provided))| {
+                    .filter_map(|(name, (readers, user_provided))| {
                        if !user_provided.contains(docid) {
                            return None;
                        }
-                        match reader.item_vectors(wtxn, docid) {
-                            Ok(vectors) if vectors.is_empty() => None,
-                            Ok(vectors) => Some(Ok((
-                                name.to_string(),
-                                serde_json::to_value(ExplicitVectors {
-                                    embeddings: Some(
-                                        VectorOrArrayOfVectors::from_array_of_vectors(vectors),
-                                    ),
-                                    regenerate: false,
-                                })
-                                .unwrap(),
-                            ))),
-                            Err(e) => Some(Err(e)),
+                        let mut vectors = Vec::new();
+                        for reader in readers {
+                            let Some(vector) = reader.item_vector(wtxn, docid).transpose() else {
+                                break;
+                            };
+
+                            match vector {
+                                Ok(vector) => vectors.push(vector),
+                                Err(error) => return Some(Err(error)),
+                            }
                        }
+                        if vectors.is_empty() {
+                            return None;
+                        }
+                        Some(Ok((
+                            name.to_string(),
+                            serde_json::to_value(ExplicitVectors {
+                                embeddings: Some(VectorOrArrayOfVectors::from_array_of_vectors(
+                                    vectors,
+                                )),
+                                regenerate: false,
+                            })
+                            .unwrap(),
+                        )))
                    })
                    .collect();

@@ -1091,9 +1107,11 @@ impl<'a, 'i> Transform<'a, 'i> {
        }

        // delete all vectors from the embedders that need removal
-        for (_, (reader, _)) in readers {
-            let dimensions = reader.dimensions(wtxn)?;
-            reader.clear(wtxn, dimensions)?;
+        for (_, (readers, _)) in readers {
+            for reader in readers {
+                let dimensions = reader.dimensions(wtxn)?;
+                reader.clear(wtxn, dimensions)?;
+            }
        }

        let grenad_params = GrenadParameters {
@@ -1137,6 +1155,8 @@ fn drop_and_reuse<U, T>(mut vec: Vec<U>) -> Vec<T> {

 #[cfg(test)]
 mod test {
+    use grenad::MergeFunction;
+
    use super::*;

    #[test]
@@ -1148,21 +1168,21 @@ mod test {
        kv_writer.insert(0_u8, [0]).unwrap();
        let buffer = kv_writer.into_inner().unwrap();
        into_del_add_obkv(
-            KvReaderU16::new(&buffer),
+            KvReaderU16::from_slice(&buffer),
            DelAddOperation::Addition,
            &mut additive_doc_0,
        )
        .unwrap();
        additive_doc_0.insert(0, Operation::Addition as u8);
        into_del_add_obkv(
-            KvReaderU16::new(&buffer),
+            KvReaderU16::from_slice(&buffer),
            DelAddOperation::Deletion,
            &mut deletive_doc_0,
        )
        .unwrap();
        deletive_doc_0.insert(0, Operation::Deletion as u8);
        into_del_add_obkv(
-            KvReaderU16::new(&buffer),
+            KvReaderU16::from_slice(&buffer),
            DelAddOperation::DeletionAndAddition,
            &mut del_add_doc_0,
        )
@@ -1174,7 +1194,7 @@ mod test {
        kv_writer.insert(1_u8, [1]).unwrap();
        let buffer = kv_writer.into_inner().unwrap();
        into_del_add_obkv(
-            KvReaderU16::new(&buffer),
+            KvReaderU16::from_slice(&buffer),
            DelAddOperation::Addition,
            &mut additive_doc_1,
        )
@@ -1187,32 +1207,39 @@ mod test {
        kv_writer.insert(1_u8, [1]).unwrap();
        let buffer = kv_writer.into_inner().unwrap();
        into_del_add_obkv(
-            KvReaderU16::new(&buffer),
+            KvReaderU16::from_slice(&buffer),
            DelAddOperation::Addition,
            &mut additive_doc_0_1,
        )
        .unwrap();
        additive_doc_0_1.insert(0, Operation::Addition as u8);

-        let ret = obkvs_merge_additions_and_deletions(&[], &[Cow::from(additive_doc_0.as_slice())])
-            .unwrap();
+        let ret = MergeFunction::merge(
+            &ObkvsMergeAdditionsAndDeletions,
+            &[],
+            &[Cow::from(additive_doc_0.as_slice())],
+        )
+        .unwrap();
        assert_eq!(*ret, additive_doc_0);

-        let ret = obkvs_merge_additions_and_deletions(
+        let ret = MergeFunction::merge(
+            &ObkvsMergeAdditionsAndDeletions,
            &[],
            &[Cow::from(deletive_doc_0.as_slice()), Cow::from(additive_doc_0.as_slice())],
        )
        .unwrap();
        assert_eq!(*ret, del_add_doc_0);

-        let ret = obkvs_merge_additions_and_deletions(
+        let ret = MergeFunction::merge(
+            &ObkvsMergeAdditionsAndDeletions,
            &[],
            &[Cow::from(additive_doc_0.as_slice()), Cow::from(deletive_doc_0.as_slice())],
        )
        .unwrap();
        assert_eq!(*ret, deletive_doc_0);

-        let ret = obkvs_merge_additions_and_deletions(
+        let ret = MergeFunction::merge(
+            &ObkvsMergeAdditionsAndDeletions,
            &[],
            &[
                Cow::from(additive_doc_1.as_slice()),
@@ -1223,21 +1250,24 @@ mod test {
        .unwrap();
        assert_eq!(*ret, del_add_doc_0);

-        let ret = obkvs_merge_additions_and_deletions(
+        let ret = MergeFunction::merge(
+            &ObkvsMergeAdditionsAndDeletions,
            &[],
            &[Cow::from(additive_doc_1.as_slice()), Cow::from(additive_doc_0.as_slice())],
        )
        .unwrap();
        assert_eq!(*ret, additive_doc_0_1);

-        let ret = obkvs_keep_last_addition_merge_deletions(
+        let ret = MergeFunction::merge(
+            &ObkvsKeepLastAdditionMergeDeletions,
            &[],
            &[Cow::from(additive_doc_1.as_slice()), Cow::from(additive_doc_0.as_slice())],
        )
        .unwrap();
        assert_eq!(*ret, additive_doc_0);

-        let ret = obkvs_keep_last_addition_merge_deletions(
+        let ret = MergeFunction::merge(
+            &ObkvsKeepLastAdditionMergeDeletions,
            &[],
            &[
                Cow::from(deletive_doc_0.as_slice()),
--- a/milli/src/update/index_documents/typed_chunk.rs
+++ b/milli/src/update/index_documents/typed_chunk.rs
@@ -4,18 +4,17 @@ use std::fs::File;
 use std::io::{self, BufReader};

 use bytemuck::allocation::pod_collect_to_vec;
-use grenad::{Merger, MergerBuilder};
+use grenad::{MergeFunction, Merger, MergerBuilder};
 use heed::types::Bytes;
 use heed::{BytesDecode, RwTxn};
 use obkv::{KvReader, KvWriter};
 use roaring::RoaringBitmap;

 use super::helpers::{
-    self, keep_first, merge_deladd_btreeset_string, merge_deladd_cbo_roaring_bitmaps,
-    merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap, merge_ignore_values, valid_lmdb_key,
-    CursorClonableMmap,
+    self, merge_deladd_cbo_roaring_bitmaps_into_cbo_roaring_bitmap, valid_lmdb_key,
+    CursorClonableMmap, KeepFirst, MergeDeladdBtreesetString, MergeDeladdCboRoaringBitmaps,
+    MergeIgnoreValues,
 };
-use super::MergeFn;
 use crate::external_documents_ids::{DocumentOperation, DocumentOperationKind};
 use crate::facet::FacetType;
 use crate::index::db_name::DOCUMENTS;
@@ -24,7 +23,7 @@ use crate::proximity::MAX_DISTANCE;
 use crate::update::del_add::{deladd_serialize_add_side, DelAdd, KvReaderDelAdd};
 use crate::update::facet::FacetsUpdate;
 use crate::update::index_documents::helpers::{
-    as_cloneable_grenad, keep_latest_obkv, try_split_array_at,
+    as_cloneable_grenad, try_split_array_at, KeepLatestObkv,
 };
 use crate::update::settings::InnerIndexSettingsDiff;
 use crate::vector::ArroyWrapper;
@@ -141,7 +140,7 @@ pub(crate) fn write_typed_chunk_into_index(
            let vectors_fid =
                fields_ids_map.id(crate::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME);

-            let mut builder = MergerBuilder::new(keep_latest_obkv as MergeFn);
+            let mut builder = MergerBuilder::new(KeepLatestObkv);
            for typed_chunk in typed_chunks {
                let TypedChunk::Documents(chunk) = typed_chunk else {
                    unreachable!();
@@ -163,7 +162,7 @@ pub(crate) fn write_typed_chunk_into_index(
            let mut vectors_buffer = Vec::new();
            while let Some((key, reader)) = iter.next()? {
                let mut writer: KvWriter<_, FieldId> = KvWriter::memory();
-                let reader: KvReader<'_, FieldId> = KvReader::new(reader);
+                let reader: &KvReader<FieldId> = reader.into();

                let (document_id_bytes, external_id_bytes) = try_split_array_at(key)
                    .ok_or(SerializationError::Decoding { db_name: Some(DOCUMENTS) })?;
@@ -171,7 +170,7 @@ pub(crate) fn write_typed_chunk_into_index(
                let external_id = std::str::from_utf8(external_id_bytes)?;

                for (field_id, value) in reader.iter() {
-                    let del_add_reader = KvReaderDelAdd::new(value);
+                    let del_add_reader = KvReaderDelAdd::from_slice(value);

                    if let Some(addition) = del_add_reader.get(DelAdd::Addition) {
                        let addition = if vectors_fid == Some(field_id) {
@@ -235,7 +234,7 @@ pub(crate) fn write_typed_chunk_into_index(
                tracing::trace_span!(target: "indexing::write_db", "field_id_word_count_docids");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            for typed_chunk in typed_chunks {
                let TypedChunk::FieldIdWordCountDocids(chunk) = typed_chunk else {
                    unreachable!();
@@ -258,13 +257,10 @@ pub(crate) fn write_typed_chunk_into_index(
            let span = tracing::trace_span!(target: "indexing::write_db", "word_docids");
            let _entered = span.enter();

-            let mut word_docids_builder =
-                MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
-            let mut exact_word_docids_builder =
-                MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
-            let mut word_fid_docids_builder =
-                MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
-            let mut fst_merger_builder = MergerBuilder::new(merge_ignore_values as MergeFn);
+            let mut word_docids_builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
+            let mut exact_word_docids_builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
+            let mut word_fid_docids_builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
+            let mut fst_merger_builder = MergerBuilder::new(MergeIgnoreValues);
            for typed_chunk in typed_chunks {
                let TypedChunk::WordDocids {
                    word_docids_reader,
@@ -329,7 +325,7 @@ pub(crate) fn write_typed_chunk_into_index(
            let span = tracing::trace_span!(target: "indexing::write_db", "word_position_docids");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            for typed_chunk in typed_chunks {
                let TypedChunk::WordPositionDocids(chunk) = typed_chunk else {
                    unreachable!();
@@ -353,7 +349,7 @@ pub(crate) fn write_typed_chunk_into_index(
                tracing::trace_span!(target: "indexing::write_db","field_id_facet_number_docids");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            let mut data_size = 0;
            for typed_chunk in typed_chunks {
                let TypedChunk::FieldIdFacetNumberDocids(facet_id_number_docids) = typed_chunk
@@ -375,10 +371,9 @@ pub(crate) fn write_typed_chunk_into_index(
                tracing::trace_span!(target: "indexing::write_db", "field_id_facet_string_docids");
            let _entered = span.enter();

-            let mut facet_id_string_builder =
-                MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut facet_id_string_builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            let mut normalized_facet_id_string_builder =
-                MergerBuilder::new(merge_deladd_btreeset_string as MergeFn);
+                MergerBuilder::new(MergeDeladdBtreesetString);
            let mut data_size = 0;
            for typed_chunk in typed_chunks {
                let TypedChunk::FieldIdFacetStringDocids((
@@ -412,7 +407,7 @@ pub(crate) fn write_typed_chunk_into_index(
                tracing::trace_span!(target: "indexing::write_db", "field_id_facet_exists_docids");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            for typed_chunk in typed_chunks {
                let TypedChunk::FieldIdFacetExistsDocids(chunk) = typed_chunk else {
                    unreachable!();
@@ -436,7 +431,7 @@ pub(crate) fn write_typed_chunk_into_index(
                tracing::trace_span!(target: "indexing::write_db", "field_id_facet_is_null_docids");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            for typed_chunk in typed_chunks {
                let TypedChunk::FieldIdFacetIsNullDocids(chunk) = typed_chunk else {
                    unreachable!();
@@ -459,7 +454,7 @@ pub(crate) fn write_typed_chunk_into_index(
            let span = tracing::trace_span!(target: "indexing::write_db", "field_id_facet_is_empty_docids");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            for typed_chunk in typed_chunks {
                let TypedChunk::FieldIdFacetIsEmptyDocids(chunk) = typed_chunk else {
                    unreachable!();
@@ -483,7 +478,7 @@ pub(crate) fn write_typed_chunk_into_index(
                tracing::trace_span!(target: "indexing::write_db", "word_pair_proximity_docids");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(merge_deladd_cbo_roaring_bitmaps as MergeFn);
+            let mut builder = MergerBuilder::new(MergeDeladdCboRoaringBitmaps);
            for typed_chunk in typed_chunks {
                let TypedChunk::WordPairProximityDocids(chunk) = typed_chunk else {
                    unreachable!();
@@ -516,7 +511,7 @@ pub(crate) fn write_typed_chunk_into_index(
                tracing::trace_span!(target: "indexing::write_db", "field_id_docid_facet_numbers");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(keep_first as MergeFn);
+            let mut builder = MergerBuilder::new(KeepFirst);
            for typed_chunk in typed_chunks {
                let TypedChunk::FieldIdDocidFacetNumbers(chunk) = typed_chunk else {
                    unreachable!();
@@ -530,7 +525,7 @@ pub(crate) fn write_typed_chunk_into_index(
                index.field_id_docid_facet_f64s.remap_types::<Bytes, Bytes>();
            let mut iter = merger.into_stream_merger_iter()?;
            while let Some((key, value)) = iter.next()? {
-                let reader = KvReaderDelAdd::new(value);
+                let reader = KvReaderDelAdd::from_slice(value);
                if valid_lmdb_key(key) {
                    match (reader.get(DelAdd::Deletion), reader.get(DelAdd::Addition)) {
                        (None, None) => {}
@@ -550,7 +545,7 @@ pub(crate) fn write_typed_chunk_into_index(
                tracing::trace_span!(target: "indexing::write_db", "field_id_docid_facet_strings");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(keep_first as MergeFn);
+            let mut builder = MergerBuilder::new(KeepFirst);
            for typed_chunk in typed_chunks {
                let TypedChunk::FieldIdDocidFacetStrings(chunk) = typed_chunk else {
                    unreachable!();
@@ -564,7 +559,7 @@ pub(crate) fn write_typed_chunk_into_index(
                index.field_id_docid_facet_strings.remap_types::<Bytes, Bytes>();
            let mut iter = merger.into_stream_merger_iter()?;
            while let Some((key, value)) = iter.next()? {
-                let reader = KvReaderDelAdd::new(value);
+                let reader = KvReaderDelAdd::from_slice(value);
                if valid_lmdb_key(key) {
                    match (reader.get(DelAdd::Deletion), reader.get(DelAdd::Addition)) {
                        (None, None) => {}
@@ -583,7 +578,7 @@ pub(crate) fn write_typed_chunk_into_index(
            let span = tracing::trace_span!(target: "indexing::write_db", "geo_points");
            let _entered = span.enter();

-            let mut builder = MergerBuilder::new(keep_first as MergeFn);
+            let mut builder = MergerBuilder::new(KeepFirst);
            for typed_chunk in typed_chunks {
                let TypedChunk::GeoPoints(chunk) = typed_chunk else {
                    unreachable!();
@@ -601,7 +596,7 @@ pub(crate) fn write_typed_chunk_into_index(
                // convert the key back to a u32 (4 bytes)
                let docid = key.try_into().map(DocumentId::from_be_bytes).unwrap();

-                let deladd_obkv = KvReaderDelAdd::new(value);
+                let deladd_obkv = KvReaderDelAdd::from_slice(value);
                if let Some(value) = deladd_obkv.get(DelAdd::Deletion) {
                    let geopoint = extract_geo_point(value, docid);
                    rtree.remove(&geopoint);
@@ -620,9 +615,9 @@ pub(crate) fn write_typed_chunk_into_index(
            let span = tracing::trace_span!(target: "indexing::write_db", "vector_points");
            let _entered = span.enter();

-            let mut remove_vectors_builder = MergerBuilder::new(keep_first as MergeFn);
-            let mut manual_vectors_builder = MergerBuilder::new(keep_first as MergeFn);
-            let mut embeddings_builder = MergerBuilder::new(keep_first as MergeFn);
+            let mut remove_vectors_builder = MergerBuilder::new(KeepFirst);
+            let mut manual_vectors_builder = MergerBuilder::new(KeepFirst);
+            let mut embeddings_builder = MergerBuilder::new(KeepFirst);
            let mut add_to_user_provided = RoaringBitmap::new();
            let mut remove_from_user_provided = RoaringBitmap::new();
            let mut params = None;
@@ -673,14 +668,22 @@ pub(crate) fn write_typed_chunk_into_index(
                .get(&embedder_name)
                .map_or(false, |conf| conf.2);
            // FIXME: allow customizing distance
-            let writer = ArroyWrapper::new(index.vector_arroy, embedder_index, binary_quantized);
+            let writers: Vec<_> = crate::vector::arroy_db_range_for_embedder(embedder_index)
+                .map(|k| ArroyWrapper::new(index.vector_arroy, k, binary_quantized))
+                .collect();

            // remove vectors for docids we want them removed
            let merger = remove_vectors_builder.build();
            let mut iter = merger.into_stream_merger_iter()?;
            while let Some((key, _)) = iter.next()? {
                let docid = key.try_into().map(DocumentId::from_be_bytes).unwrap();
-                writer.del_items(wtxn, expected_dimension, docid)?;
+
+                for writer in &writers {
+                    // Uses invariant: vectors are packed in the first writers.
+                    if !writer.del_item(wtxn, expected_dimension, docid)? {
+                        break;
+                    }
+                }
            }

            // add generated embeddings
@@ -708,7 +711,9 @@ pub(crate) fn write_typed_chunk_into_index(
                        embeddings.embedding_count(),
                    )));
                }
-                writer.add_items(wtxn, docid, &embeddings)?;
+                for (embedding, writer) in embeddings.iter().zip(&writers) {
+                    writer.add_item(wtxn, expected_dimension, docid, embedding)?;
+                }
            }

            // perform the manual diff
@@ -719,18 +724,55 @@ pub(crate) fn write_typed_chunk_into_index(
                let (left, _index) = try_split_array_at(key).unwrap();
                let docid = DocumentId::from_be_bytes(left);

-                let vector_deladd_obkv = KvReaderDelAdd::new(value);
+                let vector_deladd_obkv = KvReaderDelAdd::from_slice(value);
                if let Some(value) = vector_deladd_obkv.get(DelAdd::Deletion) {
                    let vector: Vec<f32> = pod_collect_to_vec(value);

-                    writer.del_item(wtxn, docid, &vector)?;
+                    let mut deleted_index = None;
+                    for (index, writer) in writers.iter().enumerate() {
+                        let Some(candidate) = writer.item_vector(wtxn, docid)? else {
+                            // uses invariant: vectors are packed in the first writers.
+                            break;
+                        };
+                        if candidate == vector {
+                            writer.del_item(wtxn, expected_dimension, docid)?;
+                            deleted_index = Some(index);
+                        }
+                    }
+
+                    // 🥲 enforce invariant: vectors are packed in the first writers.
+                    if let Some(deleted_index) = deleted_index {
+                        let mut last_index_with_a_vector = None;
+                        for (index, writer) in writers.iter().enumerate().skip(deleted_index) {
+                            let Some(candidate) = writer.item_vector(wtxn, docid)? else {
+                                break;
+                            };
+                            last_index_with_a_vector = Some((index, candidate));
+                        }
+                        if let Some((last_index, vector)) = last_index_with_a_vector {
+                            // unwrap: computed the index from the list of writers
+                            let writer = writers.get(last_index).unwrap();
+                            writer.del_item(wtxn, expected_dimension, docid)?;
+                            writers.get(deleted_index).unwrap().add_item(
+                                wtxn,
+                                expected_dimension,
+                                docid,
+                                &vector,
+                            )?;
+                        }
+                    }
                }

                if let Some(value) = vector_deladd_obkv.get(DelAdd::Addition) {
                    let vector = pod_collect_to_vec(value);

                    // overflow was detected during vector extraction.
-                    writer.add_item(wtxn, docid, &vector)?;
+                    for writer in &writers {
+                        if !writer.contains_item(wtxn, expected_dimension, docid)? {
+                            writer.add_item(wtxn, expected_dimension, docid, &vector)?;
+                            break;
+                        }
+                    }
                }
            }

@@ -750,9 +792,13 @@ fn extract_geo_point(value: &[u8], docid: DocumentId) -> GeoPoint {
    GeoPoint::new(xyz_point, (docid, point))
 }

-fn merge_word_docids_reader_into_fst(
-    merger: Merger<CursorClonableMmap, MergeFn>,
-) -> Result<fst::Set<Vec<u8>>> {
+fn merge_word_docids_reader_into_fst<MF>(
+    merger: Merger<CursorClonableMmap, MF>,
+) -> Result<fst::Set<Vec<u8>>>
+where
+    MF: MergeFunction,
+    crate::Error: From<MF::Error>,
+{
    let mut iter = merger.into_stream_merger_iter()?;
    let mut builder = fst::SetBuilder::memory();

@@ -766,8 +812,8 @@ fn merge_word_docids_reader_into_fst(
 /// Write provided entries in database using serialize_value function.
 /// merge_values function is used if an entry already exist in the database.
 #[tracing::instrument(level = "trace", skip_all, target = "indexing::write_db")]
-fn write_entries_into_database<R, K, V, FS, FM>(
-    merger: Merger<R, MergeFn>,
+fn write_entries_into_database<R, K, V, FS, FM, MF>(
+    merger: Merger<R, MF>,
    database: &heed::Database<K, V>,
    wtxn: &mut RwTxn<'_>,
    serialize_value: FS,
@@ -777,6 +823,8 @@ where
    R: io::Read + io::Seek,
    FS: for<'a> Fn(&'a [u8], &'a mut Vec<u8>) -> Result<&'a [u8]>,
    FM: for<'a> Fn(&[u8], &[u8], &'a mut Vec<u8>) -> Result<Option<&'a [u8]>>,
+    MF: MergeFunction,
+    crate::Error: From<MF::Error>,
 {
    let mut buffer = Vec::new();
    let database = database.remap_types::<Bytes, Bytes>();
@@ -803,20 +851,22 @@ where
 /// Akin to the `write_entries_into_database` function but specialized
 /// for the case when we only index additional searchable fields only.
 #[tracing::instrument(level = "trace", skip_all, target = "indexing::write_db")]
-fn write_proximity_entries_into_database_additional_searchables<R>(
-    merger: Merger<R, MergeFn>,
+fn write_proximity_entries_into_database_additional_searchables<R, MF>(
+    merger: Merger<R, MF>,
    database: &heed::Database<U8StrStrCodec, CboRoaringBitmapCodec>,
    wtxn: &mut RwTxn<'_>,
 ) -> Result<()>
 where
    R: io::Read + io::Seek,
+    MF: MergeFunction,
+    crate::Error: From<MF::Error>,
 {
    let mut iter = merger.into_stream_merger_iter()?;
    while let Some((key, value)) = iter.next()? {
        if valid_lmdb_key(key) {
            let (proximity_to_insert, word1, word2) =
                U8StrStrCodec::bytes_decode(key).map_err(heed::Error::Decoding)?;
-            let data_to_insert = match KvReaderDelAdd::new(value).get(DelAdd::Addition) {
+            let data_to_insert = match KvReaderDelAdd::from_slice(value).get(DelAdd::Addition) {
                Some(value) => {
                    CboRoaringBitmapCodec::bytes_decode(value).map_err(heed::Error::Decoding)?
                }
--- a/milli/src/update/mod.rs
+++ b/milli/src/update/mod.rs
@@ -1,11 +1,9 @@
-pub use self::available_documents_ids::AvailableDocumentsIds;
+pub use self::available_ids::AvailableIds;
 pub use self::clear_documents::ClearDocuments;
+pub use self::concurrent_available_ids::ConcurrentAvailableIds;
 pub use self::facet::bulk::FacetsUpdateBulk;
 pub use self::facet::incremental::FacetsUpdateIncrementalInner;
-pub use self::index_documents::{
-    merge_cbo_roaring_bitmaps, merge_roaring_bitmaps, DocumentAdditionResult, DocumentId,
-    IndexDocuments, IndexDocumentsConfig, IndexDocumentsMethod, MergeFn,
-};
+pub use self::index_documents::*;
 pub use self::indexer_config::IndexerConfig;
 pub use self::settings::{validate_embedding_settings, Setting, Settings};
 pub use self::update_step::UpdateIndexingStep;
@@ -13,12 +11,14 @@ pub use self::word_prefix_docids::WordPrefixDocids;
 pub use self::words_prefix_integer_docids::WordPrefixIntegerDocids;
 pub use self::words_prefixes_fst::WordsPrefixesFst;

-mod available_documents_ids;
+mod available_ids;
 mod clear_documents;
+mod concurrent_available_ids;
 pub(crate) mod del_add;
 pub(crate) mod facet;
 mod index_documents;
 mod indexer_config;
+pub mod new;
 mod settings;
 mod update_step;
 mod word_prefix_docids;
--- a/milli/src/update/new/channel.rs
+++ b/milli/src/update/new/channel.rs
@@ -0,0 +1,522 @@
+use std::marker::PhantomData;
+use std::sync::atomic::Ordering;
+
+use crossbeam_channel::{IntoIter, Receiver, SendError, Sender};
+use heed::types::Bytes;
+use memmap2::Mmap;
+
+use super::extract::{FacetKind, HashMapMerger};
+use super::StdResult;
+use crate::index::main_key::{DOCUMENTS_IDS_KEY, WORDS_FST_KEY};
+use crate::update::new::KvReaderFieldId;
+use crate::{DocumentId, Index};
+
+/// The capacity of the channel is currently in number of messages.
+pub fn merger_writer_channel(cap: usize) -> (MergerSender, WriterReceiver) {
+    let (sender, receiver) = crossbeam_channel::bounded(cap);
+    (
+        MergerSender {
+            sender,
+            send_count: Default::default(),
+            writer_contentious_count: Default::default(),
+            merger_contentious_count: Default::default(),
+        },
+        WriterReceiver(receiver),
+    )
+}
+
+/// The capacity of the channel is currently in number of messages.
+pub fn extractors_merger_channels(cap: usize) -> (ExtractorSender, MergerReceiver) {
+    let (sender, receiver) = crossbeam_channel::bounded(cap);
+    (ExtractorSender(sender), MergerReceiver(receiver))
+}
+
+pub enum KeyValueEntry {
+    SmallInMemory { key_length: usize, data: Box<[u8]> },
+    LargeOnDisk { key: Box<[u8]>, value: Mmap },
+}
+
+impl KeyValueEntry {
+    pub fn from_small_key_value(key: &[u8], value: &[u8]) -> Self {
+        let mut data = Vec::with_capacity(key.len() + value.len());
+        data.extend_from_slice(key);
+        data.extend_from_slice(value);
+        KeyValueEntry::SmallInMemory { key_length: key.len(), data: data.into_boxed_slice() }
+    }
+
+    pub fn from_large_key_value(key: &[u8], value: Mmap) -> Self {
+        KeyValueEntry::LargeOnDisk { key: key.to_vec().into_boxed_slice(), value }
+    }
+
+    pub fn key(&self) -> &[u8] {
+        match self {
+            KeyValueEntry::SmallInMemory { key_length, data } => &data.as_ref()[..*key_length],
+            KeyValueEntry::LargeOnDisk { key, value: _ } => key.as_ref(),
+        }
+    }
+
+    pub fn value(&self) -> &[u8] {
+        match self {
+            KeyValueEntry::SmallInMemory { key_length, data } => &data.as_ref()[*key_length..],
+            KeyValueEntry::LargeOnDisk { key: _, value } => value.as_ref(),
+        }
+    }
+}
+
+pub struct KeyEntry {
+    data: Box<[u8]>,
+}
+
+impl KeyEntry {
+    pub fn from_key(key: &[u8]) -> Self {
+        KeyEntry { data: key.to_vec().into_boxed_slice() }
+    }
+
+    pub fn entry(&self) -> &[u8] {
+        self.data.as_ref()
+    }
+}
+
+pub enum EntryOperation {
+    Delete(KeyEntry),
+    Write(KeyValueEntry),
+}
+
+pub struct DocumentEntry {
+    docid: DocumentId,
+    content: Box<[u8]>,
+}
+
+impl DocumentEntry {
+    pub fn new_uncompressed(docid: DocumentId, content: Box<KvReaderFieldId>) -> Self {
+        DocumentEntry { docid, content: content.into() }
+    }
+
+    pub fn new_compressed(docid: DocumentId, content: Box<[u8]>) -> Self {
+        DocumentEntry { docid, content }
+    }
+
+    pub fn key(&self) -> [u8; 4] {
+        self.docid.to_be_bytes()
+    }
+
+    pub fn content(&self) -> &[u8] {
+        &self.content
+    }
+}
+
+pub struct DocumentDeletionEntry(DocumentId);
+
+impl DocumentDeletionEntry {
+    pub fn key(&self) -> [u8; 4] {
+        self.0.to_be_bytes()
+    }
+}
+
+pub struct WriterOperation {
+    database: Database,
+    entry: EntryOperation,
+}
+
+pub enum Database {
+    Documents,
+    ExactWordDocids,
+    FidWordCountDocids,
+    Main,
+    WordDocids,
+    WordFidDocids,
+    WordPairProximityDocids,
+    WordPositionDocids,
+    FacetIdIsNullDocids,
+    FacetIdIsEmptyDocids,
+    FacetIdExistsDocids,
+    FacetIdF64NumberDocids,
+    FacetIdStringDocids,
+}
+
+impl Database {
+    pub fn database(&self, index: &Index) -> heed::Database<Bytes, Bytes> {
+        match self {
+            Database::Documents => index.documents.remap_types(),
+            Database::ExactWordDocids => index.exact_word_docids.remap_types(),
+            Database::Main => index.main.remap_types(),
+            Database::WordDocids => index.word_docids.remap_types(),
+            Database::WordFidDocids => index.word_fid_docids.remap_types(),
+            Database::WordPositionDocids => index.word_position_docids.remap_types(),
+            Database::FidWordCountDocids => index.field_id_word_count_docids.remap_types(),
+            Database::WordPairProximityDocids => index.word_pair_proximity_docids.remap_types(),
+            Database::FacetIdIsNullDocids => index.facet_id_is_null_docids.remap_types(),
+            Database::FacetIdIsEmptyDocids => index.facet_id_is_empty_docids.remap_types(),
+            Database::FacetIdExistsDocids => index.facet_id_exists_docids.remap_types(),
+            Database::FacetIdF64NumberDocids => index.facet_id_f64_docids.remap_types(),
+            Database::FacetIdStringDocids => index.facet_id_string_docids.remap_types(),
+        }
+    }
+}
+
+impl WriterOperation {
+    pub fn database(&self, index: &Index) -> heed::Database<Bytes, Bytes> {
+        self.database.database(index)
+    }
+
+    pub fn entry(self) -> EntryOperation {
+        self.entry
+    }
+}
+
+pub struct WriterReceiver(Receiver<WriterOperation>);
+
+impl IntoIterator for WriterReceiver {
+    type Item = WriterOperation;
+    type IntoIter = IntoIter<Self::Item>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        self.0.into_iter()
+    }
+}
+
+pub struct MergerSender {
+    sender: Sender<WriterOperation>,
+    /// The number of message we send in total in the channel.
+    send_count: std::sync::atomic::AtomicUsize,
+    /// The number of times we sent something in a channel that was full.
+    writer_contentious_count: std::sync::atomic::AtomicUsize,
+    /// The number of times we sent something in a channel that was empty.
+    merger_contentious_count: std::sync::atomic::AtomicUsize,
+}
+
+impl Drop for MergerSender {
+    fn drop(&mut self) {
+        eprintln!(
+            "Merger channel stats: {} sends, {} writer contentions ({}%), {} merger contentions ({}%)",
+            self.send_count.load(Ordering::SeqCst),
+            self.writer_contentious_count.load(Ordering::SeqCst),
+            (self.writer_contentious_count.load(Ordering::SeqCst) as f32 / self.send_count.load(Ordering::SeqCst) as f32) * 100.0,
+            self.merger_contentious_count.load(Ordering::SeqCst),
+            (self.merger_contentious_count.load(Ordering::SeqCst) as f32 / self.send_count.load(Ordering::SeqCst) as f32) * 100.0
+        )
+    }
+}
+
+impl MergerSender {
+    pub fn main(&self) -> MainSender<'_> {
+        MainSender(self)
+    }
+
+    pub fn docids<D: DatabaseType>(&self) -> WordDocidsSender<'_, D> {
+        WordDocidsSender { sender: self, _marker: PhantomData }
+    }
+
+    pub fn facet_docids(&self) -> FacetDocidsSender<'_> {
+        FacetDocidsSender { sender: self }
+    }
+
+    pub fn documents(&self) -> DocumentsSender<'_> {
+        DocumentsSender(self)
+    }
+
+    pub fn send_documents_ids(&self, bitmap: &[u8]) -> StdResult<(), SendError<()>> {
+        let entry = EntryOperation::Write(KeyValueEntry::from_small_key_value(
+            DOCUMENTS_IDS_KEY.as_bytes(),
+            bitmap,
+        ));
+        match self.send(WriterOperation { database: Database::Main, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+
+    fn send(&self, op: WriterOperation) -> StdResult<(), SendError<()>> {
+        if self.sender.is_full() {
+            self.writer_contentious_count.fetch_add(1, Ordering::SeqCst);
+        }
+        if self.sender.is_empty() {
+            self.merger_contentious_count.fetch_add(1, Ordering::SeqCst);
+        }
+        self.send_count.fetch_add(1, Ordering::SeqCst);
+        match self.sender.send(op) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+}
+
+pub struct MainSender<'a>(&'a MergerSender);
+
+impl MainSender<'_> {
+    pub fn write_words_fst(&self, value: Mmap) -> StdResult<(), SendError<()>> {
+        let entry = EntryOperation::Write(KeyValueEntry::from_large_key_value(
+            WORDS_FST_KEY.as_bytes(),
+            value,
+        ));
+        match self.0.send(WriterOperation { database: Database::Main, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+
+    pub fn delete(&self, key: &[u8]) -> StdResult<(), SendError<()>> {
+        let entry = EntryOperation::Delete(KeyEntry::from_key(key));
+        match self.0.send(WriterOperation { database: Database::Main, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+}
+
+pub enum ExactWordDocids {}
+pub enum FidWordCountDocids {}
+pub enum WordDocids {}
+pub enum WordFidDocids {}
+pub enum WordPairProximityDocids {}
+pub enum WordPositionDocids {}
+pub enum FacetDocids {}
+
+pub trait DatabaseType {
+    const DATABASE: Database;
+}
+
+pub trait MergerOperationType {
+    fn new_merger_operation(merger: HashMapMerger) -> MergerOperation;
+}
+
+impl DatabaseType for ExactWordDocids {
+    const DATABASE: Database = Database::ExactWordDocids;
+}
+
+impl MergerOperationType for ExactWordDocids {
+    fn new_merger_operation(merger: HashMapMerger) -> MergerOperation {
+        MergerOperation::ExactWordDocidsMerger(merger)
+    }
+}
+
+impl DatabaseType for FidWordCountDocids {
+    const DATABASE: Database = Database::FidWordCountDocids;
+}
+
+impl MergerOperationType for FidWordCountDocids {
+    fn new_merger_operation(merger: HashMapMerger) -> MergerOperation {
+        MergerOperation::FidWordCountDocidsMerger(merger)
+    }
+}
+
+impl DatabaseType for WordDocids {
+    const DATABASE: Database = Database::WordDocids;
+}
+
+impl MergerOperationType for WordDocids {
+    fn new_merger_operation(merger: HashMapMerger) -> MergerOperation {
+        MergerOperation::WordDocidsMerger(merger)
+    }
+}
+
+impl DatabaseType for WordFidDocids {
+    const DATABASE: Database = Database::WordFidDocids;
+}
+
+impl MergerOperationType for WordFidDocids {
+    fn new_merger_operation(merger: HashMapMerger) -> MergerOperation {
+        MergerOperation::WordFidDocidsMerger(merger)
+    }
+}
+
+impl DatabaseType for WordPairProximityDocids {
+    const DATABASE: Database = Database::WordPairProximityDocids;
+}
+
+impl MergerOperationType for WordPairProximityDocids {
+    fn new_merger_operation(merger: HashMapMerger) -> MergerOperation {
+        MergerOperation::WordPairProximityDocidsMerger(merger)
+    }
+}
+
+impl DatabaseType for WordPositionDocids {
+    const DATABASE: Database = Database::WordPositionDocids;
+}
+
+impl MergerOperationType for WordPositionDocids {
+    fn new_merger_operation(merger: HashMapMerger) -> MergerOperation {
+        MergerOperation::WordPositionDocidsMerger(merger)
+    }
+}
+
+impl MergerOperationType for FacetDocids {
+    fn new_merger_operation(merger: HashMapMerger) -> MergerOperation {
+        MergerOperation::FacetDocidsMerger(merger)
+    }
+}
+
+pub trait DocidsSender {
+    fn write(&self, key: &[u8], value: &[u8]) -> StdResult<(), SendError<()>>;
+    fn delete(&self, key: &[u8]) -> StdResult<(), SendError<()>>;
+}
+
+pub struct WordDocidsSender<'a, D> {
+    sender: &'a MergerSender,
+    _marker: PhantomData<D>,
+}
+
+impl<D: DatabaseType> DocidsSender for WordDocidsSender<'_, D> {
+    fn write(&self, key: &[u8], value: &[u8]) -> StdResult<(), SendError<()>> {
+        let entry = EntryOperation::Write(KeyValueEntry::from_small_key_value(key, value));
+        match self.sender.send(WriterOperation { database: D::DATABASE, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+
+    fn delete(&self, key: &[u8]) -> StdResult<(), SendError<()>> {
+        let entry = EntryOperation::Delete(KeyEntry::from_key(key));
+        match self.sender.send(WriterOperation { database: D::DATABASE, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+}
+
+pub struct FacetDocidsSender<'a> {
+    sender: &'a MergerSender,
+}
+
+impl DocidsSender for FacetDocidsSender<'_> {
+    fn write(&self, key: &[u8], value: &[u8]) -> StdResult<(), SendError<()>> {
+        let (database, key) = self.extract_database(key);
+        let entry = EntryOperation::Write(KeyValueEntry::from_small_key_value(key, value));
+        match self.sender.send(WriterOperation { database, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+
+    fn delete(&self, key: &[u8]) -> StdResult<(), SendError<()>> {
+        let (database, key) = self.extract_database(key);
+        let entry = EntryOperation::Delete(KeyEntry::from_key(key));
+        match self.sender.send(WriterOperation { database, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+}
+
+impl FacetDocidsSender<'_> {
+    fn extract_database<'a>(&self, key: &'a [u8]) -> (Database, &'a [u8]) {
+        let database = match FacetKind::from(key[0]) {
+            FacetKind::Number => Database::FacetIdF64NumberDocids,
+            FacetKind::String => Database::FacetIdStringDocids,
+            FacetKind::Null => Database::FacetIdIsNullDocids,
+            FacetKind::Empty => Database::FacetIdIsEmptyDocids,
+            FacetKind::Exists => Database::FacetIdExistsDocids,
+        };
+        (database, &key[1..])
+    }
+}
+
+pub struct DocumentsSender<'a>(&'a MergerSender);
+
+impl DocumentsSender<'_> {
+    /// TODO do that efficiently
+    pub fn uncompressed(
+        &self,
+        docid: DocumentId,
+        document: &KvReaderFieldId,
+    ) -> StdResult<(), SendError<()>> {
+        let entry = EntryOperation::Write(KeyValueEntry::from_small_key_value(
+            &docid.to_be_bytes(),
+            document.as_bytes(),
+        ));
+        match self.0.send(WriterOperation { database: Database::Documents, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+
+    pub fn delete(&self, docid: DocumentId) -> StdResult<(), SendError<()>> {
+        let entry = EntryOperation::Delete(KeyEntry::from_key(&docid.to_be_bytes()));
+        match self.0.send(WriterOperation { database: Database::Documents, entry }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+}
+
+pub enum MergerOperation {
+    ExactWordDocidsMerger(HashMapMerger),
+    FidWordCountDocidsMerger(HashMapMerger),
+    WordDocidsMerger(HashMapMerger),
+    WordFidDocidsMerger(HashMapMerger),
+    WordPairProximityDocidsMerger(HashMapMerger),
+    WordPositionDocidsMerger(HashMapMerger),
+    FacetDocidsMerger(HashMapMerger),
+    DeleteDocument { docid: DocumentId },
+    InsertDocument { docid: DocumentId, document: Box<KvReaderFieldId> },
+    FinishedDocument,
+}
+
+pub struct MergerReceiver(Receiver<MergerOperation>);
+
+impl IntoIterator for MergerReceiver {
+    type Item = MergerOperation;
+    type IntoIter = IntoIter<Self::Item>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        self.0.into_iter()
+    }
+}
+
+pub struct ExtractorSender(Sender<MergerOperation>);
+
+impl ExtractorSender {
+    pub fn document_sender(&self) -> DocumentSender<'_> {
+        DocumentSender(Some(&self.0))
+    }
+
+    pub fn send_searchable<D: MergerOperationType>(
+        &self,
+        merger: HashMapMerger,
+    ) -> StdResult<(), SendError<()>> {
+        match self.0.send(D::new_merger_operation(merger)) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+}
+
+pub struct DocumentSender<'a>(Option<&'a Sender<MergerOperation>>);
+
+impl DocumentSender<'_> {
+    pub fn insert(
+        &self,
+        docid: DocumentId,
+        document: Box<KvReaderFieldId>,
+    ) -> StdResult<(), SendError<()>> {
+        let sender = self.0.unwrap();
+        match sender.send(MergerOperation::InsertDocument { docid, document }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+
+    pub fn delete(&self, docid: DocumentId) -> StdResult<(), SendError<()>> {
+        let sender = self.0.unwrap();
+        match sender.send(MergerOperation::DeleteDocument { docid }) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+
+    pub fn finish(mut self) -> StdResult<(), SendError<()>> {
+        let sender = self.0.take().unwrap();
+        match sender.send(MergerOperation::FinishedDocument) {
+            Ok(()) => Ok(()),
+            Err(SendError(_)) => Err(SendError(())),
+        }
+    }
+}
+
+impl Drop for DocumentSender<'_> {
+    fn drop(&mut self) {
+        if let Some(sender) = self.0.take() {
+            sender.send(MergerOperation::FinishedDocument);
+        }
+    }
+}
--- a/milli/src/update/new/document_change.rs
+++ b/milli/src/update/new/document_change.rs
@@ -0,0 +1,96 @@
+use heed::RoTxn;
+use obkv::KvReader;
+
+use crate::update::new::KvReaderFieldId;
+use crate::{DocumentId, FieldId, Index, Result};
+
+pub enum DocumentChange {
+    Deletion(Deletion),
+    Update(Update),
+    Insertion(Insertion),
+}
+
+pub struct Deletion {
+    docid: DocumentId,
+    current: Box<KvReaderFieldId>,
+}
+
+pub struct Update {
+    docid: DocumentId,
+    current: Box<KvReaderFieldId>,
+    new: Box<KvReaderFieldId>,
+}
+
+pub struct Insertion {
+    docid: DocumentId,
+    new: Box<KvReaderFieldId>,
+}
+
+impl DocumentChange {
+    pub fn docid(&self) -> DocumentId {
+        match &self {
+            Self::Deletion(inner) => inner.docid(),
+            Self::Update(inner) => inner.docid(),
+            Self::Insertion(inner) => inner.docid(),
+        }
+    }
+}
+
+impl Deletion {
+    pub fn create(docid: DocumentId, current: Box<KvReaderFieldId>) -> Self {
+        Self { docid, current }
+    }
+
+    pub fn docid(&self) -> DocumentId {
+        self.docid
+    }
+
+    // TODO shouldn't we use the one in self?
+    pub fn current<'a>(
+        &self,
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<&'a KvReader<FieldId>>> {
+        index.documents.get(rtxn, &self.docid).map_err(crate::Error::from)
+    }
+}
+
+impl Insertion {
+    pub fn create(docid: DocumentId, new: Box<KvReaderFieldId>) -> Self {
+        Insertion { docid, new }
+    }
+
+    pub fn docid(&self) -> DocumentId {
+        self.docid
+    }
+
+    pub fn new(&self) -> &KvReader<FieldId> {
+        self.new.as_ref()
+    }
+}
+
+impl Update {
+    pub fn create(
+        docid: DocumentId,
+        current: Box<KvReaderFieldId>,
+        new: Box<KvReaderFieldId>,
+    ) -> Self {
+        Update { docid, current, new }
+    }
+
+    pub fn docid(&self) -> DocumentId {
+        self.docid
+    }
+
+    pub fn current<'a>(
+        &self,
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<&'a KvReader<FieldId>>> {
+        index.documents.get(rtxn, &self.docid).map_err(crate::Error::from)
+    }
+
+    pub fn new(&self) -> &KvReader<FieldId> {
+        self.new.as_ref()
+    }
+}
--- a/milli/src/update/new/extract/cache.rs
+++ b/milli/src/update/new/extract/cache.rs
@@ -0,0 +1,149 @@
+use std::collections::HashMap;
+
+use roaring::RoaringBitmap;
+use smallvec::SmallVec;
+
+pub const KEY_SIZE: usize = 12;
+
+#[derive(Debug)]
+pub struct CboCachedSorter {
+    cache: HashMap<SmallVec<[u8; KEY_SIZE]>, DelAddRoaringBitmap>,
+    total_insertions: usize,
+    fitted_in_key: usize,
+}
+
+impl CboCachedSorter {
+    pub fn new() -> Self {
+        CboCachedSorter { cache: HashMap::new(), total_insertions: 0, fitted_in_key: 0 }
+    }
+}
+
+impl CboCachedSorter {
+    pub fn insert_del_u32(&mut self, key: &[u8], n: u32) {
+        match self.cache.get_mut(key) {
+            Some(DelAddRoaringBitmap { del, add: _ }) => {
+                del.get_or_insert_with(RoaringBitmap::default).insert(n);
+            }
+            None => {
+                self.total_insertions += 1;
+                self.fitted_in_key += (key.len() <= KEY_SIZE) as usize;
+                let value = DelAddRoaringBitmap::new_del_u32(n);
+                assert!(self.cache.insert(key.into(), value).is_none());
+            }
+        }
+    }
+
+    pub fn insert_del(&mut self, key: &[u8], bitmap: RoaringBitmap) {
+        match self.cache.get_mut(key) {
+            Some(DelAddRoaringBitmap { del, add: _ }) => {
+                *del.get_or_insert_with(RoaringBitmap::default) |= bitmap;
+            }
+            None => {
+                self.total_insertions += 1;
+                self.fitted_in_key += (key.len() <= KEY_SIZE) as usize;
+                let value = DelAddRoaringBitmap::new_del(bitmap);
+                assert!(self.cache.insert(key.into(), value).is_none());
+            }
+        }
+    }
+
+    pub fn insert_add_u32(&mut self, key: &[u8], n: u32) {
+        match self.cache.get_mut(key) {
+            Some(DelAddRoaringBitmap { del: _, add }) => {
+                add.get_or_insert_with(RoaringBitmap::default).insert(n);
+            }
+            None => {
+                self.total_insertions += 1;
+                self.fitted_in_key += (key.len() <= KEY_SIZE) as usize;
+                let value = DelAddRoaringBitmap::new_add_u32(n);
+                assert!(self.cache.insert(key.into(), value).is_none());
+            }
+        }
+    }
+
+    pub fn insert_add(&mut self, key: &[u8], bitmap: RoaringBitmap) {
+        match self.cache.get_mut(key) {
+            Some(DelAddRoaringBitmap { del: _, add }) => {
+                *add.get_or_insert_with(RoaringBitmap::default) |= bitmap;
+            }
+            None => {
+                self.total_insertions += 1;
+                self.fitted_in_key += (key.len() <= KEY_SIZE) as usize;
+                let value = DelAddRoaringBitmap::new_add(bitmap);
+                assert!(self.cache.insert(key.into(), value).is_none());
+            }
+        }
+    }
+
+    pub fn insert_del_add_u32(&mut self, key: &[u8], n: u32) {
+        match self.cache.get_mut(key) {
+            Some(DelAddRoaringBitmap { del, add }) => {
+                del.get_or_insert_with(RoaringBitmap::default).insert(n);
+                add.get_or_insert_with(RoaringBitmap::default).insert(n);
+            }
+            None => {
+                self.total_insertions += 1;
+                self.fitted_in_key += (key.len() <= KEY_SIZE) as usize;
+                let value = DelAddRoaringBitmap::new_del_add_u32(n);
+                assert!(self.cache.insert(key.into(), value).is_none());
+            }
+        }
+    }
+
+    pub fn into_sorter(self) -> HashMap<SmallVec<[u8; KEY_SIZE]>, DelAddRoaringBitmap> {
+        eprintln!(
+            "LruCache stats: {} <= {KEY_SIZE} bytes ({}%) on a total of {} insertions",
+            self.fitted_in_key,
+            (self.fitted_in_key as f32 / self.total_insertions as f32) * 100.0,
+            self.total_insertions,
+        );
+
+        self.cache
+    }
+}
+
+#[derive(Debug, Clone, Default)]
+pub struct DelAddRoaringBitmap {
+    pub(crate) del: Option<RoaringBitmap>,
+    pub(crate) add: Option<RoaringBitmap>,
+}
+
+impl DelAddRoaringBitmap {
+    fn new_del_add_u32(n: u32) -> Self {
+        DelAddRoaringBitmap {
+            del: Some(RoaringBitmap::from([n])),
+            add: Some(RoaringBitmap::from([n])),
+        }
+    }
+
+    fn new_del(bitmap: RoaringBitmap) -> Self {
+        DelAddRoaringBitmap { del: Some(bitmap), add: None }
+    }
+
+    fn new_del_u32(n: u32) -> Self {
+        DelAddRoaringBitmap { del: Some(RoaringBitmap::from([n])), add: None }
+    }
+
+    fn new_add(bitmap: RoaringBitmap) -> Self {
+        DelAddRoaringBitmap { del: None, add: Some(bitmap) }
+    }
+
+    fn new_add_u32(n: u32) -> Self {
+        DelAddRoaringBitmap { del: None, add: Some(RoaringBitmap::from([n])) }
+    }
+
+    pub fn merge_with(&mut self, other: DelAddRoaringBitmap) {
+        self.del = match (self.del.take(), other.del) {
+            (None, None) => None,
+            (None, Some(other)) => Some(other),
+            (Some(this), None) => Some(this),
+            (Some(this), Some(other)) => Some(this | other),
+        };
+        self.add = match (self.add.take(), other.add) {
+            (None, None) => None,
+            (None, Some(other)) => Some(other),
+            (Some(this), None) => Some(this),
+            (Some(this), Some(other)) => Some(this | other),
+        };
+    }
+}
--- a/milli/src/update/new/extract/faceted/extract_facets.rs
+++ b/milli/src/update/new/extract/faceted/extract_facets.rs
@@ -0,0 +1,240 @@
+use std::collections::HashSet;
+
+use heed::RoTxn;
+use rayon::iter::{IntoParallelIterator, ParallelBridge, ParallelIterator};
+use serde_json::Value;
+
+use super::super::cache::CboCachedSorter;
+use super::facet_document::extract_document_facets;
+use super::FacetKind;
+use crate::facet::value_encoding::f64_into_bytes;
+use crate::update::new::extract::{DocidsExtractor, HashMapMerger};
+use crate::update::new::{DocumentChange, ItemsPool};
+use crate::update::GrenadParameters;
+use crate::{DocumentId, FieldId, GlobalFieldsIdsMap, Index, Result, MAX_FACET_VALUE_LENGTH};
+pub struct FacetedDocidsExtractor;
+
+impl FacetedDocidsExtractor {
+    fn extract_document_change(
+        rtxn: &RoTxn,
+        index: &Index,
+        buffer: &mut Vec<u8>,
+        fields_ids_map: &mut GlobalFieldsIdsMap,
+        attributes_to_extract: &[&str],
+        cached_sorter: &mut CboCachedSorter,
+        document_change: DocumentChange,
+    ) -> Result<()> {
+        match document_change {
+            DocumentChange::Deletion(inner) => extract_document_facets(
+                attributes_to_extract,
+                inner.current(rtxn, index)?.unwrap(),
+                fields_ids_map,
+                &mut |fid, value| {
+                    Self::facet_fn_with_options(
+                        buffer,
+                        cached_sorter,
+                        CboCachedSorter::insert_del_u32,
+                        inner.docid(),
+                        fid,
+                        value,
+                    )
+                },
+            ),
+            DocumentChange::Update(inner) => {
+                extract_document_facets(
+                    attributes_to_extract,
+                    inner.current(rtxn, index)?.unwrap(),
+                    fields_ids_map,
+                    &mut |fid, value| {
+                        Self::facet_fn_with_options(
+                            buffer,
+                            cached_sorter,
+                            CboCachedSorter::insert_del_u32,
+                            inner.docid(),
+                            fid,
+                            value,
+                        )
+                    },
+                )?;
+
+                extract_document_facets(
+                    attributes_to_extract,
+                    inner.new(),
+                    fields_ids_map,
+                    &mut |fid, value| {
+                        Self::facet_fn_with_options(
+                            buffer,
+                            cached_sorter,
+                            CboCachedSorter::insert_add_u32,
+                            inner.docid(),
+                            fid,
+                            value,
+                        )
+                    },
+                )
+            }
+            DocumentChange::Insertion(inner) => extract_document_facets(
+                attributes_to_extract,
+                inner.new(),
+                fields_ids_map,
+                &mut |fid, value| {
+                    Self::facet_fn_with_options(
+                        buffer,
+                        cached_sorter,
+                        CboCachedSorter::insert_add_u32,
+                        inner.docid(),
+                        fid,
+                        value,
+                    )
+                },
+            ),
+        }
+    }
+
+    fn facet_fn_with_options(
+        buffer: &mut Vec<u8>,
+        cached_sorter: &mut CboCachedSorter,
+        cache_fn: impl Fn(&mut CboCachedSorter, &[u8], u32),
+        docid: DocumentId,
+        fid: FieldId,
+        value: &Value,
+    ) -> Result<()> {
+        // Exists
+        // key: fid
+        buffer.clear();
+        buffer.push(FacetKind::Exists as u8);
+        buffer.extend_from_slice(&fid.to_be_bytes());
+        cache_fn(cached_sorter, &*buffer, docid);
+
+        match value {
+            // Number
+            // key: fid - level - orderedf64 - orignalf64
+            Value::Number(number) => {
+                if let Some((n, ordered)) =
+                    number.as_f64().and_then(|n| f64_into_bytes(n).map(|ordered| (n, ordered)))
+                {
+                    buffer.clear();
+                    buffer.push(FacetKind::Number as u8);
+                    buffer.extend_from_slice(&fid.to_be_bytes());
+                    buffer.push(1); // level 0
+                    buffer.extend_from_slice(&ordered);
+                    buffer.extend_from_slice(&n.to_be_bytes());
+
+                    Ok(cache_fn(cached_sorter, &*buffer, docid))
+                } else {
+                    Ok(())
+                }
+            }
+            // String
+            // key: fid - level - truncated_string
+            Value::String(s) => {
+                let truncated = truncate_str(s);
+                buffer.clear();
+                buffer.push(FacetKind::String as u8);
+                buffer.extend_from_slice(&fid.to_be_bytes());
+                buffer.push(1); // level 0
+                buffer.extend_from_slice(truncated.as_bytes());
+                Ok(cache_fn(cached_sorter, &*buffer, docid))
+            }
+            // Null
+            // key: fid
+            Value::Null => {
+                buffer.clear();
+                buffer.push(FacetKind::Null as u8);
+                buffer.extend_from_slice(&fid.to_be_bytes());
+                Ok(cache_fn(cached_sorter, &*buffer, docid))
+            }
+            // Empty
+            // key: fid
+            Value::Array(a) if a.is_empty() => {
+                buffer.clear();
+                buffer.push(FacetKind::Empty as u8);
+                buffer.extend_from_slice(&fid.to_be_bytes());
+                Ok(cache_fn(cached_sorter, &*buffer, docid))
+            }
+            Value::Object(o) if o.is_empty() => {
+                buffer.clear();
+                buffer.push(FacetKind::Empty as u8);
+                buffer.extend_from_slice(&fid.to_be_bytes());
+                Ok(cache_fn(cached_sorter, &*buffer, docid))
+            }
+            // Otherwise, do nothing
+            /// TODO: What about Value::Bool?
+            _ => Ok(()),
+        }
+    }
+
+    fn attributes_to_extract<'a>(rtxn: &'a RoTxn, index: &'a Index) -> Result<HashSet<String>> {
+        index.user_defined_faceted_fields(rtxn)
+    }
+}
+
+/// Truncates a string to the biggest valid LMDB key size.
+fn truncate_str(s: &str) -> &str {
+    let index = s
+        .char_indices()
+        .map(|(idx, _)| idx)
+        .chain(std::iter::once(s.len()))
+        .take_while(|idx| idx <= &MAX_FACET_VALUE_LENGTH)
+        .last();
+
+    &s[..index.unwrap_or(0)]
+}
+
+impl DocidsExtractor for FacetedDocidsExtractor {
+    #[tracing::instrument(level = "trace", skip_all, target = "indexing::extract::faceted")]
+    fn run_extraction(
+        index: &Index,
+        fields_ids_map: &GlobalFieldsIdsMap,
+        indexer: GrenadParameters,
+        document_changes: impl IntoParallelIterator<Item = Result<DocumentChange>>,
+    ) -> Result<HashMapMerger> {
+        let max_memory = indexer.max_memory_by_thread();
+
+        let rtxn = index.read_txn()?;
+        let attributes_to_extract = Self::attributes_to_extract(&rtxn, index)?;
+        let attributes_to_extract: Vec<_> =
+            attributes_to_extract.iter().map(|s| s.as_ref()).collect();
+
+        let context_pool = ItemsPool::new(|| {
+            Ok((index.read_txn()?, fields_ids_map.clone(), Vec::new(), CboCachedSorter::new()))
+        });
+
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::extract", "docids_extraction");
+            let _entered = span.enter();
+            document_changes.into_par_iter().try_for_each(|document_change| {
+                context_pool.with(|(rtxn, fields_ids_map, buffer, cached_sorter)| {
+                    Self::extract_document_change(
+                        &*rtxn,
+                        index,
+                        buffer,
+                        fields_ids_map,
+                        &attributes_to_extract,
+                        cached_sorter,
+                        document_change?,
+                    )
+                })
+            })?;
+        }
+        {
+            let mut builder = HashMapMerger::new();
+            let span =
+                tracing::trace_span!(target: "indexing::documents::extract", "merger_building");
+            let _entered = span.enter();
+
+            let readers: Vec<_> = context_pool
+                .into_items()
+                .par_bridge()
+                .map(|(_rtxn, _tokenizer, _fields_ids_map, cached_sorter)| {
+                    cached_sorter.into_sorter()
+                })
+                .collect();
+
+            builder.extend(readers);
+
+            Ok(builder)
+        }
+    }
+}
--- a/milli/src/update/new/extract/faceted/facet_document.rs
+++ b/milli/src/update/new/extract/faceted/facet_document.rs
@@ -0,0 +1,52 @@
+use serde_json::Value;
+
+use crate::update::new::extract::perm_json_p;
+use crate::update::new::KvReaderFieldId;
+use crate::{FieldId, GlobalFieldsIdsMap, InternalError, Result, UserError};
+
+pub fn extract_document_facets(
+    attributes_to_extract: &[&str],
+    obkv: &KvReaderFieldId,
+    field_id_map: &mut GlobalFieldsIdsMap,
+    facet_fn: &mut impl FnMut(FieldId, &Value) -> Result<()>,
+) -> Result<()> {
+    let mut field_name = String::new();
+    for (field_id, field_bytes) in obkv {
+        let Some(field_name) = field_id_map.name(field_id).map(|s| {
+            field_name.clear();
+            field_name.push_str(s);
+            &field_name
+        }) else {
+            unreachable!("field id not found in field id map");
+        };
+
+        let mut tokenize_field = |name: &str, value: &Value| match field_id_map.id_or_insert(name) {
+            Some(field_id) => facet_fn(field_id, value),
+            None => Err(UserError::AttributeLimitReached.into()),
+        };
+
+        // if the current field is searchable or contains a searchable attribute
+        if perm_json_p::select_field(field_name, Some(attributes_to_extract), &[]) {
+            // parse json.
+            match serde_json::from_slice(field_bytes).map_err(InternalError::SerdeJson)? {
+                Value::Object(object) => perm_json_p::seek_leaf_values_in_object(
+                    &object,
+                    Some(attributes_to_extract),
+                    &[], // skip no attributes
+                    field_name,
+                    &mut tokenize_field,
+                )?,
+                Value::Array(array) => perm_json_p::seek_leaf_values_in_array(
+                    &array,
+                    Some(attributes_to_extract),
+                    &[], // skip no attributes
+                    field_name,
+                    &mut tokenize_field,
+                )?,
+                value => tokenize_field(field_name, &value)?,
+            }
+        }
+    }
+
+    Ok(())
+}
--- a/milli/src/update/new/extract/faceted/mod.rs
+++ b/milli/src/update/new/extract/faceted/mod.rs
@@ -0,0 +1,26 @@
+mod extract_facets;
+mod facet_document;
+
+pub use extract_facets::FacetedDocidsExtractor;
+
+#[repr(u8)]
+pub enum FacetKind {
+    Number = 0,
+    String = 1,
+    Null = 2,
+    Empty = 3,
+    Exists,
+}
+
+impl From<u8> for FacetKind {
+    fn from(value: u8) -> Self {
+        match value {
+            0 => Self::Number,
+            1 => Self::String,
+            2 => Self::Null,
+            3 => Self::Empty,
+            4 => Self::Exists,
+            _ => unreachable!(),
+        }
+    }
+}
--- a/milli/src/update/new/extract/lru.rs
+++ b/milli/src/update/new/extract/lru.rs
@@ -0,0 +1,234 @@
+use std::borrow::Borrow;
+use std::hash::{BuildHasher, Hash};
+use std::iter::repeat_with;
+use std::mem;
+use std::num::NonZeroUsize;
+
+use hashbrown::hash_map::{DefaultHashBuilder, Entry};
+use hashbrown::HashMap;
+
+#[derive(Debug)]
+pub struct Lru<K, V, S = DefaultHashBuilder> {
+    lookup: HashMap<K, usize, S>,
+    storage: FixedSizeList<LruNode<K, V>>,
+}
+
+impl<K: Eq + Hash, V> Lru<K, V> {
+    /// Creates a new LRU cache that holds at most `capacity` elements.
+    pub fn new(capacity: NonZeroUsize) -> Self {
+        Self { lookup: HashMap::new(), storage: FixedSizeList::new(capacity.get()) }
+    }
+}
+
+impl<K: Eq + Hash, V, S: BuildHasher> Lru<K, V, S> {
+    /// Creates a new LRU cache that holds at most `capacity` elements
+    /// and uses the provided hash builder to hash keys.
+    pub fn with_hasher(capacity: NonZeroUsize, hash_builder: S) -> Lru<K, V, S> {
+        Self {
+            lookup: HashMap::with_hasher(hash_builder),
+            storage: FixedSizeList::new(capacity.get()),
+        }
+    }
+}
+
+impl<K: Eq + Hash, V, S: BuildHasher> Lru<K, V, S> {
+    /// Returns a mutable reference to the value of the key in the cache or `None` if it is not present in the cache.
+    ///
+    /// Moves the key to the head of the LRU list if it exists.
+    pub fn get_mut<Q>(&mut self, key: &Q) -> Option<&mut V>
+    where
+        K: Borrow<Q>,
+        Q: Hash + Eq + ?Sized,
+    {
+        let idx = *self.lookup.get(key)?;
+        self.storage.move_front(idx).map(|node| &mut node.value)
+    }
+}
+
+impl<K: Clone + Eq + Hash, V, S: BuildHasher> Lru<K, V, S> {
+    pub fn push(&mut self, key: K, value: V) -> Option<(K, V)> {
+        match self.lookup.entry(key) {
+            Entry::Occupied(occ) => {
+                // It's fine to unwrap here because:
+                // * the entry already exists
+                let node = self.storage.move_front(*occ.get()).unwrap();
+                let old_value = mem::replace(&mut node.value, value);
+                let old_key = occ.replace_key();
+                Some((old_key, old_value))
+            }
+            Entry::Vacant(vac) => {
+                let key = vac.key().clone();
+                if self.storage.is_full() {
+                    // It's fine to unwrap here because:
+                    // * the cache capacity is non zero
+                    // * the cache is full
+                    let idx = self.storage.back_idx();
+                    let node = self.storage.move_front(idx).unwrap();
+                    let LruNode { key, value } = mem::replace(node, LruNode { key, value });
+                    vac.insert(idx);
+                    self.lookup.remove(&key);
+                    Some((key, value))
+                } else {
+                    // It's fine to unwrap here because:
+                    // * the cache capacity is non zero
+                    // * the cache is not full
+                    let (idx, _) = self.storage.push_front(LruNode { key, value }).unwrap();
+                    vac.insert(idx);
+                    None
+                }
+            }
+        }
+    }
+}
+
+impl<K, V, S> IntoIterator for Lru<K, V, S> {
+    type Item = (K, V);
+    type IntoIter = IntoIter<K, V>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        IntoIter { lookup_iter: self.lookup.into_iter(), nodes: self.storage.nodes }
+    }
+}
+
+pub struct IntoIter<K, V> {
+    lookup_iter: hashbrown::hash_map::IntoIter<K, usize>,
+    nodes: Box<[Option<FixedSizeListNode<LruNode<K, V>>>]>,
+}
+
+impl<K, V> Iterator for IntoIter<K, V> {
+    type Item = (K, V);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let (_key, idx) = self.lookup_iter.next()?;
+        let LruNode { key, value } = self.nodes.get_mut(idx)?.take()?.data;
+        Some((key, value))
+    }
+}
+
+#[derive(Debug)]
+struct LruNode<K, V> {
+    key: K,
+    value: V,
+}
+
+#[derive(Debug)]
+struct FixedSizeListNode<T> {
+    prev: usize,
+    next: usize,
+    data: T,
+}
+
+#[derive(Debug)]
+struct FixedSizeList<T> {
+    nodes: Box<[Option<FixedSizeListNode<T>>]>,
+    /// Also corresponds to the first `None` in the nodes.
+    length: usize,
+    // TODO Also, we probably do not need one of the front and back cursors.
+    front: usize,
+    back: usize,
+}
+
+impl<T> FixedSizeList<T> {
+    fn new(capacity: usize) -> Self {
+        Self {
+            nodes: repeat_with(|| None).take(capacity).collect::<Vec<_>>().into_boxed_slice(),
+            length: 0,
+            front: usize::MAX,
+            back: usize::MAX,
+        }
+    }
+
+    #[inline]
+    fn capacity(&self) -> usize {
+        self.nodes.len()
+    }
+
+    #[inline]
+    fn len(&self) -> usize {
+        self.length
+    }
+
+    #[inline]
+    fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+
+    #[inline]
+    fn is_full(&self) -> bool {
+        self.len() == self.capacity()
+    }
+
+    #[inline]
+    fn back_idx(&self) -> usize {
+        self.back
+    }
+
+    #[inline]
+    fn next(&mut self) -> Option<usize> {
+        if self.is_full() {
+            None
+        } else {
+            let current_free = self.length;
+            self.length += 1;
+            Some(current_free)
+        }
+    }
+
+    #[inline]
+    fn node_mut(&mut self, idx: usize) -> Option<&mut FixedSizeListNode<T>> {
+        self.nodes.get_mut(idx).and_then(|node| node.as_mut())
+    }
+
+    #[inline]
+    fn node_ref(&self, idx: usize) -> Option<&FixedSizeListNode<T>> {
+        self.nodes.get(idx).and_then(|node| node.as_ref())
+    }
+
+    #[inline]
+    fn move_front(&mut self, idx: usize) -> Option<&mut T> {
+        let node = self.nodes.get_mut(idx)?.take()?;
+        if let Some(prev) = self.node_mut(node.prev) {
+            prev.next = node.next;
+        } else {
+            self.front = node.next;
+        }
+        if let Some(next) = self.node_mut(node.next) {
+            next.prev = node.prev;
+        } else {
+            self.back = node.prev;
+        }
+
+        if let Some(front) = self.node_mut(self.front) {
+            front.prev = idx;
+        }
+        if self.node_ref(self.back).is_none() {
+            self.back = idx;
+        }
+
+        let node = self.nodes.get_mut(idx).unwrap().insert(FixedSizeListNode {
+            prev: usize::MAX,
+            next: self.front,
+            data: node.data,
+        });
+        self.front = idx;
+        Some(&mut node.data)
+    }
+
+    #[inline]
+    fn push_front(&mut self, data: T) -> Option<(usize, &mut T)> {
+        let idx = self.next()?;
+        if let Some(front) = self.node_mut(self.front) {
+            front.prev = idx;
+        }
+        if self.node_ref(self.back).is_none() {
+            self.back = idx;
+        }
+        let node = self.nodes.get_mut(idx).unwrap().insert(FixedSizeListNode {
+            prev: usize::MAX,
+            next: self.front,
+            data,
+        });
+        self.front = idx;
+        Some((idx, &mut node.data))
+    }
+}
--- a/milli/src/update/new/extract/mod.rs
+++ b/milli/src/update/new/extract/mod.rs
@@ -0,0 +1,218 @@
+mod cache;
+mod faceted;
+mod lru;
+mod searchable;
+
+use std::collections::HashMap;
+use std::mem;
+
+pub use faceted::*;
+use grenad::MergeFunction;
+use rayon::iter::{IntoParallelIterator, ParallelIterator as _};
+use rayon::slice::ParallelSliceMut as _;
+pub use searchable::*;
+use smallvec::SmallVec;
+
+use super::DocumentChange;
+use crate::update::{GrenadParameters, MergeDeladdCboRoaringBitmaps};
+use crate::{GlobalFieldsIdsMap, Index, Result};
+
+pub trait DocidsExtractor {
+    fn run_extraction(
+        index: &Index,
+        fields_ids_map: &GlobalFieldsIdsMap,
+        indexer: GrenadParameters,
+        document_changes: impl IntoParallelIterator<Item = Result<DocumentChange>>,
+    ) -> Result<HashMapMerger>;
+}
+
+pub struct HashMapMerger {
+    maps: Vec<HashMap<SmallVec<[u8; cache::KEY_SIZE]>, cache::DelAddRoaringBitmap>>,
+}
+
+impl HashMapMerger {
+    pub fn new() -> HashMapMerger {
+        HashMapMerger { maps: Vec::new() }
+    }
+
+    pub fn extend<I>(&mut self, iter: I)
+    where
+        I: IntoIterator<
+            Item = HashMap<SmallVec<[u8; cache::KEY_SIZE]>, cache::DelAddRoaringBitmap>,
+        >,
+    {
+        self.maps.extend(iter);
+    }
+}
+
+impl IntoIterator for HashMapMerger {
+    type Item = (SmallVec<[u8; 12]>, cache::DelAddRoaringBitmap);
+    type IntoIter = IntoIter;
+
+    fn into_iter(self) -> Self::IntoIter {
+        let mut entries = {
+            let span = tracing::trace_span!(target: "indexing::documents::merge", "into_par_iter");
+            let _entered = span.enter();
+            let entries: Vec<_> =
+                self.maps.into_par_iter().flat_map(|m| m.into_par_iter()).collect();
+            eprintln!("There are {} entries in the HashMapMerger", entries.len());
+            entries
+        };
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::merge", "par_sort_unstable_by");
+            let _entered = span.enter();
+            entries.par_sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
+            IntoIter {
+                sorted_entries: entries.into_iter(),
+                current_key: None,
+                current_deladd: cache::DelAddRoaringBitmap::default(),
+            }
+        }
+    }
+}
+
+pub struct IntoIter {
+    sorted_entries: std::vec::IntoIter<(SmallVec<[u8; 12]>, cache::DelAddRoaringBitmap)>,
+    current_key: Option<SmallVec<[u8; 12]>>,
+    current_deladd: cache::DelAddRoaringBitmap,
+}
+
+impl Iterator for IntoIter {
+    type Item = (SmallVec<[u8; 12]>, cache::DelAddRoaringBitmap);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        loop {
+            match self.sorted_entries.next() {
+                Some((k, deladd)) => {
+                    if self.current_key.as_deref() == Some(k.as_slice()) {
+                        self.current_deladd.merge_with(deladd);
+                    } else {
+                        let previous_key = self.current_key.replace(k);
+                        let previous_deladd = mem::replace(&mut self.current_deladd, deladd);
+                        if let Some(previous_key) = previous_key {
+                            return Some((previous_key, previous_deladd));
+                        }
+                    }
+                }
+                None => {
+                    let current_deladd = mem::take(&mut self.current_deladd);
+                    return self.current_key.take().map(|ck| (ck, current_deladd));
+                }
+            }
+        }
+    }
+}
+
+/// TODO move in permissive json pointer
+pub mod perm_json_p {
+    use serde_json::{Map, Value};
+
+    use crate::Result;
+    const SPLIT_SYMBOL: char = '.';
+
+    /// Returns `true` if the `selector` match the `key`.
+    ///
+    /// ```text
+    /// Example:
+    /// `animaux`           match `animaux`
+    /// `animaux.chien`     match `animaux`
+    /// `animaux.chien`     match `animaux`
+    /// `animaux.chien.nom` match `animaux`
+    /// `animaux.chien.nom` match `animaux.chien`
+    /// -----------------------------------------
+    /// `animaux`    doesn't match `animaux.chien`
+    /// `animaux.`   doesn't match `animaux`
+    /// `animaux.ch` doesn't match `animaux.chien`
+    /// `animau`     doesn't match `animaux`
+    /// ```
+    pub fn contained_in(selector: &str, key: &str) -> bool {
+        selector.starts_with(key)
+            && selector[key.len()..].chars().next().map(|c| c == SPLIT_SYMBOL).unwrap_or(true)
+    }
+
+    pub fn seek_leaf_values_in_object(
+        value: &Map<String, Value>,
+        selectors: Option<&[&str]>,
+        skip_selectors: &[&str],
+        base_key: &str,
+        seeker: &mut impl FnMut(&str, &Value) -> Result<()>,
+    ) -> Result<()> {
+        if value.is_empty() {
+            seeker(&base_key, &Value::Object(Map::with_capacity(0)))?;
+        }
+
+        for (key, value) in value.iter() {
+            let base_key = if base_key.is_empty() {
+                key.to_string()
+            } else {
+                format!("{}{}{}", base_key, SPLIT_SYMBOL, key)
+            };
+
+            // here if the user only specified `doggo` we need to iterate in all the fields of `doggo`
+            // so we check the contained_in on both side
+            let should_continue = select_field(&base_key, selectors, skip_selectors);
+            if should_continue {
+                match value {
+                    Value::Object(object) => seek_leaf_values_in_object(
+                        object,
+                        selectors,
+                        skip_selectors,
+                        &base_key,
+                        seeker,
+                    ),
+                    Value::Array(array) => seek_leaf_values_in_array(
+                        array,
+                        selectors,
+                        skip_selectors,
+                        &base_key,
+                        seeker,
+                    ),
+                    value => seeker(&base_key, value),
+                }?;
+            }
+        }
+
+        Ok(())
+    }
+
+    pub fn seek_leaf_values_in_array(
+        values: &[Value],
+        selectors: Option<&[&str]>,
+        skip_selectors: &[&str],
+        base_key: &str,
+        seeker: &mut impl FnMut(&str, &Value) -> Result<()>,
+    ) -> Result<()> {
+        if values.is_empty() {
+            seeker(&base_key, &Value::Array(vec![]))?;
+        }
+
+        for value in values {
+            match value {
+                Value::Object(object) => {
+                    seek_leaf_values_in_object(object, selectors, skip_selectors, base_key, seeker)
+                }
+                Value::Array(array) => {
+                    seek_leaf_values_in_array(array, selectors, skip_selectors, base_key, seeker)
+                }
+                value => seeker(base_key, value),
+            }?;
+        }
+
+        Ok(())
+    }
+
+    pub fn select_field(
+        field_name: &str,
+        selectors: Option<&[&str]>,
+        skip_selectors: &[&str],
+    ) -> bool {
+        selectors.map_or(true, |selectors| {
+            selectors.iter().any(|selector| {
+                contained_in(selector, &field_name) || contained_in(&field_name, selector)
+            })
+        }) && !skip_selectors.iter().any(|skip_selector| {
+            contained_in(skip_selector, &field_name) || contained_in(&field_name, skip_selector)
+        })
+    }
+}
--- a/milli/src/update/new/extract/searchable/extract_fid_word_count_docids.rs
+++ b/milli/src/update/new/extract/searchable/extract_fid_word_count_docids.rs
@@ -0,0 +1,124 @@
+use std::collections::HashMap;
+
+use heed::RoTxn;
+
+use super::tokenize_document::DocumentTokenizer;
+use super::SearchableExtractor;
+use crate::update::new::extract::cache::CboCachedSorter;
+use crate::update::new::DocumentChange;
+use crate::update::MergeDeladdCboRoaringBitmaps;
+use crate::{FieldId, GlobalFieldsIdsMap, Index, Result};
+
+const MAX_COUNTED_WORDS: usize = 30;
+
+pub struct FidWordCountDocidsExtractor;
+impl SearchableExtractor for FidWordCountDocidsExtractor {
+    fn attributes_to_extract<'a>(
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>> {
+        index.user_defined_searchable_fields(rtxn).map_err(Into::into)
+    }
+
+    fn attributes_to_skip<'a>(_rtxn: &'a RoTxn, _index: &'a Index) -> Result<Vec<&'a str>> {
+        Ok(vec![])
+    }
+
+    // This method is reimplemented to count the number of words in the document in each field
+    // and to store the docids of the documents that have a number of words in a given field equal to or under than MAX_COUNTED_WORDS.
+    fn extract_document_change(
+        rtxn: &RoTxn,
+        index: &Index,
+        document_tokenizer: &DocumentTokenizer,
+        fields_ids_map: &mut GlobalFieldsIdsMap,
+        cached_sorter: &mut CboCachedSorter,
+        document_change: DocumentChange,
+    ) -> Result<()> {
+        let mut key_buffer = Vec::new();
+        match document_change {
+            DocumentChange::Deletion(inner) => {
+                let mut fid_word_count = HashMap::new();
+                let mut token_fn = |_fname: &str, fid: FieldId, _pos: u16, _word: &str| {
+                    fid_word_count.entry(fid).and_modify(|count| *count += 1).or_insert(1);
+                    Ok(())
+                };
+                document_tokenizer.tokenize_document(
+                    inner.current(rtxn, index)?.unwrap(),
+                    fields_ids_map,
+                    &mut token_fn,
+                )?;
+
+                // The docids of the documents that have a number of words in a given field equal to or under than MAX_COUNTED_WORDS are deleted.
+                for (fid, count) in fid_word_count.iter() {
+                    if *count <= MAX_COUNTED_WORDS {
+                        let key = build_key(*fid, *count as u8, &mut key_buffer);
+                        cached_sorter.insert_del_u32(key, inner.docid());
+                    }
+                }
+            }
+            DocumentChange::Update(inner) => {
+                let mut fid_word_count = HashMap::new();
+                let mut token_fn = |_fname: &str, fid: FieldId, _pos: u16, _word: &str| {
+                    fid_word_count
+                        .entry(fid)
+                        .and_modify(|(current_count, _new_count)| *current_count += 1)
+                        .or_insert((1, 0));
+                    Ok(())
+                };
+                document_tokenizer.tokenize_document(
+                    inner.current(rtxn, index)?.unwrap(),
+                    fields_ids_map,
+                    &mut token_fn,
+                )?;
+
+                let mut token_fn = |_fname: &str, fid: FieldId, _pos: u16, _word: &str| {
+                    fid_word_count
+                        .entry(fid)
+                        .and_modify(|(_current_count, new_count)| *new_count += 1)
+                        .or_insert((0, 1));
+                    Ok(())
+                };
+                document_tokenizer.tokenize_document(inner.new(), fields_ids_map, &mut token_fn)?;
+
+                // Only the fields that have a change in the number of words are updated.
+                for (fid, (current_count, new_count)) in fid_word_count.iter() {
+                    if *current_count != *new_count {
+                        if *current_count <= MAX_COUNTED_WORDS {
+                            let key = build_key(*fid, *current_count as u8, &mut key_buffer);
+                            cached_sorter.insert_del_u32(key, inner.docid());
+                        }
+                        if *new_count <= MAX_COUNTED_WORDS {
+                            let key = build_key(*fid, *new_count as u8, &mut key_buffer);
+                            cached_sorter.insert_add_u32(key, inner.docid());
+                        }
+                    }
+                }
+            }
+            DocumentChange::Insertion(inner) => {
+                let mut fid_word_count = HashMap::new();
+                let mut token_fn = |_fname: &str, fid: FieldId, _pos: u16, _word: &str| {
+                    fid_word_count.entry(fid).and_modify(|count| *count += 1).or_insert(1);
+                    Ok(())
+                };
+                document_tokenizer.tokenize_document(inner.new(), fields_ids_map, &mut token_fn)?;
+
+                // The docids of the documents that have a number of words in a given field equal to or under than MAX_COUNTED_WORDS are stored.
+                for (fid, count) in fid_word_count.iter() {
+                    if *count <= MAX_COUNTED_WORDS {
+                        let key = build_key(*fid, *count as u8, &mut key_buffer);
+                        cached_sorter.insert_add_u32(key, inner.docid());
+                    }
+                }
+            }
+        }
+
+        Ok(())
+    }
+}
+
+fn build_key(fid: FieldId, count: u8, key_buffer: &mut Vec<u8>) -> &[u8] {
+    key_buffer.clear();
+    key_buffer.extend_from_slice(&fid.to_be_bytes());
+    key_buffer.push(count);
+    key_buffer.as_slice()
+}
--- a/milli/src/update/new/extract/searchable/extract_word_docids.rs
+++ b/milli/src/update/new/extract/searchable/extract_word_docids.rs
@@ -0,0 +1,594 @@
+use std::borrow::Cow;
+use std::collections::HashMap;
+use std::fs::File;
+use std::num::NonZero;
+
+use grenad::{Merger, MergerBuilder};
+use heed::RoTxn;
+use rayon::iter::{IntoParallelIterator, ParallelIterator};
+
+use super::tokenize_document::{tokenizer_builder, DocumentTokenizer};
+use super::SearchableExtractor;
+use crate::update::new::extract::cache::CboCachedSorter;
+use crate::update::new::extract::perm_json_p::contained_in;
+use crate::update::new::extract::HashMapMerger;
+use crate::update::new::{DocumentChange, ItemsPool};
+use crate::update::{create_sorter, GrenadParameters, MergeDeladdCboRoaringBitmaps};
+use crate::{
+    bucketed_position, DocumentId, FieldId, GlobalFieldsIdsMap, Index, Result,
+    MAX_POSITION_PER_ATTRIBUTE,
+};
+
+const MAX_COUNTED_WORDS: usize = 30;
+
+trait ProtoWordDocidsExtractor {
+    fn build_key(field_id: FieldId, position: u16, word: &str) -> Cow<'_, [u8]>;
+    fn attributes_to_extract<'a>(
+        _rtxn: &'a RoTxn,
+        _index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>>;
+
+    fn attributes_to_skip<'a>(rtxn: &'a RoTxn, index: &'a Index) -> Result<Vec<&'a str>>;
+}
+
+impl<T> SearchableExtractor for T
+where
+    T: ProtoWordDocidsExtractor,
+{
+    fn extract_document_change(
+        rtxn: &RoTxn,
+        index: &Index,
+        document_tokenizer: &DocumentTokenizer,
+        fields_ids_map: &mut GlobalFieldsIdsMap,
+        cached_sorter: &mut CboCachedSorter,
+        document_change: DocumentChange,
+    ) -> Result<()> {
+        match document_change {
+            DocumentChange::Deletion(inner) => {
+                let mut token_fn = |_fname: &str, fid, pos, word: &str| {
+                    let key = Self::build_key(fid, pos, word);
+                    Ok(cached_sorter.insert_del_u32(&key, inner.docid()))
+                };
+                document_tokenizer.tokenize_document(
+                    inner.current(rtxn, index)?.unwrap(),
+                    fields_ids_map,
+                    &mut token_fn,
+                )?;
+            }
+            DocumentChange::Update(inner) => {
+                let mut token_fn = |_fname: &str, fid, pos, word: &str| {
+                    let key = Self::build_key(fid, pos, word);
+                    Ok(cached_sorter.insert_del_u32(&key, inner.docid()))
+                };
+                document_tokenizer.tokenize_document(
+                    inner.current(rtxn, index)?.unwrap(),
+                    fields_ids_map,
+                    &mut token_fn,
+                )?;
+
+                let mut token_fn = |_fname: &str, fid, pos, word: &str| {
+                    let key = Self::build_key(fid, pos, word);
+                    Ok(cached_sorter.insert_add_u32(&key, inner.docid()))
+                };
+                document_tokenizer.tokenize_document(inner.new(), fields_ids_map, &mut token_fn)?;
+            }
+            DocumentChange::Insertion(inner) => {
+                let mut token_fn = |_fname: &str, fid, pos, word: &str| {
+                    let key = Self::build_key(fid, pos, word);
+                    Ok(cached_sorter.insert_add_u32(&key, inner.docid()))
+                };
+                document_tokenizer.tokenize_document(inner.new(), fields_ids_map, &mut token_fn)?;
+            }
+        }
+
+        Ok(())
+    }
+
+    fn attributes_to_extract<'a>(
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>> {
+        Self::attributes_to_extract(rtxn, index)
+    }
+
+    fn attributes_to_skip<'a>(rtxn: &'a RoTxn, index: &'a Index) -> Result<Vec<&'a str>> {
+        Self::attributes_to_skip(rtxn, index)
+    }
+}
+
+pub struct WordDocidsExtractor;
+impl ProtoWordDocidsExtractor for WordDocidsExtractor {
+    fn attributes_to_extract<'a>(
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>> {
+        index.user_defined_searchable_fields(rtxn).map_err(Into::into)
+    }
+
+    fn attributes_to_skip<'a>(rtxn: &'a RoTxn, index: &'a Index) -> Result<Vec<&'a str>> {
+        // exact attributes must be skipped and stored in a separate DB, see `ExactWordDocidsExtractor`.
+        index.exact_attributes(rtxn).map_err(Into::into)
+    }
+
+    /// TODO write in an external Vec buffer
+    fn build_key(_field_id: FieldId, _position: u16, word: &str) -> Cow<[u8]> {
+        Cow::Borrowed(word.as_bytes())
+    }
+}
+
+pub struct ExactWordDocidsExtractor;
+impl ProtoWordDocidsExtractor for ExactWordDocidsExtractor {
+    fn attributes_to_extract<'a>(
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>> {
+        let exact_attributes = index.exact_attributes(rtxn)?;
+        // If there are no user-defined searchable fields, we return all exact attributes.
+        // Otherwise, we return the intersection of exact attributes and user-defined searchable fields.
+        if let Some(searchable_attributes) = index.user_defined_searchable_fields(rtxn)? {
+            let attributes = exact_attributes
+                .into_iter()
+                .filter(|attr| searchable_attributes.contains(attr))
+                .collect();
+            Ok(Some(attributes))
+        } else {
+            Ok(Some(exact_attributes))
+        }
+    }
+
+    fn attributes_to_skip<'a>(_rtxn: &'a RoTxn, _index: &'a Index) -> Result<Vec<&'a str>> {
+        Ok(vec![])
+    }
+
+    fn build_key(_field_id: FieldId, _position: u16, word: &str) -> Cow<[u8]> {
+        Cow::Borrowed(word.as_bytes())
+    }
+}
+
+pub struct WordFidDocidsExtractor;
+impl ProtoWordDocidsExtractor for WordFidDocidsExtractor {
+    fn attributes_to_extract<'a>(
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>> {
+        index.user_defined_searchable_fields(rtxn).map_err(Into::into)
+    }
+
+    fn attributes_to_skip<'a>(_rtxn: &'a RoTxn, _index: &'a Index) -> Result<Vec<&'a str>> {
+        Ok(vec![])
+    }
+
+    fn build_key(field_id: FieldId, _position: u16, word: &str) -> Cow<[u8]> {
+        let mut key = Vec::new();
+        key.extend_from_slice(word.as_bytes());
+        key.push(0);
+        key.extend_from_slice(&field_id.to_be_bytes());
+        Cow::Owned(key)
+    }
+}
+
+pub struct WordPositionDocidsExtractor;
+impl ProtoWordDocidsExtractor for WordPositionDocidsExtractor {
+    fn attributes_to_extract<'a>(
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>> {
+        index.user_defined_searchable_fields(rtxn).map_err(Into::into)
+    }
+
+    fn attributes_to_skip<'a>(_rtxn: &'a RoTxn, _index: &'a Index) -> Result<Vec<&'a str>> {
+        Ok(vec![])
+    }
+
+    fn build_key(_field_id: FieldId, position: u16, word: &str) -> Cow<[u8]> {
+        // position must be bucketed to reduce the number of keys in the DB.
+        let position = bucketed_position(position);
+        let mut key = Vec::new();
+        key.extend_from_slice(word.as_bytes());
+        key.push(0);
+        key.extend_from_slice(&position.to_be_bytes());
+        Cow::Owned(key)
+    }
+}
+
+// V2
+
+struct WordDocidsCachedSorters {
+    word_fid_docids: CboCachedSorter,
+    word_docids: CboCachedSorter,
+    exact_word_docids: CboCachedSorter,
+    word_position_docids: CboCachedSorter,
+    fid_word_count_docids: CboCachedSorter,
+    fid_word_count: HashMap<FieldId, (usize, usize)>,
+    current_docid: Option<DocumentId>,
+}
+
+impl WordDocidsCachedSorters {
+    pub fn new(
+        indexer: GrenadParameters,
+        max_memory: Option<usize>,
+        capacity: NonZero<usize>,
+    ) -> Self {
+        let max_memory = max_memory.map(|max_memory| max_memory / 4);
+
+        let word_fid_docids = CboCachedSorter::new();
+        let word_docids = CboCachedSorter::new();
+        let exact_word_docids = CboCachedSorter::new();
+        let word_position_docids = CboCachedSorter::new();
+        let fid_word_count_docids = CboCachedSorter::new();
+
+        Self {
+            word_fid_docids,
+            word_docids,
+            exact_word_docids,
+            word_position_docids,
+            fid_word_count_docids,
+            fid_word_count: HashMap::new(),
+            current_docid: None,
+        }
+    }
+
+    fn insert_add_u32(
+        &mut self,
+        field_id: FieldId,
+        position: u16,
+        word: &str,
+        exact: bool,
+        docid: u32,
+        buffer: &mut Vec<u8>,
+    ) -> Result<()> {
+        let key = word.as_bytes();
+        if exact {
+            self.exact_word_docids.insert_add_u32(key, docid);
+        } else {
+            self.word_docids.insert_add_u32(key, docid);
+        }
+
+        buffer.clear();
+        buffer.extend_from_slice(word.as_bytes());
+        buffer.push(0);
+        buffer.extend_from_slice(&position.to_be_bytes());
+        self.word_fid_docids.insert_add_u32(buffer, docid);
+
+        buffer.clear();
+        buffer.extend_from_slice(word.as_bytes());
+        buffer.push(0);
+        buffer.extend_from_slice(&field_id.to_be_bytes());
+        self.word_position_docids.insert_add_u32(buffer, docid);
+
+        if self.current_docid.map_or(false, |id| docid != id) {
+            self.flush_fid_word_count(buffer)?;
+        }
+
+        self.fid_word_count
+            .entry(field_id)
+            .and_modify(|(_current_count, new_count)| *new_count += 1)
+            .or_insert((0, 1));
+        self.current_docid = Some(docid);
+
+        Ok(())
+    }
+
+    fn insert_del_u32(
+        &mut self,
+        field_id: FieldId,
+        position: u16,
+        word: &str,
+        exact: bool,
+        docid: u32,
+        buffer: &mut Vec<u8>,
+    ) -> Result<()> {
+        let key = word.as_bytes();
+        if exact {
+            self.exact_word_docids.insert_del_u32(key, docid);
+        } else {
+            self.word_docids.insert_del_u32(key, docid);
+        }
+
+        buffer.clear();
+        buffer.extend_from_slice(word.as_bytes());
+        buffer.push(0);
+        buffer.extend_from_slice(&position.to_be_bytes());
+        self.word_fid_docids.insert_del_u32(buffer, docid);
+
+        buffer.clear();
+        buffer.extend_from_slice(word.as_bytes());
+        buffer.push(0);
+        buffer.extend_from_slice(&field_id.to_be_bytes());
+        self.word_position_docids.insert_del_u32(buffer, docid);
+
+        if self.current_docid.map_or(false, |id| docid != id) {
+            self.flush_fid_word_count(buffer)?;
+        }
+
+        self.fid_word_count
+            .entry(field_id)
+            .and_modify(|(current_count, _new_count)| *current_count += 1)
+            .or_insert((1, 0));
+        self.current_docid = Some(docid);
+
+        Ok(())
+    }
+
+    fn flush_fid_word_count(&mut self, buffer: &mut Vec<u8>) -> Result<()> {
+        for (fid, (current_count, new_count)) in self.fid_word_count.drain() {
+            if current_count != new_count {
+                if current_count <= MAX_COUNTED_WORDS {
+                    buffer.clear();
+                    buffer.extend_from_slice(&fid.to_be_bytes());
+                    buffer.push(current_count as u8);
+                    self.fid_word_count_docids.insert_del_u32(buffer, self.current_docid.unwrap());
+                }
+                if new_count <= MAX_COUNTED_WORDS {
+                    buffer.clear();
+                    buffer.extend_from_slice(&fid.to_be_bytes());
+                    buffer.push(new_count as u8);
+                    self.fid_word_count_docids.insert_add_u32(buffer, self.current_docid.unwrap());
+                }
+            }
+        }
+
+        Ok(())
+    }
+}
+
+struct WordDocidsMergerBuilders {
+    word_fid_docids: HashMapMerger,
+    word_docids: HashMapMerger,
+    exact_word_docids: HashMapMerger,
+    word_position_docids: HashMapMerger,
+    fid_word_count_docids: HashMapMerger,
+}
+
+pub struct WordDocidsMergers {
+    pub word_fid_docids: HashMapMerger,
+    pub word_docids: HashMapMerger,
+    pub exact_word_docids: HashMapMerger,
+    pub word_position_docids: HashMapMerger,
+    pub fid_word_count_docids: HashMapMerger,
+}
+
+impl WordDocidsMergerBuilders {
+    fn new() -> Self {
+        Self {
+            word_fid_docids: HashMapMerger::new(),
+            word_docids: HashMapMerger::new(),
+            exact_word_docids: HashMapMerger::new(),
+            word_position_docids: HashMapMerger::new(),
+            fid_word_count_docids: HashMapMerger::new(),
+        }
+    }
+
+    fn add_sorters(&mut self, other: WordDocidsCachedSorters) -> Result<()> {
+        let WordDocidsCachedSorters {
+            word_fid_docids,
+            word_docids,
+            exact_word_docids,
+            word_position_docids,
+            fid_word_count_docids,
+            fid_word_count: _,
+            current_docid: _,
+        } = other;
+
+        let mut word_fid_docids_readers = HashMap::new();
+        let mut word_docids_readers = HashMap::new();
+        let mut exact_word_docids_readers = HashMap::new();
+        let mut word_position_docids_readers = HashMap::new();
+        let mut fid_word_count_docids_readers = HashMap::new();
+        rayon::scope(|s| {
+            s.spawn(|_| {
+                word_fid_docids_readers = word_fid_docids.into_sorter();
+            });
+            s.spawn(|_| {
+                word_docids_readers = word_docids.into_sorter();
+            });
+            s.spawn(|_| {
+                exact_word_docids_readers = exact_word_docids.into_sorter();
+            });
+            s.spawn(|_| {
+                word_position_docids_readers = word_position_docids.into_sorter();
+            });
+            s.spawn(|_| {
+                fid_word_count_docids_readers = fid_word_count_docids.into_sorter();
+            });
+        });
+        self.word_fid_docids.extend([word_fid_docids_readers]);
+        self.word_docids.extend([word_docids_readers]);
+        self.exact_word_docids.extend([exact_word_docids_readers]);
+        self.word_position_docids.extend([word_position_docids_readers]);
+        self.fid_word_count_docids.extend([fid_word_count_docids_readers]);
+
+        Ok(())
+    }
+
+    fn build(self) -> WordDocidsMergers {
+        WordDocidsMergers {
+            word_fid_docids: self.word_fid_docids,
+            word_docids: self.word_docids,
+            exact_word_docids: self.exact_word_docids,
+            word_position_docids: self.word_position_docids,
+            fid_word_count_docids: self.fid_word_count_docids,
+        }
+    }
+}
+
+pub struct WordDocidsExtractors;
+
+impl WordDocidsExtractors {
+    pub fn run_extraction(
+        index: &Index,
+        fields_ids_map: &GlobalFieldsIdsMap,
+        indexer: GrenadParameters,
+        document_changes: impl IntoParallelIterator<Item = Result<DocumentChange>>,
+    ) -> Result<WordDocidsMergers> {
+        let max_memory = indexer.max_memory_by_thread();
+
+        let rtxn = index.read_txn()?;
+        let stop_words = index.stop_words(&rtxn)?;
+        let allowed_separators = index.allowed_separators(&rtxn)?;
+        let allowed_separators: Option<Vec<_>> =
+            allowed_separators.as_ref().map(|s| s.iter().map(String::as_str).collect());
+        let dictionary = index.dictionary(&rtxn)?;
+        let dictionary: Option<Vec<_>> =
+            dictionary.as_ref().map(|s| s.iter().map(String::as_str).collect());
+        let builder = tokenizer_builder(
+            stop_words.as_ref(),
+            allowed_separators.as_deref(),
+            dictionary.as_deref(),
+        );
+        let tokenizer = builder.into_tokenizer();
+
+        let attributes_to_extract = Self::attributes_to_extract(&rtxn, index)?;
+        let attributes_to_skip = Self::attributes_to_skip(&rtxn, index)?;
+        let localized_attributes_rules =
+            index.localized_attributes_rules(&rtxn)?.unwrap_or_default();
+
+        let document_tokenizer = DocumentTokenizer {
+            tokenizer: &tokenizer,
+            attribute_to_extract: attributes_to_extract.as_deref(),
+            attribute_to_skip: attributes_to_skip.as_slice(),
+            localized_attributes_rules: &localized_attributes_rules,
+            max_positions_per_attributes: MAX_POSITION_PER_ATTRIBUTE,
+        };
+
+        let context_pool = ItemsPool::new(|| {
+            Ok((
+                index.read_txn()?,
+                &document_tokenizer,
+                fields_ids_map.clone(),
+                WordDocidsCachedSorters::new(
+                    indexer,
+                    max_memory,
+                    // TODO use a better value
+                    200_000.try_into().unwrap(),
+                ),
+            ))
+        });
+
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::extract", "docids_extraction");
+            let _entered = span.enter();
+            document_changes.into_par_iter().try_for_each(|document_change| {
+                context_pool.with(|(rtxn, document_tokenizer, fields_ids_map, cached_sorter)| {
+                    Self::extract_document_change(
+                        &*rtxn,
+                        index,
+                        document_tokenizer,
+                        fields_ids_map,
+                        cached_sorter,
+                        document_change?,
+                    )
+                })
+            })?;
+        }
+
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::extract", "merger_building");
+            let _entered = span.enter();
+            let mut builder = WordDocidsMergerBuilders::new();
+            for (_rtxn, _tokenizer, _fields_ids_map, cache) in context_pool.into_items() {
+                builder.add_sorters(cache)?;
+            }
+
+            Ok(builder.build())
+        }
+    }
+
+    fn extract_document_change(
+        rtxn: &RoTxn,
+        index: &Index,
+        document_tokenizer: &DocumentTokenizer,
+        fields_ids_map: &mut GlobalFieldsIdsMap,
+        cached_sorter: &mut WordDocidsCachedSorters,
+        document_change: DocumentChange,
+    ) -> Result<()> {
+        let exact_attributes = index.exact_attributes(rtxn)?;
+        let is_exact_attribute =
+            |fname: &str| exact_attributes.iter().any(|attr| contained_in(fname, attr));
+        let mut buffer = Vec::new();
+        match document_change {
+            DocumentChange::Deletion(inner) => {
+                let mut token_fn = |fname: &str, fid, pos, word: &str| {
+                    cached_sorter
+                        .insert_del_u32(
+                            fid,
+                            pos,
+                            word,
+                            is_exact_attribute(fname),
+                            inner.docid(),
+                            &mut buffer,
+                        )
+                        .map_err(crate::Error::from)
+                };
+                document_tokenizer.tokenize_document(
+                    inner.current(rtxn, index)?.unwrap(),
+                    fields_ids_map,
+                    &mut token_fn,
+                )?;
+            }
+            DocumentChange::Update(inner) => {
+                let mut token_fn = |fname: &str, fid, pos, word: &str| {
+                    cached_sorter
+                        .insert_del_u32(
+                            fid,
+                            pos,
+                            word,
+                            is_exact_attribute(fname),
+                            inner.docid(),
+                            &mut buffer,
+                        )
+                        .map_err(crate::Error::from)
+                };
+                document_tokenizer.tokenize_document(
+                    inner.current(rtxn, index)?.unwrap(),
+                    fields_ids_map,
+                    &mut token_fn,
+                )?;
+
+                let mut token_fn = |fname: &str, fid, pos, word: &str| {
+                    cached_sorter
+                        .insert_add_u32(
+                            fid,
+                            pos,
+                            word,
+                            is_exact_attribute(fname),
+                            inner.docid(),
+                            &mut buffer,
+                        )
+                        .map_err(crate::Error::from)
+                };
+                document_tokenizer.tokenize_document(inner.new(), fields_ids_map, &mut token_fn)?;
+            }
+            DocumentChange::Insertion(inner) => {
+                let mut token_fn = |fname: &str, fid, pos, word: &str| {
+                    cached_sorter
+                        .insert_add_u32(
+                            fid,
+                            pos,
+                            word,
+                            is_exact_attribute(fname),
+                            inner.docid(),
+                            &mut buffer,
+                        )
+                        .map_err(crate::Error::from)
+                };
+                document_tokenizer.tokenize_document(inner.new(), fields_ids_map, &mut token_fn)?;
+            }
+        }
+
+        cached_sorter.flush_fid_word_count(&mut buffer)
+    }
+
+    fn attributes_to_extract<'a>(
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>> {
+        index.user_defined_searchable_fields(rtxn).map_err(Into::into)
+    }
+
+    fn attributes_to_skip<'a>(_rtxn: &'a RoTxn, _index: &'a Index) -> Result<Vec<&'a str>> {
+        Ok(vec![])
+    }
+}
--- a/milli/src/update/new/extract/searchable/extract_word_pair_proximity_docids.rs
+++ b/milli/src/update/new/extract/searchable/extract_word_pair_proximity_docids.rs
@@ -0,0 +1,160 @@
+use std::collections::VecDeque;
+use std::rc::Rc;
+
+use heed::RoTxn;
+use obkv::KvReader;
+
+use super::tokenize_document::DocumentTokenizer;
+use super::SearchableExtractor;
+use crate::proximity::{index_proximity, MAX_DISTANCE};
+use crate::update::new::extract::cache::CboCachedSorter;
+use crate::update::new::DocumentChange;
+use crate::{FieldId, GlobalFieldsIdsMap, Index, Result};
+
+pub struct WordPairProximityDocidsExtractor;
+impl SearchableExtractor for WordPairProximityDocidsExtractor {
+    fn attributes_to_extract<'a>(
+        rtxn: &'a RoTxn,
+        index: &'a Index,
+    ) -> Result<Option<Vec<&'a str>>> {
+        index.user_defined_searchable_fields(rtxn).map_err(Into::into)
+    }
+
+    fn attributes_to_skip<'a>(_rtxn: &'a RoTxn, _index: &'a Index) -> Result<Vec<&'a str>> {
+        Ok(vec![])
+    }
+
+    // This method is reimplemented to count the number of words in the document in each field
+    // and to store the docids of the documents that have a number of words in a given field equal to or under than MAX_COUNTED_WORDS.
+    fn extract_document_change(
+        rtxn: &RoTxn,
+        index: &Index,
+        document_tokenizer: &DocumentTokenizer,
+        fields_ids_map: &mut GlobalFieldsIdsMap,
+        cached_sorter: &mut CboCachedSorter,
+        document_change: DocumentChange,
+    ) -> Result<()> {
+        let mut key_buffer = Vec::new();
+        let mut del_word_pair_proximity = Vec::new();
+        let mut add_word_pair_proximity = Vec::new();
+        let mut word_positions: VecDeque<(Rc<str>, u16)> =
+            VecDeque::with_capacity(MAX_DISTANCE as usize);
+
+        let docid = document_change.docid();
+        match document_change {
+            DocumentChange::Deletion(inner) => {
+                let document = inner.current(rtxn, index)?.unwrap();
+                process_document_tokens(
+                    document,
+                    document_tokenizer,
+                    fields_ids_map,
+                    &mut word_positions,
+                    &mut |(w1, w2), prox| {
+                        del_word_pair_proximity.push(((w1, w2), prox));
+                    },
+                )?;
+            }
+            DocumentChange::Update(inner) => {
+                let document = inner.current(rtxn, index)?.unwrap();
+                process_document_tokens(
+                    document,
+                    document_tokenizer,
+                    fields_ids_map,
+                    &mut word_positions,
+                    &mut |(w1, w2), prox| {
+                        del_word_pair_proximity.push(((w1, w2), prox));
+                    },
+                )?;
+                let document = inner.new();
+                process_document_tokens(
+                    document,
+                    document_tokenizer,
+                    fields_ids_map,
+                    &mut word_positions,
+                    &mut |(w1, w2), prox| {
+                        add_word_pair_proximity.push(((w1, w2), prox));
+                    },
+                )?;
+            }
+            DocumentChange::Insertion(inner) => {
+                let document = inner.new();
+                process_document_tokens(
+                    document,
+                    document_tokenizer,
+                    fields_ids_map,
+                    &mut word_positions,
+                    &mut |(w1, w2), prox| {
+                        add_word_pair_proximity.push(((w1, w2), prox));
+                    },
+                )?;
+            }
+        }
+
+        del_word_pair_proximity.sort_unstable();
+        del_word_pair_proximity.dedup_by(|(k1, _), (k2, _)| k1 == k2);
+        for ((w1, w2), prox) in del_word_pair_proximity.iter() {
+            let key = build_key(*prox, w1, w2, &mut key_buffer);
+            cached_sorter.insert_del_u32(key, docid);
+        }
+
+        add_word_pair_proximity.sort_unstable();
+        add_word_pair_proximity.dedup_by(|(k1, _), (k2, _)| k1 == k2);
+        for ((w1, w2), prox) in add_word_pair_proximity.iter() {
+            let key = build_key(*prox, w1, w2, &mut key_buffer);
+            cached_sorter.insert_add_u32(key, docid);
+        }
+        Ok(())
+    }
+}
+
+fn build_key<'a>(prox: u8, w1: &str, w2: &str, key_buffer: &'a mut Vec<u8>) -> &'a [u8] {
+    key_buffer.clear();
+    key_buffer.push(prox);
+    key_buffer.extend_from_slice(w1.as_bytes());
+    key_buffer.push(0);
+    key_buffer.extend_from_slice(w2.as_bytes());
+    key_buffer.as_slice()
+}
+
+fn word_positions_into_word_pair_proximity(
+    word_positions: &mut VecDeque<(Rc<str>, u16)>,
+    word_pair_proximity: &mut impl FnMut((Rc<str>, Rc<str>), u8),
+) -> Result<()> {
+    let (head_word, head_position) = word_positions.pop_front().unwrap();
+    for (word, position) in word_positions.iter() {
+        let prox = index_proximity(head_position as u32, *position as u32) as u8;
+        if prox > 0 && prox < MAX_DISTANCE as u8 {
+            word_pair_proximity((head_word.clone(), word.clone()), prox);
+        }
+    }
+    Ok(())
+}
+
+fn process_document_tokens(
+    document: &KvReader<FieldId>,
+    document_tokenizer: &DocumentTokenizer,
+    fields_ids_map: &mut GlobalFieldsIdsMap,
+    word_positions: &mut VecDeque<(Rc<str>, u16)>,
+    word_pair_proximity: &mut impl FnMut((Rc<str>, Rc<str>), u8),
+) -> Result<()> {
+    let mut token_fn = |_fname: &str, _fid: FieldId, pos: u16, word: &str| {
+        // drain the proximity window until the head word is considered close to the word we are inserting.
+        while word_positions
+            .front()
+            .map_or(false, |(_w, p)| index_proximity(*p as u32, pos as u32) >= MAX_DISTANCE)
+        {
+            word_positions_into_word_pair_proximity(word_positions, word_pair_proximity)?;
+        }
+
+        // insert the new word.
+        word_positions.push_back((Rc::from(word), pos));
+        Ok(())
+    };
+    document_tokenizer.tokenize_document(document, fields_ids_map, &mut token_fn)?;
+
+    while !word_positions.is_empty() {
+        word_positions_into_word_pair_proximity(word_positions, word_pair_proximity)?;
+    }
+
+    Ok(())
+}
--- a/milli/src/update/new/extract/searchable/mod.rs
+++ b/milli/src/update/new/extract/searchable/mod.rs
@@ -0,0 +1,126 @@
+mod extract_fid_word_count_docids;
+mod extract_word_docids;
+mod extract_word_pair_proximity_docids;
+mod tokenize_document;
+
+use std::fs::File;
+
+pub use extract_word_docids::{WordDocidsExtractors, WordDocidsMergers};
+pub use extract_word_pair_proximity_docids::WordPairProximityDocidsExtractor;
+use grenad::Merger;
+use heed::RoTxn;
+use rayon::iter::{IntoParallelIterator, ParallelBridge, ParallelIterator};
+use tokenize_document::{tokenizer_builder, DocumentTokenizer};
+
+use super::cache::CboCachedSorter;
+use super::{DocidsExtractor, HashMapMerger};
+use crate::update::new::{DocumentChange, ItemsPool};
+use crate::update::{GrenadParameters, MergeDeladdCboRoaringBitmaps};
+use crate::{GlobalFieldsIdsMap, Index, Result, MAX_POSITION_PER_ATTRIBUTE};
+
+pub trait SearchableExtractor {
+    fn run_extraction(
+        index: &Index,
+        fields_ids_map: &GlobalFieldsIdsMap,
+        indexer: GrenadParameters,
+        document_changes: impl IntoParallelIterator<Item = Result<DocumentChange>>,
+    ) -> Result<HashMapMerger> {
+        let max_memory = indexer.max_memory_by_thread();
+
+        let rtxn = index.read_txn()?;
+        let stop_words = index.stop_words(&rtxn)?;
+        let allowed_separators = index.allowed_separators(&rtxn)?;
+        let allowed_separators: Option<Vec<_>> =
+            allowed_separators.as_ref().map(|s| s.iter().map(String::as_str).collect());
+        let dictionary = index.dictionary(&rtxn)?;
+        let dictionary: Option<Vec<_>> =
+            dictionary.as_ref().map(|s| s.iter().map(String::as_str).collect());
+        let builder = tokenizer_builder(
+            stop_words.as_ref(),
+            allowed_separators.as_deref(),
+            dictionary.as_deref(),
+        );
+        let tokenizer = builder.into_tokenizer();
+
+        let attributes_to_extract = Self::attributes_to_extract(&rtxn, index)?;
+        let attributes_to_skip = Self::attributes_to_skip(&rtxn, index)?;
+        let localized_attributes_rules =
+            index.localized_attributes_rules(&rtxn)?.unwrap_or_default();
+
+        let document_tokenizer = DocumentTokenizer {
+            tokenizer: &tokenizer,
+            attribute_to_extract: attributes_to_extract.as_deref(),
+            attribute_to_skip: attributes_to_skip.as_slice(),
+            localized_attributes_rules: &localized_attributes_rules,
+            max_positions_per_attributes: MAX_POSITION_PER_ATTRIBUTE,
+        };
+
+        let context_pool = ItemsPool::new(|| {
+            Ok((
+                index.read_txn()?,
+                &document_tokenizer,
+                fields_ids_map.clone(),
+                CboCachedSorter::new(),
+            ))
+        });
+
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::extract", "docids_extraction");
+            let _entered = span.enter();
+            document_changes.into_par_iter().try_for_each(|document_change| {
+                context_pool.with(|(rtxn, document_tokenizer, fields_ids_map, cached_sorter)| {
+                    Self::extract_document_change(
+                        &*rtxn,
+                        index,
+                        document_tokenizer,
+                        fields_ids_map,
+                        cached_sorter,
+                        document_change?,
+                    )
+                })
+            })?;
+        }
+        {
+            let mut builder = HashMapMerger::new();
+            let span =
+                tracing::trace_span!(target: "indexing::documents::extract", "merger_building");
+            let _entered = span.enter();
+
+            let readers: Vec<_> = context_pool
+                .into_items()
+                .par_bridge()
+                .map(|(_rtxn, _tokenizer, _fields_ids_map, cached_sorter)| {
+                    cached_sorter.into_sorter()
+                })
+                .collect();
+            builder.extend(readers);
+            Ok(builder)
+        }
+    }
+
+    fn extract_document_change(
+        rtxn: &RoTxn,
+        index: &Index,
+        document_tokenizer: &DocumentTokenizer,
+        fields_ids_map: &mut GlobalFieldsIdsMap,
+        cached_sorter: &mut CboCachedSorter,
+        document_change: DocumentChange,
+    ) -> Result<()>;
+
+    fn attributes_to_extract<'a>(rtxn: &'a RoTxn, index: &'a Index)
+        -> Result<Option<Vec<&'a str>>>;
+
+    fn attributes_to_skip<'a>(rtxn: &'a RoTxn, index: &'a Index) -> Result<Vec<&'a str>>;
+}
+
+impl<T: SearchableExtractor> DocidsExtractor for T {
+    fn run_extraction(
+        index: &Index,
+        fields_ids_map: &GlobalFieldsIdsMap,
+        indexer: GrenadParameters,
+        document_changes: impl IntoParallelIterator<Item = Result<DocumentChange>>,
+    ) -> Result<HashMapMerger> {
+        Self::run_extraction(index, fields_ids_map, indexer, document_changes)
+    }
+}
--- a/milli/src/update/new/extract/searchable/tokenize_document.rs
+++ b/milli/src/update/new/extract/searchable/tokenize_document.rs
@@ -0,0 +1,273 @@
+use std::collections::HashMap;
+
+use charabia::{SeparatorKind, Token, TokenKind, Tokenizer, TokenizerBuilder};
+use serde_json::Value;
+
+use crate::proximity::MAX_DISTANCE;
+use crate::update::new::extract::perm_json_p::{
+    seek_leaf_values_in_array, seek_leaf_values_in_object, select_field,
+};
+use crate::update::new::KvReaderFieldId;
+use crate::{
+    FieldId, GlobalFieldsIdsMap, InternalError, LocalizedAttributesRule, Result, UserError,
+    MAX_WORD_LENGTH,
+};
+
+pub struct DocumentTokenizer<'a> {
+    pub tokenizer: &'a Tokenizer<'a>,
+    pub attribute_to_extract: Option<&'a [&'a str]>,
+    pub attribute_to_skip: &'a [&'a str],
+    pub localized_attributes_rules: &'a [LocalizedAttributesRule],
+    pub max_positions_per_attributes: u32,
+}
+
+impl<'a> DocumentTokenizer<'a> {
+    pub fn tokenize_document(
+        &self,
+        obkv: &KvReaderFieldId,
+        field_id_map: &mut GlobalFieldsIdsMap,
+        token_fn: &mut impl FnMut(&str, FieldId, u16, &str) -> Result<()>,
+    ) -> Result<()> {
+        let mut field_position = HashMap::new();
+        let mut field_name = String::new();
+        for (field_id, field_bytes) in obkv {
+            let Some(field_name) = field_id_map.name(field_id).map(|s| {
+                field_name.clear();
+                field_name.push_str(s);
+                &field_name
+            }) else {
+                unreachable!("field id not found in field id map");
+            };
+
+            let mut tokenize_field = |name: &str, value: &Value| {
+                let Some(field_id) = field_id_map.id_or_insert(name) else {
+                    return Err(UserError::AttributeLimitReached.into());
+                };
+
+                let position = field_position
+                    .entry(field_id)
+                    .and_modify(|counter| *counter += MAX_DISTANCE)
+                    .or_insert(0);
+                if *position as u32 >= self.max_positions_per_attributes {
+                    return Ok(());
+                }
+
+                match value {
+                    Value::Number(n) => {
+                        let token = n.to_string();
+                        if let Ok(position) = (*position).try_into() {
+                            token_fn(name, field_id, position, token.as_str())?;
+                        }
+
+                        Ok(())
+                    }
+                    Value::String(text) => {
+                        // create an iterator of token with their positions.
+                        let locales = self
+                            .localized_attributes_rules
+                            .iter()
+                            .find(|rule| rule.match_str(field_name))
+                            .map(|rule| rule.locales());
+                        let tokens = process_tokens(
+                            *position,
+                            self.tokenizer.tokenize_with_allow_list(text.as_str(), locales),
+                        )
+                        .take_while(|(p, _)| (*p as u32) < self.max_positions_per_attributes);
+
+                        for (index, token) in tokens {
+                            // keep a word only if it is not empty and fit in a LMDB key.
+                            let token = token.lemma().trim();
+                            if !token.is_empty() && token.len() <= MAX_WORD_LENGTH {
+                                *position = index;
+                                if let Ok(position) = (*position).try_into() {
+                                    token_fn(name, field_id, position, token)?;
+                                }
+                            }
+                        }
+
+                        Ok(())
+                    }
+                    _ => Ok(()),
+                }
+            };
+
+            // if the current field is searchable or contains a searchable attribute
+            if select_field(field_name, self.attribute_to_extract, self.attribute_to_skip) {
+                // parse json.
+                match serde_json::from_slice(field_bytes).map_err(InternalError::SerdeJson)? {
+                    Value::Object(object) => seek_leaf_values_in_object(
+                        &object,
+                        self.attribute_to_extract,
+                        self.attribute_to_skip,
+                        field_name,
+                        &mut tokenize_field,
+                    )?,
+                    Value::Array(array) => seek_leaf_values_in_array(
+                        &array,
+                        self.attribute_to_extract,
+                        self.attribute_to_skip,
+                        field_name,
+                        &mut tokenize_field,
+                    )?,
+                    value => tokenize_field(field_name, &value)?,
+                }
+            }
+        }
+
+        Ok(())
+    }
+}
+
+/// take an iterator on tokens and compute their relative position depending on separator kinds
+/// if it's an `Hard` separator we add an additional relative proximity of MAX_DISTANCE between words,
+/// else we keep the standard proximity of 1 between words.
+fn process_tokens<'a>(
+    start_offset: u32,
+    tokens: impl Iterator<Item = Token<'a>>,
+) -> impl Iterator<Item = (u32, Token<'a>)> {
+    tokens
+        .skip_while(|token| token.is_separator())
+        .scan((start_offset, None), |(offset, prev_kind), mut token| {
+            match token.kind {
+                TokenKind::Word | TokenKind::StopWord if !token.lemma().is_empty() => {
+                    *offset += match *prev_kind {
+                        Some(TokenKind::Separator(SeparatorKind::Hard)) => MAX_DISTANCE,
+                        Some(_) => 1,
+                        None => 0,
+                    };
+                    *prev_kind = Some(token.kind)
+                }
+                TokenKind::Separator(SeparatorKind::Hard) => {
+                    *prev_kind = Some(token.kind);
+                }
+                TokenKind::Separator(SeparatorKind::Soft)
+                    if *prev_kind != Some(TokenKind::Separator(SeparatorKind::Hard)) =>
+                {
+                    *prev_kind = Some(token.kind);
+                }
+                _ => token.kind = TokenKind::Unknown,
+            }
+            Some((*offset, token))
+        })
+        .filter(|(_, t)| t.is_word())
+}
+
+/// Factorize tokenizer building.
+pub fn tokenizer_builder<'a>(
+    stop_words: Option<&'a fst::Set<&'a [u8]>>,
+    allowed_separators: Option<&'a [&str]>,
+    dictionary: Option<&'a [&str]>,
+) -> TokenizerBuilder<'a, &'a [u8]> {
+    let mut tokenizer_builder = TokenizerBuilder::new();
+    if let Some(stop_words) = stop_words {
+        tokenizer_builder.stop_words(stop_words);
+    }
+    if let Some(dictionary) = dictionary {
+        tokenizer_builder.words_dict(dictionary);
+    }
+    if let Some(separators) = allowed_separators {
+        tokenizer_builder.separators(separators);
+    }
+
+    tokenizer_builder
+}
+
+#[cfg(test)]
+mod test {
+    use charabia::TokenizerBuilder;
+    use meili_snap::snapshot;
+    use obkv::KvReader;
+    use serde_json::json;
+
+    use super::*;
+    use crate::FieldsIdsMap;
+
+    #[test]
+    fn test_tokenize_document() {
+        let mut fields_ids_map = FieldsIdsMap::new();
+
+        let field_1 = json!({
+                "name": "doggo",
+                "age": 10,
+        });
+
+        let field_2 = json!({
+                "catto": {
+                    "name": "pesti",
+                    "age": 23,
+                }
+        });
+
+        let field_3 = json!(["doggo", "catto"]);
+        let field_4 = json!("UNSEARCHABLE");
+        let field_5 = json!({"nope": "unsearchable"});
+
+        let mut obkv = obkv::KvWriter::memory();
+        let field_1_id = fields_ids_map.insert("doggo").unwrap();
+        let field_1 = serde_json::to_string(&field_1).unwrap();
+        obkv.insert(field_1_id, field_1.as_bytes()).unwrap();
+        let field_2_id = fields_ids_map.insert("catto").unwrap();
+        let field_2 = serde_json::to_string(&field_2).unwrap();
+        obkv.insert(field_2_id, field_2.as_bytes()).unwrap();
+        let field_3_id = fields_ids_map.insert("doggo.name").unwrap();
+        let field_3 = serde_json::to_string(&field_3).unwrap();
+        obkv.insert(field_3_id, field_3.as_bytes()).unwrap();
+        let field_4_id = fields_ids_map.insert("not-me").unwrap();
+        let field_4 = serde_json::to_string(&field_4).unwrap();
+        obkv.insert(field_4_id, field_4.as_bytes()).unwrap();
+        let field_5_id = fields_ids_map.insert("me-nether").unwrap();
+        let field_5 = serde_json::to_string(&field_5).unwrap();
+        obkv.insert(field_5_id, field_5.as_bytes()).unwrap();
+        let value = obkv.into_inner().unwrap();
+        let obkv = KvReader::from_slice(value.as_slice());
+
+        let mut tb = TokenizerBuilder::default();
+        let document_tokenizer = DocumentTokenizer {
+            tokenizer: &tb.build(),
+            attribute_to_extract: None,
+            attribute_to_skip: &["not-me", "me-nether.nope"],
+            localized_attributes_rules: &[],
+            max_positions_per_attributes: 1000,
+        };
+
+        let fields_ids_map_lock = std::sync::RwLock::new(fields_ids_map);
+        let mut global_fields_ids_map = GlobalFieldsIdsMap::new(&fields_ids_map_lock);
+
+        let mut words = std::collections::BTreeMap::new();
+        document_tokenizer
+            .tokenize_document(obkv, &mut global_fields_ids_map, &mut |_fname, fid, pos, word| {
+                words.insert([fid, pos], word.to_string());
+                Ok(())
+            })
+            .unwrap();
+
+        snapshot!(format!("{:#?}", words), @r###"
+        {
+            [
+                2,
+                0,
+            ]: "doggo",
+            [
+                2,
+                MAX_DISTANCE,
+            ]: "doggo",
+            [
+                2,
+                16,
+            ]: "catto",
+            [
+                3,
+                0,
+            ]: "10",
+            [
+                4,
+                0,
+            ]: "pesti",
+            [
+                5,
+                0,
+            ]: "23",
+        }
+        "###);
+    }
+}
--- a/milli/src/update/new/indexer/document_deletion.rs
+++ b/milli/src/update/new/indexer/document_deletion.rs
@@ -0,0 +1,42 @@
+use std::sync::Arc;
+
+use rayon::iter::{IndexedParallelIterator, IntoParallelIterator, ParallelIterator};
+use roaring::RoaringBitmap;
+
+use super::DocumentChanges;
+use crate::update::new::{Deletion, DocumentChange, ItemsPool};
+use crate::{FieldsIdsMap, Index, Result};
+
+pub struct DocumentDeletion {
+    pub to_delete: RoaringBitmap,
+}
+
+impl DocumentDeletion {
+    pub fn new() -> Self {
+        Self { to_delete: Default::default() }
+    }
+
+    pub fn delete_documents_by_docids(&mut self, docids: RoaringBitmap) {
+        self.to_delete |= docids;
+    }
+}
+
+impl<'p> DocumentChanges<'p> for DocumentDeletion {
+    type Parameter = &'p Index;
+
+    fn document_changes(
+        self,
+        _fields_ids_map: &mut FieldsIdsMap,
+        param: Self::Parameter,
+    ) -> Result<impl IndexedParallelIterator<Item = Result<DocumentChange>> + Clone + 'p> {
+        let index = param;
+        let items = Arc::new(ItemsPool::new(|| index.read_txn().map_err(crate::Error::from)));
+        let to_delete: Vec<_> = self.to_delete.into_iter().collect();
+        Ok(to_delete.into_par_iter().map_with(items, |items, docid| {
+            items.with(|rtxn| {
+                let current = index.document(rtxn, docid)?;
+                Ok(DocumentChange::Deletion(Deletion::create(docid, current.boxed())))
+            })
+        }))
+    }
+}
--- a/milli/src/update/new/indexer/document_operation.rs
+++ b/milli/src/update/new/indexer/document_operation.rs
@@ -0,0 +1,392 @@
+use std::borrow::Cow;
+use std::collections::{BTreeMap, HashMap};
+use std::sync::Arc;
+
+use heed::types::Bytes;
+use heed::RoTxn;
+use memmap2::Mmap;
+use rayon::iter::{IndexedParallelIterator, IntoParallelIterator, ParallelIterator};
+use IndexDocumentsMethod as Idm;
+
+use super::super::document_change::DocumentChange;
+use super::super::items_pool::ItemsPool;
+use super::super::{CowStr, TopLevelMap};
+use super::DocumentChanges;
+use crate::documents::{DocumentIdExtractionError, PrimaryKey};
+use crate::update::new::{Deletion, Insertion, KvReaderFieldId, KvWriterFieldId, Update};
+use crate::update::{AvailableIds, IndexDocumentsMethod};
+use crate::{DocumentId, Error, FieldsIdsMap, Index, Result, UserError};
+
+pub struct DocumentOperation<'pl> {
+    operations: Vec<Payload<'pl>>,
+    index_documents_method: IndexDocumentsMethod,
+}
+
+pub enum Payload<'pl> {
+    Addition(&'pl [u8]),
+    Deletion(Vec<String>),
+}
+
+pub struct PayloadStats {
+    pub document_count: usize,
+    pub bytes: u64,
+}
+
+#[derive(Clone)]
+enum InnerDocOp<'pl> {
+    Addition(DocumentOffset<'pl>),
+    Deletion,
+}
+
+/// Represents an offset where a document lives
+/// in an mmapped grenad reader file.
+#[derive(Clone)]
+pub struct DocumentOffset<'pl> {
+    /// The mmapped payload files.
+    pub content: &'pl [u8],
+}
+
+impl<'pl> DocumentOperation<'pl> {
+    pub fn new(method: IndexDocumentsMethod) -> Self {
+        Self { operations: Default::default(), index_documents_method: method }
+    }
+
+    /// TODO please give me a type
+    /// The payload is expected to be in the grenad format
+    pub fn add_documents(&mut self, payload: &'pl Mmap) -> Result<PayloadStats> {
+        payload.advise(memmap2::Advice::Sequential)?;
+        let document_count =
+            memchr::memmem::find_iter(&payload[..], "}{").count().saturating_add(1);
+        self.operations.push(Payload::Addition(&payload[..]));
+        Ok(PayloadStats { bytes: payload.len() as u64, document_count })
+    }
+
+    pub fn delete_documents(&mut self, to_delete: Vec<String>) {
+        self.operations.push(Payload::Deletion(to_delete))
+    }
+}
+
+impl<'p, 'pl: 'p> DocumentChanges<'p> for DocumentOperation<'pl> {
+    type Parameter = (&'p Index, &'p RoTxn<'p>, &'p PrimaryKey<'p>);
+
+    fn document_changes(
+        self,
+        fields_ids_map: &mut FieldsIdsMap,
+        param: Self::Parameter,
+    ) -> Result<impl IndexedParallelIterator<Item = Result<DocumentChange>> + Clone + 'p> {
+        let (index, rtxn, primary_key) = param;
+
+        let documents_ids = index.documents_ids(rtxn)?;
+        let mut available_docids = AvailableIds::new(&documents_ids);
+        let mut docids_version_offsets = HashMap::<CowStr<'pl>, _>::new();
+
+        for operation in self.operations {
+            match operation {
+                Payload::Addition(payload) => {
+                    let mut iter =
+                        serde_json::Deserializer::from_slice(payload).into_iter::<TopLevelMap>();
+
+                    /// TODO manage the error
+                    let mut previous_offset = 0;
+                    while let Some(document) = iter.next().transpose().unwrap() {
+                        // TODO Fetch all document fields to fill the fields ids map
+                        document.0.keys().for_each(|key| {
+                            fields_ids_map.insert(key.as_ref());
+                        });
+
+                        // TODO we must manage the TooManyDocumentIds,InvalidDocumentId
+                        //      we must manage the unwrap
+                        let external_document_id =
+                            match primary_key.document_id_from_top_level_map(&document)? {
+                                Ok(document_id) => Ok(document_id),
+                                Err(DocumentIdExtractionError::InvalidDocumentId(e)) => Err(e),
+                                Err(DocumentIdExtractionError::MissingDocumentId) => {
+                                    Err(UserError::MissingDocumentId {
+                                        primary_key: primary_key.name().to_string(),
+                                        document: document.try_into().unwrap(),
+                                    })
+                                }
+                                Err(DocumentIdExtractionError::TooManyDocumentIds(_)) => {
+                                    Err(UserError::TooManyDocumentIds {
+                                        primary_key: primary_key.name().to_string(),
+                                        document: document.try_into().unwrap(),
+                                    })
+                                }
+                            }?;
+
+                        let current_offset = iter.byte_offset();
+                        let document_operation = InnerDocOp::Addition(DocumentOffset {
+                            content: &payload[previous_offset..current_offset],
+                        });
+
+                        match docids_version_offsets.get_mut(external_document_id.as_ref()) {
+                            None => {
+                                let docid = match index
+                                    .external_documents_ids()
+                                    .get(rtxn, &external_document_id)?
+                                {
+                                    Some(docid) => docid,
+                                    None => available_docids
+                                        .next()
+                                        .ok_or(Error::UserError(UserError::DocumentLimitReached))?,
+                                };
+
+                                docids_version_offsets.insert(
+                                    external_document_id,
+                                    (docid, vec![document_operation]),
+                                );
+                            }
+                            Some((_, offsets)) => {
+                                let useless_previous_addition = match self.index_documents_method {
+                                    IndexDocumentsMethod::ReplaceDocuments => {
+                                        MergeDocumentForReplacement::USELESS_PREVIOUS_CHANGES
+                                    }
+                                    IndexDocumentsMethod::UpdateDocuments => {
+                                        MergeDocumentForUpdates::USELESS_PREVIOUS_CHANGES
+                                    }
+                                };
+
+                                if useless_previous_addition {
+                                    offsets.clear();
+                                }
+
+                                offsets.push(document_operation);
+                            }
+                        }
+
+                        previous_offset = iter.byte_offset();
+                    }
+                }
+                Payload::Deletion(to_delete) => {
+                    for external_document_id in to_delete {
+                        match docids_version_offsets.get_mut(external_document_id.as_str()) {
+                            None => {
+                                let docid = match index
+                                    .external_documents_ids()
+                                    .get(rtxn, &external_document_id)?
+                                {
+                                    Some(docid) => docid,
+                                    None => available_docids
+                                        .next()
+                                        .ok_or(Error::UserError(UserError::DocumentLimitReached))?,
+                                };
+
+                                docids_version_offsets.insert(
+                                    CowStr(external_document_id.into()),
+                                    (docid, vec![InnerDocOp::Deletion]),
+                                );
+                            }
+                            Some((_, offsets)) => {
+                                offsets.clear();
+                                offsets.push(InnerDocOp::Deletion);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        /// TODO is it the best way to provide FieldsIdsMap to the parallel iterator?
+        let fields_ids_map = fields_ids_map.clone();
+        // TODO We must drain the HashMap into a Vec because rayon::hash_map::IntoIter: !Clone
+        let mut docids_version_offsets: Vec<_> = docids_version_offsets.drain().collect();
+        // Reorder the offsets to make sure we iterate on the file sequentially
+        let sort_function_key = match self.index_documents_method {
+            Idm::ReplaceDocuments => MergeDocumentForReplacement::sort_key,
+            Idm::UpdateDocuments => MergeDocumentForUpdates::sort_key,
+        };
+
+        // And finally sort them
+        docids_version_offsets.sort_unstable_by_key(|(_, (_, docops))| sort_function_key(docops));
+
+        Ok(docids_version_offsets.into_par_iter().map_with(
+            Arc::new(ItemsPool::new(|| index.read_txn().map_err(crate::Error::from))),
+            move |context_pool, (external_docid, (internal_docid, operations))| {
+                context_pool.with(|rtxn| {
+                    let document_merge_function = match self.index_documents_method {
+                        Idm::ReplaceDocuments => MergeDocumentForReplacement::merge,
+                        Idm::UpdateDocuments => MergeDocumentForUpdates::merge,
+                    };
+
+                    document_merge_function(
+                        rtxn,
+                        index,
+                        &fields_ids_map,
+                        internal_docid,
+                        external_docid.to_string(), // TODO do not clone
+                        &operations,
+                    )
+                })
+            },
+        ))
+    }
+}
+
+trait MergeChanges {
+    /// Wether the payloads in the list of operations are useless or not.
+    const USELESS_PREVIOUS_CHANGES: bool;
+
+    /// Returns a key that is used to order the payloads the right way.
+    fn sort_key(docops: &[InnerDocOp]) -> usize;
+
+    fn merge(
+        rtxn: &RoTxn,
+        index: &Index,
+        fields_ids_map: &FieldsIdsMap,
+        docid: DocumentId,
+        external_docid: String,
+        operations: &[InnerDocOp],
+    ) -> Result<DocumentChange>;
+}
+
+struct MergeDocumentForReplacement;
+
+impl MergeChanges for MergeDocumentForReplacement {
+    const USELESS_PREVIOUS_CHANGES: bool = true;
+
+    /// Reorders to read only the last change.
+    fn sort_key(docops: &[InnerDocOp]) -> usize {
+        let f = |ido: &_| match ido {
+            InnerDocOp::Addition(add) => Some(add.content.as_ptr() as usize),
+            InnerDocOp::Deletion => None,
+        };
+        docops.iter().rev().find_map(f).unwrap_or(0)
+    }
+
+    /// Returns only the most recent version of a document based on the updates from the payloads.
+    ///
+    /// This function is only meant to be used when doing a replacement and not an update.
+    fn merge(
+        rtxn: &RoTxn,
+        index: &Index,
+        fields_ids_map: &FieldsIdsMap,
+        docid: DocumentId,
+        external_docid: String,
+        operations: &[InnerDocOp],
+    ) -> Result<DocumentChange> {
+        let current = index.documents.remap_data_type::<Bytes>().get(rtxn, &docid)?;
+        let current: Option<&KvReaderFieldId> = current.map(Into::into);
+
+        match operations.last() {
+            Some(InnerDocOp::Addition(DocumentOffset { content })) => {
+                let map: TopLevelMap = serde_json::from_slice(content).unwrap();
+                let mut document_entries = Vec::new();
+                for (key, v) in map.0 {
+                    let id = fields_ids_map.id(key.as_ref()).unwrap();
+                    document_entries.push((id, v));
+                }
+
+                document_entries.sort_unstable_by_key(|(id, _)| *id);
+
+                let mut writer = KvWriterFieldId::memory();
+                document_entries
+                    .into_iter()
+                    .for_each(|(id, value)| writer.insert(id, value.get()).unwrap());
+                let new = writer.into_boxed();
+
+                match current {
+                    Some(current) => {
+                        let update = Update::create(docid, current.boxed(), new);
+                        Ok(DocumentChange::Update(update))
+                    }
+                    None => Ok(DocumentChange::Insertion(Insertion::create(docid, new))),
+                }
+            }
+            Some(InnerDocOp::Deletion) => {
+                let deletion = match current {
+                    Some(current) => Deletion::create(docid, current.boxed()),
+                    None => todo!("Do that with Louis"),
+                };
+                Ok(DocumentChange::Deletion(deletion))
+            }
+            None => unreachable!("We must not have empty set of operations on a document"),
+        }
+    }
+}
+
+struct MergeDocumentForUpdates;
+
+impl MergeChanges for MergeDocumentForUpdates {
+    const USELESS_PREVIOUS_CHANGES: bool = false;
+
+    /// Reorders to read the first changes first so that it's faster to read the first one and then the rest.
+    fn sort_key(docops: &[InnerDocOp]) -> usize {
+        let f = |ido: &_| match ido {
+            InnerDocOp::Addition(add) => Some(add.content.as_ptr() as usize),
+            InnerDocOp::Deletion => None,
+        };
+        docops.iter().find_map(f).unwrap_or(0)
+    }
+
+    /// Reads the previous version of a document from the database, the new versions
+    /// in the grenad update files and merges them to generate a new boxed obkv.
+    ///
+    /// This function is only meant to be used when doing an update and not a replacement.
+    fn merge(
+        rtxn: &RoTxn,
+        index: &Index,
+        fields_ids_map: &FieldsIdsMap,
+        docid: DocumentId,
+        external_docid: String,
+        operations: &[InnerDocOp],
+    ) -> Result<DocumentChange> {
+        let mut document = BTreeMap::<_, Cow<_>>::new();
+        let current = index.documents.remap_data_type::<Bytes>().get(rtxn, &docid)?;
+        let current: Option<&KvReaderFieldId> = current.map(Into::into);
+
+        if operations.is_empty() {
+            unreachable!("We must not have empty set of operations on a document");
+        }
+
+        let last_deletion = operations.iter().rposition(|op| matches!(op, InnerDocOp::Deletion));
+        let operations = &operations[last_deletion.map_or(0, |i| i + 1)..];
+
+        // If there was a deletion we must not start
+        // from the original document but from scratch.
+        if last_deletion.is_none() {
+            if let Some(current) = current {
+                current.into_iter().for_each(|(k, v)| {
+                    document.insert(k, v.into());
+                });
+            }
+        }
+
+        if operations.is_empty() {
+            let deletion = match current {
+                Some(current) => Deletion::create(docid, current.boxed()),
+                None => todo!("Do that with Louis"),
+            };
+            return Ok(DocumentChange::Deletion(deletion));
+        }
+
+        for operation in operations {
+            let DocumentOffset { content } = match operation {
+                InnerDocOp::Addition(offset) => offset,
+                InnerDocOp::Deletion => {
+                    unreachable!("Deletion in document operations")
+                }
+            };
+
+            let map: TopLevelMap = serde_json::from_slice(content).unwrap();
+            for (key, v) in map.0 {
+                let id = fields_ids_map.id(key.as_ref()).unwrap();
+                document.insert(id, v.get().as_bytes().to_vec().into());
+            }
+        }
+
+        let mut writer = KvWriterFieldId::memory();
+        document.into_iter().for_each(|(id, value)| writer.insert(id, value).unwrap());
+        let new = writer.into_boxed();
+
+        match current {
+            Some(current) => {
+                let update = Update::create(docid, current.boxed(), new);
+                Ok(DocumentChange::Update(update))
+            }
+            None => {
+                let insertion = Insertion::create(docid, new);
+                Ok(DocumentChange::Insertion(insertion))
+            }
+        }
+    }
+}
--- a/milli/src/update/new/indexer/mod.rs
+++ b/milli/src/update/new/indexer/mod.rs
@@ -0,0 +1,316 @@
+use std::sync::RwLock;
+use std::thread::{self, Builder};
+
+use big_s::S;
+pub use document_deletion::DocumentDeletion;
+pub use document_operation::DocumentOperation;
+use heed::{RoTxn, RwTxn};
+pub use partial_dump::PartialDump;
+use rayon::iter::{IndexedParallelIterator, IntoParallelIterator, ParallelIterator};
+use rayon::ThreadPool;
+pub use update_by_function::UpdateByFunction;
+
+use super::channel::*;
+use super::document_change::DocumentChange;
+use super::extract::*;
+use super::merger::merge_grenad_entries;
+use super::{ItemsPool, StdResult, TopLevelMap};
+use crate::documents::{PrimaryKey, DEFAULT_PRIMARY_KEY};
+use crate::update::new::channel::ExtractorSender;
+use crate::update::GrenadParameters;
+use crate::{FieldsIdsMap, GlobalFieldsIdsMap, Index, Result, UserError};
+
+mod document_deletion;
+mod document_operation;
+mod partial_dump;
+mod update_by_function;
+
+pub trait DocumentChanges<'p> {
+    type Parameter: 'p;
+
+    fn document_changes(
+        self,
+        fields_ids_map: &mut FieldsIdsMap,
+        param: Self::Parameter,
+    ) -> Result<impl IndexedParallelIterator<Item = Result<DocumentChange>> + Clone + 'p>;
+}
+
+/// This is the main function of this crate.
+///
+/// Give it the output of the [`Indexer::document_changes`] method and it will execute it in the [`rayon::ThreadPool`].
+///
+/// TODO return stats
+pub fn index<PI>(
+    wtxn: &mut RwTxn,
+    index: &Index,
+    fields_ids_map: FieldsIdsMap,
+    pool: &ThreadPool,
+    document_changes: PI,
+) -> Result<()>
+where
+    PI: IndexedParallelIterator<Item = Result<DocumentChange>> + Send + Clone,
+{
+    let (merger_sender, writer_receiver) = merger_writer_channel(10_000);
+    // This channel acts as a rendezvous point to ensure that we are one task ahead
+    let (extractor_sender, merger_receiver) = extractors_merger_channels(4);
+
+    let fields_ids_map_lock = RwLock::new(fields_ids_map);
+    let global_fields_ids_map = GlobalFieldsIdsMap::new(&fields_ids_map_lock);
+    let global_fields_ids_map_clone = global_fields_ids_map.clone();
+
+    thread::scope(|s| {
+        // TODO manage the errors correctly
+        let current_span = tracing::Span::current();
+        let handle = Builder::new().name(S("indexer-extractors")).spawn_scoped(s, move || {
+            pool.in_place_scope(|_s| {
+                    let span = tracing::trace_span!(target: "indexing::documents", parent: &current_span, "extract");
+                    let _entered = span.enter();
+                    let document_changes = document_changes.into_par_iter();
+
+                    // document but we need to create a function that collects and compresses documents.
+                    let document_sender = extractor_sender.document_sender();
+                    document_changes.clone().into_par_iter().try_for_each(|result| {
+                        match result? {
+                            DocumentChange::Deletion(deletion) => {
+                                let docid = deletion.docid();
+                                document_sender.delete(docid).unwrap();
+                            }
+                            DocumentChange::Update(update) => {
+                                let docid = update.docid();
+                                let content = update.new();
+                                document_sender.insert(docid, content.boxed()).unwrap();
+                            }
+                            DocumentChange::Insertion(insertion) => {
+                                let docid = insertion.docid();
+                                let content = insertion.new();
+                                document_sender.insert(docid, content.boxed()).unwrap();
+                                // extracted_dictionary_sender.send(self, dictionary: &[u8]);
+                            }
+                        }
+                        Ok(()) as Result<_>
+                    })?;
+
+                    document_sender.finish().unwrap();
+
+                    const TEN_GIB: usize = 10 * 1024 * 1024 * 1024;
+                    let max_memory = TEN_GIB / dbg!(rayon::current_num_threads());
+                    let grenad_parameters = GrenadParameters {
+                        max_memory: Some(max_memory),
+                        ..GrenadParameters::default()
+                    };
+
+                    {
+                        let span = tracing::trace_span!(target: "indexing::documents::extract", "faceted");
+                        let _entered = span.enter();
+                        extract_and_send_docids::<
+                            FacetedDocidsExtractor,
+                            FacetDocids,
+                        >(
+                            index,
+                            &global_fields_ids_map,
+                            grenad_parameters,
+                            document_changes.clone(),
+                            &extractor_sender,
+                        )?;
+                    }
+
+                    {
+                        let span = tracing::trace_span!(target: "indexing::documents::extract", "word_docids");
+                        let _entered = span.enter();
+
+                        let WordDocidsMergers {
+                            word_fid_docids,
+                            word_docids,
+                            exact_word_docids,
+                            word_position_docids,
+                            fid_word_count_docids,
+                        } = WordDocidsExtractors::run_extraction(index, &global_fields_ids_map, grenad_parameters, document_changes.clone())?;
+                        extractor_sender.send_searchable::<WordDocids>(word_docids).unwrap();
+                        extractor_sender.send_searchable::<WordFidDocids>(word_fid_docids).unwrap();
+                        extractor_sender.send_searchable::<ExactWordDocids>(exact_word_docids).unwrap();
+                        extractor_sender.send_searchable::<WordPositionDocids>(word_position_docids).unwrap();
+                        extractor_sender.send_searchable::<FidWordCountDocids>(fid_word_count_docids).unwrap();
+                    }
+
+                    // {
+                    //     let span = tracing::trace_span!(target: "indexing::documents::extract", "exact_word_docids");
+                    //     let _entered = span.enter();
+                    //     extract_and_send_docids::<ExactWordDocidsExtractor, ExactWordDocids>(
+                    //         index,
+                    //         &global_fields_ids_map,
+                    //         grenad_parameters,
+                    //         document_changes.clone(),
+                    //         &extractor_sender,
+                    //     )?;
+                    // }
+
+                    // {
+                    //     let span = tracing::trace_span!(target: "indexing::documents::extract", "word_position_docids");
+                    //     let _entered = span.enter();
+                    //     extract_and_send_docids::<WordPositionDocidsExtractor, WordPositionDocids>(
+                    //         index,
+                    //         &global_fields_ids_map,
+                    //         grenad_parameters,
+                    //         document_changes.clone(),
+                    //         &extractor_sender,
+                    //     )?;
+                    // }
+
+                    // {
+                    //     let span = tracing::trace_span!(target: "indexing::documents::extract", "fid_word_count_docids");
+                    //     let _entered = span.enter();
+                    //     extract_and_send_docids::<FidWordCountDocidsExtractor, FidWordCountDocids>(
+                    //         index,
+                    //         &global_fields_ids_map,
+                    //         GrenadParameters::default(),
+                    //         document_changes.clone(),
+                    //         &extractor_sender,
+                    //     )?;
+                    // }
+
+                    {
+                        let span = tracing::trace_span!(target: "indexing::documents::extract", "word_pair_proximity_docids");
+                        let _entered = span.enter();
+                        extract_and_send_docids::<
+                            WordPairProximityDocidsExtractor,
+                            WordPairProximityDocids,
+                        >(
+                            index,
+                            &global_fields_ids_map,
+                            grenad_parameters,
+                            document_changes.clone(),
+                            &extractor_sender,
+                        )?;
+                    }
+
+                    {
+                        let span = tracing::trace_span!(target: "indexing::documents::extract", "FINISH");
+                        let _entered = span.enter();
+                    }
+
+                    // TODO THIS IS TOO MUCH
+                    // - [ ] Extract fieldid docid facet number
+                    // - [ ] Extract fieldid docid facet string
+                    // - [ ] Extract facetid string fst
+                    // - [ ] Extract facetid normalized string strings
+
+                    // TODO Inverted Indexes again
+                    // - [x] Extract fieldid facet isempty docids
+                    // - [x] Extract fieldid facet isnull docids
+                    // - [x] Extract fieldid facet exists docids
+
+                    // TODO This is the normal system
+                    // - [x] Extract fieldid facet number docids
+                    // - [x] Extract fieldid facet string docids
+
+                    Ok(()) as Result<_>
+                })
+        })?;
+
+        // TODO manage the errors correctly
+        let current_span = tracing::Span::current();
+        let handle2 = Builder::new().name(S("indexer-merger")).spawn_scoped(s, move || {
+            let span =
+                tracing::trace_span!(target: "indexing::documents", parent: &current_span, "merge");
+            let _entered = span.enter();
+
+            let rtxn_pool = ItemsPool::new(|| index.read_txn().map_err(Into::into));
+            merge_grenad_entries(
+                merger_receiver,
+                merger_sender,
+                &rtxn_pool,
+                index,
+                global_fields_ids_map_clone,
+            )
+        })?;
+
+        let mut entries_count = 0;
+        for operation in writer_receiver {
+            let database = operation.database(index);
+            match operation.entry() {
+                EntryOperation::Delete(e) => {
+                    if !database.delete(wtxn, e.entry())? {
+                        unreachable!("We tried to delete an unknown key")
+                    }
+                }
+                EntryOperation::Write(e) => {
+                    entries_count += 1;
+                    database.put(wtxn, e.key(), e.value())?
+                }
+            }
+        }
+
+        eprintln!("We saw {entries_count}");
+
+        /// TODO handle the panicking threads
+        handle.join().unwrap()?;
+        handle2.join().unwrap()?;
+
+        Ok(()) as Result<_>
+    })?;
+
+    let fields_ids_map = fields_ids_map_lock.into_inner().unwrap();
+    index.put_fields_ids_map(wtxn, &fields_ids_map)?;
+
+    Ok(())
+}
+
+/// TODO: GrenadParameters::default() should be removed in favor a passed parameter
+/// TODO: manage the errors correctly
+/// TODO: we must have a single trait that also gives the extractor type
+fn extract_and_send_docids<E: DocidsExtractor, D: MergerOperationType>(
+    index: &Index,
+    fields_ids_map: &GlobalFieldsIdsMap,
+    indexer: GrenadParameters,
+    document_changes: impl IntoParallelIterator<Item = Result<DocumentChange>>,
+    sender: &ExtractorSender,
+) -> Result<()> {
+    let merger = E::run_extraction(index, fields_ids_map, indexer, document_changes)?;
+    Ok(sender.send_searchable::<D>(merger).unwrap())
+}
+
+/// Returns the primary key *field id* that has already been set for this index or the
+/// one we will guess by searching for the first key that contains "id" as a substring.
+/// TODO move this elsewhere
+pub fn retrieve_or_guess_primary_key<'a>(
+    rtxn: &'a RoTxn<'a>,
+    index: &Index,
+    fields_ids_map: &mut FieldsIdsMap,
+    first_document: Option<&'a TopLevelMap<'_>>,
+) -> Result<StdResult<PrimaryKey<'a>, UserError>> {
+    match index.primary_key(rtxn)? {
+        Some(primary_key) => match PrimaryKey::new(primary_key, fields_ids_map) {
+            Some(primary_key) => Ok(Ok(primary_key)),
+            None => unreachable!("Why is the primary key not in the fidmap?"),
+        },
+        None => {
+            let first_document = match first_document {
+                Some(document) => document,
+                None => return Ok(Err(UserError::NoPrimaryKeyCandidateFound)),
+            };
+
+            let mut guesses: Vec<&str> = first_document
+                .keys()
+                .map(AsRef::as_ref)
+                .filter(|name| name.to_lowercase().ends_with(DEFAULT_PRIMARY_KEY))
+                .collect();
+
+            // sort the keys in lexicographical order, so that fields are always in the same order.
+            guesses.sort_unstable();
+
+            match guesses.as_slice() {
+                [] => Ok(Err(UserError::NoPrimaryKeyCandidateFound)),
+                [name] => {
+                    tracing::info!("Primary key was not specified in index. Inferred to '{name}'");
+                    match fields_ids_map.insert(name) {
+                        Some(field_id) => Ok(Ok(PrimaryKey::Flat { name, field_id })),
+                        None => Ok(Err(UserError::AttributeLimitReached)),
+                    }
+                }
+                multiple => Ok(Err(UserError::MultiplePrimaryKeyCandidatesFound {
+                    candidates: multiple.iter().map(|candidate| candidate.to_string()).collect(),
+                })),
+            }
+        }
+    }
+}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Clément Renault	bdd363dd94	Add spans	2024-09-26 17:20:32 +02:00
Clément Renault	d6b3aae8a6	WIP add more logs	2024-09-26 16:37:38 +02:00
Clément Renault	ac2d54b27c	Make the merger multithreaded	2024-09-26 11:09:06 +02:00
Clément Renault	7d61697f19	Fix another iteration bug on hashmap entries	2024-09-25 22:42:41 +02:00
Clément Renault	97d2860998	Fix iterating on hashmap entries	2024-09-25 22:15:15 +02:00
Clément Renault	15bf556291	Write the inverted indexes in memory and never on disk	2024-09-25 18:13:19 +02:00
Clément Renault	3d244451df	Reduce the lru key size from 8 to 12 bytes	2024-09-25 16:14:13 +02:00
Clément Renault	5f53935c8a	Fix a bug in the Lru	2024-09-25 16:09:34 +02:00
Clément Renault	29a7623c3f	Fxi some logs	2024-09-25 15:57:50 +02:00
Clément Renault	e97041f7d0	Replace the Lru free list by a simple increment	2024-09-25 15:55:52 +02:00
Clément Renault	52d7f3ed1c	Reduce the lru key size from 20 to 8 bytes	2024-09-25 15:37:13 +02:00
Clément Renault	86d5e6d9ff	Use the new Lru	2024-09-25 14:54:56 +02:00
Clément Renault	759b9b1546	Introduce a new custom Lru	2024-09-25 14:49:12 +02:00
ManyTheFish	3f7a500f3b	Build prefix fst	2024-09-25 14:36:06 +02:00
ManyTheFish	974272f2e9	Merge branch 'main' into indexer-edition-2024	2024-09-25 07:41:16 +02:00
Clément Renault	7ad037841f	Move the tracing info to eprintln	2024-09-24 18:21:58 +02:00
Clément Renault	e0c7067355	Expose an IndexedParallelIterator to the index function	2024-09-24 17:24:59 +02:00
ManyTheFish	6e87332410	Change the way the FST is built	2024-09-24 16:28:31 +02:00
Clément Renault	2d1caf27df	Use eprintln to log	2024-09-24 15:59:50 +02:00
Clément Renault	92678383d6	Update charabia	2024-09-24 15:37:56 +02:00
Clément Renault	7f148c127c	Measure the SmallVec efficacity	2024-09-24 15:32:15 +02:00
Clément Renault	4ce5d3d66d	Do not check before pushing in bitmaps	2024-09-24 09:43:16 +02:00
Clément Renault	ff931edb55	Update roaring to inline max calls	2024-09-23 16:53:42 +02:00
Clément Renault	42b093687d	Introduce the new PushOptimizedBitmap	2024-09-23 16:38:21 +02:00
Clément Renault	835c5f98f9	Remove the debug symbols	2024-09-23 15:49:24 +02:00
Clément Renault	f00664247d	Add more stats about the channel message sent	2024-09-23 15:13:52 +02:00
Clément Renault	3c63d4a1e5	Fix charabia Zho	2024-09-23 14:50:17 +02:00
Clément Renault	4551abf6d4	Update roaring to the latest version	2024-09-23 14:35:33 +02:00
Clément Renault	193d7f5d34	Add the mutualized charabia normalization	2024-09-23 14:24:25 +02:00
Clément Renault	013acb3d93	Measure merger writer channel contention	2024-09-23 11:07:59 +02:00
Clément Renault	f4ab1f168e	Prefer using Rc<str> than String when cloning a lot	2024-09-16 15:41:29 +02:00
ManyTheFish	1a0e962299	Replace hashmap by vectors in wpp	2024-09-16 15:01:20 +02:00
ManyTheFish	f13e076b8a	Use hashmap instead of Btree in wpp extractor	2024-09-16 14:40:40 +02:00
ManyTheFish	7ba49b849e	Extract and write facet databases	2024-09-16 09:35:16 +02:00
Clément Renault	f7652186e1	WIP geo fields	2024-09-12 18:01:02 +02:00
Clément Renault	b2f4e67c9a	Do not store useless updates	2024-09-12 15:38:31 +02:00
Clément Renault	ff5d3b59f5	Move the document id extraction to the primary key code	2024-09-12 12:01:42 +02:00
ManyTheFish	aa69308e45	Use a bufWriter to build word FSTs	2024-09-12 11:48:00 +02:00
ManyTheFish	eb9a20ff0b	Fix fid_word_docids extraction	2024-09-12 11:08:18 +02:00
Clément Renault	0d868f36d7	Make sure we always use a BufWriter to write the update files	2024-09-11 18:38:04 +02:00
Clément Renault	e7d9db078f	Use the right key name when convertir from CSV to NDJSON	2024-09-11 18:27:00 +02:00
Clément Renault	3e9198ebaa	Support guessing primary key again	2024-09-11 17:25:40 +02:00
Clément Renault	2a0ad0982f	Fix the document counter	2024-09-11 15:59:36 +02:00
ManyTheFish	2b317c681b	Build mergers in parallel	2024-09-11 11:49:26 +02:00
ManyTheFish	39b5990f64	Mutualize tokenization	2024-09-11 10:22:38 +02:00
Clément Renault	3848adf5a2	Improve error management and simplify JSON read	2024-09-11 10:10:51 +02:00
Clément Renault	b4de06259e	Better CSV support	2024-09-11 10:02:00 +02:00
Clément Renault	8287c2644f	Support CSV again	2024-09-10 21:10:28 +01:00
Clément Renault	c1c44a0b81	Impl serialize on TopLevelMap	2024-09-10 19:32:03 +01:00
Clément Renault	04596f3616	Move the TopLevelMap into a dedicated module	2024-09-10 18:01:17 +01:00
Clément Renault	24cb5839ad	Move the document changes sorting logic to a new trait	2024-09-10 17:37:52 +01:00
Clément Renault	8d97b7b28c	Support JSON payloads again (not perfectly though)	2024-09-10 17:09:49 +01:00
ManyTheFish	f69688e8f7	Fix several warnings in extractors and remove unreachable macros	2024-09-09 14:52:50 +02:00
Clément Renault	8fd0afaaaa	Make sure we iterate over the payload documents in order	2024-09-06 08:09:08 +02:00
Clément Renault	72c6a21a30	Use raw JSON to read the payloads	2024-09-05 20:08:23 +02:00
Clément Renault	8412be4a7d	Cleanup CowStr and TopLevelMap struct	2024-09-05 18:32:55 +02:00
Louis Dureuil	10f09c531f	add some commented code to read from json with raw values	2024-09-05 18:22:16 +02:00
ManyTheFish	8fd99b111b	Add tracing timers logs	2024-09-05 18:00:22 +02:00
Clément Renault	f6b3d1f9a5	Increase some channel sizes	2024-09-05 15:12:07 +02:00
Clément Renault	73ce67862d	Use the word pair proximity and fid word count docids extractors Co-authored-by: ManyTheFish <many@meilisearch.com>	2024-09-05 10:56:22 +02:00
Clément Renault	0fc02f7351	Move the facet extraction to dedicated modules	2024-09-05 10:32:27 +02:00
ManyTheFish	34f11e3380	Implement word count and word pair proximity extractors	2024-09-05 10:30:39 +02:00
Clément Renault	27308eaab1	Import the facet extractors	2024-09-04 17:58:15 +02:00
Clément Renault	b33ec9ba3f	Introduce the FieldIdFacetIsNullDocidsExtractor	2024-09-04 17:50:08 +02:00
Clément Renault	9c0a1cd9fd	Introduce the FieldIdFacetExistsDocidsExtractor	2024-09-04 17:48:49 +02:00
Clément Renault	0b061f1e70	Introduce the FieldIdFacetIsEmptyDocidsExtractor	2024-09-04 17:40:24 +02:00
Clément Renault	19d937ab21	Introduce the facet extractors	2024-09-04 17:03:54 +02:00
Clément Renault	1d59c19cd2	Send the WordsFst by using an Mmap	2024-09-04 14:30:09 +02:00
Clément Renault	98e48371c3	Factorize some stuff	2024-09-04 12:17:13 +02:00
Clément Renault	6d74fb0229	Introduce the WordFidWordDocids database	2024-09-04 11:40:55 +02:00
ManyTheFish	1eb75a1040	remove milli/src/update/new/extract/tokenize_document.rs	2024-09-04 11:40:26 +02:00
Clément Renault	3b82d8b5b9	Fix the cache to serialize entries correctly	2024-09-04 10:55:36 +02:00
ManyTheFish	781a186f75	remove milli/src/update/new/extract/extract_word_docids.rs	2024-09-04 10:28:31 +02:00
ManyTheFish	6a399556b5	Implement more searchable extractor	2024-09-04 10:20:18 +02:00
Clément Renault	27b4cab857	Extract and write the documents and words fst in the database	2024-09-04 09:59:19 +02:00
Clément Renault	52d32b4ee9	Move the channel sender in the closure to stop the merger thread	2024-09-03 16:08:33 +02:00
ManyTheFish	da61408e52	Remove unimplemented from document changes	2024-09-03 15:14:16 +02:00
ManyTheFish	fe69385bd7	Fix tokenizer test	2024-09-03 14:24:37 +02:00
Clément Renault	c1557734dc	Use the GlobalFieldsIdsMap everywhere and write it to disk Co-authored-by: Dureuill <louis@meilisearch.com> Co-authored-by: ManyTheFish <many@meilisearch.com>	2024-09-03 12:01:01 +02:00
ManyTheFish	c50d3edc4a	Integrate first searchable exctrator	2024-09-03 11:02:39 +02:00
Clément Renault	5369bf4a62	Change some lifetimes	2024-09-02 19:51:22 +02:00
Clément Renault	bcb1aa3d22	Find a temporary solution to par into iter on an HashMap Spoiler: Do not use an HashMap but drain it into a Vec	2024-09-02 19:39:48 +02:00
Clément Renault	9b7858fb90	Expose the new indexer	2024-09-02 15:21:59 +02:00
Clément Renault	ab01679a8f	Remove the useless option from the document changes	2024-09-02 15:21:00 +02:00
Clément Renault	521775f788	I push for Many	2024-09-02 15:10:21 +02:00
Clément Renault	72e7b7846e	Renaming the indexers	2024-09-02 14:42:27 +02:00
Clément Renault	6526ce1208	Fix the merging of documents	2024-09-02 14:41:20 +02:00
Clément Renault	e639ec79d1	Move the indexers into their own modules	2024-09-02 10:42:19 +02:00
Clément Renault	bb885a5810	Fix the merge for roaring bitmap	2024-09-01 23:20:19 +02:00
Clément Renault	b625d31c7d	Introduce the PartialDumpIndexer indexer that generates document ids in parallel	2024-08-30 15:07:21 +02:00
Clément Renault	6487a67f2b	Introduce the ConcurrentAvailableIds struct and rename the other to AvailableIds	2024-08-30 15:06:50 +02:00
Clément Renault	271ce91b3b	Add the rayon Threadpool to the index function parameter	2024-08-30 14:34:24 +02:00
Clément Renault	54f2eb4507	Remove duplication of grenad merger	2024-08-30 14:34:05 +02:00
Clément Renault	794ebcd582	Replace grenad with the new grenad various-improvement branch	2024-08-30 11:53:59 +02:00
Clément Renault	b7c77c7a39	Use the latest version of the obkv crate	2024-08-30 11:53:59 +02:00
Clément Renault	0c57cf7565	Replace obkv with the temporary new version of it	2024-08-30 11:53:58 +02:00
Clément Renault	27df9e6c73	Introduce the indexer::index function that runs the indexation	2024-08-30 11:53:58 +02:00
Clément Renault	45c060831e	Introduce typed channels and the merger loop	2024-08-30 11:53:58 +02:00
Clément Renault	874c1ac538	First channels types	2024-08-30 11:53:58 +02:00
Clément Renault	e6ffa4d454	Implement the document merge function for the replace method	2024-08-30 11:53:58 +02:00
Clément Renault	637a9c8bdd	Implement the document merge function for the update method	2024-08-30 11:53:58 +02:00
Louis Dureuil	c683fa98e6	WIP Co-authored-by: Kerollmops <clement@meilisearch.com> Co-authored-by: ManyTheFish <many@meilisearch.com>	2024-08-30 11:53:57 +02:00