Remove the unwraps, asserts and return actual io errors

Update Charabia v0.9.8

Cargo.lock

@@ -589,7 +589,7 @@ source = "git+https://github.com/meilisearch/bbqueue#cbb87cc707b5af415ef203bdaf2
 
 [[package]]
 name = "benchmarks"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "anyhow",
  "bumpalo",
@@ -799,7 +799,7 @@ dependencies = [
 
 [[package]]
 name = "build-info"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "anyhow",
  "time",
@@ -1128,9 +1128,9 @@ dependencies = [
 
 [[package]]
 name = "charabia"
-version = "0.9.7"
+version = "0.9.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7c2f456825b7f15eac01a1cae40c12c3f55e931d4327e6e4fa59508d664e9568"
+checksum = "bbdc8cd8f999e8b8b13ed71d30962bbf98cf39e2f2a9f1ae1ba354199239d66e"
 dependencies = [
  "aho-corasick",
  "csv",
@@ -1139,7 +1139,6 @@ dependencies = [
  "irg-kvariants",
  "jieba-rs",
  "lindera",
- "once_cell",
  "pinyin",
  "serde",
  "slice-group-by",
@@ -1824,12 +1823,12 @@ version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "415b6ec780d34dcf624666747194393603d0373b7141eef01d12ee58881507d9"
 dependencies = [
- "phf",
+ "phf 0.11.3",
 ]
 
 [[package]]
 name = "dump"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "anyhow",
  "big_s",
@@ -2009,6 +2008,16 @@ dependencies = [
  "syn 2.0.106",
 ]
 
+[[package]]
+name = "env_logger"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a19187fea3ac7e84da7dacf48de0c45d63c6a76f9490dae389aead16c243fce3"
+dependencies = [
+ "log",
+ "regex",
+]
+
 [[package]]
 name = "equivalent"
 version = "1.0.2"
@@ -2072,7 +2081,7 @@ checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
 
 [[package]]
 name = "file-store"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "tempfile",
  "thiserror 2.0.16",
@@ -2094,7 +2103,7 @@ dependencies = [
 
 [[package]]
 name = "filter-parser"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "insta",
  "levenshtein_automata",
@@ -2122,7 +2131,7 @@ dependencies = [
 
 [[package]]
 name = "flatten-serde-json"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "criterion",
  "serde_json",
@@ -2279,7 +2288,7 @@ dependencies = [
 
 [[package]]
 name = "fuzzers"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "arbitrary",
  "bumpalo",
@@ -2758,9 +2767,9 @@ dependencies = [
 
 [[package]]
 name = "hannoy"
-version = "0.0.9-nested-rtxns"
+version = "0.0.9-nested-rtxns-2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cc5a945b92b063e677d658cfcc7cb6dec2502fe44631f017084938f14d6ce30e"
+checksum = "06eda090938d9dcd568c8c2a5de383047ed9191578ebf4a342d2975d16e621f2"
 dependencies = [
  "bytemuck",
  "byteorder",
@@ -2838,9 +2847,9 @@ checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
 
 [[package]]
 name = "heed"
-version = "0.22.1-nested-rtxns"
+version = "0.22.1-nested-rtxns-6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ff115ba5712b1f1fc7617b195f5c2f139e29c397ff79da040cd19db75ccc240"
+checksum = "c69e07cd539834bedcfa938f3d7d8520cce1ad2b0776c122b5ccdf8fd5bafe12"
 dependencies = [
  "bitflags 2.9.4",
  "byteorder",
@@ -3233,7 +3242,7 @@ dependencies = [
 
 [[package]]
 name = "index-scheduler"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "anyhow",
  "backoff",
@@ -3242,6 +3251,7 @@ dependencies = [
  "bumpalo",
  "bumparaw-collections",
  "byte-unit",
+ "bytes",
  "convert_case 0.8.0",
  "crossbeam-channel",
  "csv",
@@ -3259,13 +3269,17 @@ dependencies = [
  "memmap2",
  "page_size",
  "rayon",
+ "reqwest",
  "roaring 0.10.12",
+ "rusty-s3",
  "serde",
  "serde_json",
  "synchronoise",
+ "tar",
  "tempfile",
  "thiserror 2.0.16",
  "time",
+ "tokio",
  "tracing",
  "ureq",
  "uuid",
@@ -3443,26 +3457,49 @@ checksum = "4a5f13b858c8d314ee3e8f639011f7ccefe71f97f96e50151fb991f267928e2c"
 
 [[package]]
 name = "jieba-macros"
-version = "0.7.1"
+version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7c676b32a471d3cfae8dac2ad2f8334cd52e53377733cca8c1fb0a5062fec192"
+checksum = "348294e44ee7e3c42685da656490f8febc7359632544019621588902216da95c"
 dependencies = [
  "phf_codegen",
 ]
 
 [[package]]
 name = "jieba-rs"
-version = "0.7.4"
+version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f5dd552bbb95d578520ee68403bf8aaf0dbbb2ce55b0854d019f9350ad61040a"
+checksum = "766bd7012aa5ba49411ebdf4e93bddd59b182d2918e085d58dec5bb9b54b7105"
 dependencies = [
  "cedarwood",
- "fxhash",
  "include-flate",
  "jieba-macros",
- "lazy_static",
- "phf",
+ "phf 0.13.1",
  "regex",
+ "rustc-hash 2.1.1",
+]
+
+[[package]]
+name = "jiff"
+version = "0.2.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "be1f93b8b1eb69c77f24bbb0afdf66f54b632ee39af40ca21c4365a1d7347e49"
+dependencies = [
+ "jiff-static",
+ "log",
+ "portable-atomic",
+ "portable-atomic-util",
+ "serde",
+]
+
+[[package]]
+name = "jiff-static"
+version = "0.2.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "03343451ff899767262ec32146f6d559dd759fdadf42ff0e227c7c48f72594b4"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.106",
 ]
 
 [[package]]
@@ -3487,7 +3524,7 @@ dependencies = [
 
 [[package]]
 name = "json-depth-checker"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "criterion",
  "serde_json",
@@ -3888,9 +3925,9 @@ checksum = "241eaef5fd12c88705a01fc1066c48c4b36e0dd4377dcdc7ec3942cea7a69956"
 
 [[package]]
 name = "lmdb-master-sys"
-version = "0.2.6-nested-rtxns"
+version = "0.2.6-nested-rtxns-6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f4ff85130e3c994b36877045fbbb138d521dea7197bfc19dc3d5d95101a8e20a"
+checksum = "e113d9bf240f974fbe7fd516cbfd8c422e925c0655495501c7237548425493d0"
 dependencies = [
  "cc",
  "doxygen-rs",
@@ -3988,6 +4025,16 @@ version = "1.0.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3e2e65a1a2e43cfcb47a895c4c8b10d1f4a61097f9f254f183aee60cad9c651d"
 
+[[package]]
+name = "md-5"
+version = "0.10.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d89e7ee0cfbedfc4da3340218492196241d89eefb6dab27de5df917a6d2e78cf"
+dependencies = [
+ "cfg-if",
+ "digest",
+]
+
 [[package]]
 name = "md5"
 version = "0.7.0"
@@ -3996,7 +4043,7 @@ checksum = "490cc448043f947bae3cbee9c203358d62dbee0db12107a74be5c30ccfd09771"
 
 [[package]]
 name = "meili-snap"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "insta",
  "md5",
@@ -4007,7 +4054,7 @@ dependencies = [
 
 [[package]]
 name = "meilisearch"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "actix-cors",
  "actix-http",
@@ -4104,7 +4151,7 @@ dependencies = [
 
 [[package]]
 name = "meilisearch-auth"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "base64 0.22.1",
  "enum-iterator",
@@ -4123,7 +4170,7 @@ dependencies = [
 
 [[package]]
 name = "meilisearch-types"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "actix-web",
  "anyhow",
@@ -4158,7 +4205,7 @@ dependencies = [
 
 [[package]]
 name = "meilitool"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "anyhow",
  "clap",
@@ -4192,7 +4239,7 @@ dependencies = [
 
 [[package]]
 name = "milli"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "allocator-api2 0.3.1",
  "arroy",
@@ -4200,6 +4247,7 @@ dependencies = [
  "big_s",
  "bimap",
  "bincode 1.3.3",
+ "bitpacking",
  "bstr",
  "bumpalo",
  "bumparaw-collections",
@@ -4246,6 +4294,7 @@ dependencies = [
  "obkv",
  "once_cell",
  "ordered-float 5.0.0",
+ "quickcheck",
  "rand 0.8.5",
  "rayon",
  "rhai",
@@ -4773,7 +4822,7 @@ checksum = "9b4f627cb1b25917193a259e49bdad08f671f8d9708acfd5fe0a8c1455d87220"
 
 [[package]]
 name = "permissive-json-pointer"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "big_s",
  "serde_json",
@@ -4830,17 +4879,27 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1fd6780a80ae0c52cc120a26a1a42c1ae51b247a253e4e06113d23d2c2edd078"
 dependencies = [
  "phf_macros",
- "phf_shared",
+ "phf_shared 0.11.3",
+]
+
+[[package]]
+name = "phf"
+version = "0.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c1562dc717473dbaa4c1f85a36410e03c047b2e7df7f45ee938fbef64ae7fadf"
+dependencies = [
+ "phf_shared 0.13.1",
+ "serde",
 ]
 
 [[package]]
 name = "phf_codegen"
-version = "0.11.3"
+version = "0.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "aef8048c789fa5e851558d709946d6d79a8ff88c0440c587967f8e94bfb1216a"
+checksum = "49aa7f9d80421bca176ca8dbfebe668cc7a2684708594ec9f3c0db0805d5d6e1"
 dependencies = [
- "phf_generator",
- "phf_shared",
+ "phf_generator 0.13.1",
+ "phf_shared 0.13.1",
 ]
 
 [[package]]
@@ -4849,18 +4908,28 @@ version = "0.11.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3c80231409c20246a13fddb31776fb942c38553c51e871f8cbd687a4cfb5843d"
 dependencies = [
- "phf_shared",
+ "phf_shared 0.11.3",
  "rand 0.8.5",
 ]
 
+[[package]]
+name = "phf_generator"
+version = "0.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "135ace3a761e564ec88c03a77317a7c6b80bb7f7135ef2544dbe054243b89737"
+dependencies = [
+ "fastrand",
+ "phf_shared 0.13.1",
+]
+
 [[package]]
 name = "phf_macros"
 version = "0.11.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f84ac04429c13a7ff43785d75ad27569f2951ce0ffd30a3321230db2fc727216"
 dependencies = [
- "phf_generator",
- "phf_shared",
+ "phf_generator 0.11.3",
+ "phf_shared 0.11.3",
  "proc-macro2",
  "quote",
  "syn 2.0.106",
@@ -4875,6 +4944,15 @@ dependencies = [
  "siphasher",
 ]
 
+[[package]]
+name = "phf_shared"
+version = "0.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e57fef6bc5981e38c2ce2d63bfa546861309f875b8a75f092d1d54ae2d64f266"
+dependencies = [
+ "siphasher",
+]
+
 [[package]]
 name = "pin-project"
 version = "1.1.10"
@@ -4962,6 +5040,15 @@ version = "1.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f84267b20a16ea918e43c6a88433c2d54fa145c92a811b5b047ccbe153674483"
 
+[[package]]
+name = "portable-atomic-util"
+version = "0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d8a2f0d8d040d7848a709caf78912debcc3f33ee4b3cac47d73d1e1069e83507"
+dependencies = [
+ "portable-atomic",
+]
+
 [[package]]
 name = "potential_utf"
 version = "0.1.3"
@@ -5139,6 +5226,27 @@ dependencies = [
  "version_check",
 ]
 
+[[package]]
+name = "quick-xml"
+version = "0.38.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "42a232e7487fc2ef313d96dde7948e7a3c05101870d8985e4fd8d26aedd27b89"
+dependencies = [
+ "memchr",
+ "serde",
+]
+
+[[package]]
+name = "quickcheck"
+version = "1.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "588f6378e4dd99458b60ec275b4477add41ce4fa9f64dcba6f15adccb19b50d6"
+dependencies = [
+ "env_logger",
+ "log",
+ "rand 0.8.5",
+]
+
 [[package]]
 name = "quinn"
 version = "0.11.9"
@@ -5414,6 +5522,7 @@ dependencies = [
  "futures-channel",
  "futures-core",
  "futures-util",
+ "h2 0.4.12",
  "http 1.3.1",
  "http-body",
  "http-body-util",
@@ -5704,6 +5813,25 @@ version = "1.0.22"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b39cdef0fa800fc44525c84ccb54a029961a8215f9619753635a9c0d2538d46d"
 
+[[package]]
+name = "rusty-s3"
+version = "0.8.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fac2edd2f0b56bd79a7343f49afc01c2d41010df480538a510e0abc56044f66c"
+dependencies = [
+ "base64 0.22.1",
+ "hmac",
+ "jiff",
+ "md-5",
+ "percent-encoding",
+ "quick-xml",
+ "serde",
+ "serde_json",
+ "sha2",
+ "url",
+ "zeroize",
+]
+
 [[package]]
 name = "ryu"
 version = "1.0.20"
@@ -7820,7 +7948,7 @@ dependencies = [
 
 [[package]]
 name = "xtask"
-version = "1.23.0"
+version = "1.24.0"
 dependencies = [
  "anyhow",
  "build-info",

Cargo.toml

@@ -23,7 +23,7 @@ members = [
 ]
 
 [workspace.package]
-version = "1.23.0"
+version = "1.24.0"
 authors = [
     "Quentin de Quelen <quentin@dequelen.me>",
     "Clément Renault <clement@meilisearch.com>",

Dockerfile

@@ -1,5 +1,5 @@
 # Compile
-FROM    rust:1.89-alpine3.20 AS compiler
+FROM    rust:1.89-alpine3.22 AS compiler
 
 RUN     apk add -q --no-cache build-base openssl-dev
 
@@ -20,7 +20,7 @@ RUN     set -eux; \
         cargo build --release -p meilisearch -p meilitool
 
 # Run
-FROM    alpine:3.20
+FROM    alpine:3.22
 LABEL   org.opencontainers.image.source="https://github.com/meilisearch/meilisearch"
 
 ENV     MEILI_HTTP_ADDR 0.0.0.0:7700

README.md

@@ -39,6 +39,7 @@
 ## 🖥 Examples
 
 - [**Movies**](https://where2watch.meilisearch.com/?utm_campaign=oss&utm_source=github&utm_medium=organization) — An application to help you find streaming platforms to watch movies using [hybrid search](https://www.meilisearch.com/solutions/hybrid-search?utm_campaign=oss&utm_source=github&utm_medium=meilisearch&utm_content=demos).
+- [**Flickr**](https://flickr.meilisearch.com/?utm_campaign=oss&utm_source=github&utm_medium=organization) — Search and explore one hundred million Flickr images with semantic search.
 - [**Ecommerce**](https://ecommerce.meilisearch.com/?utm_campaign=oss&utm_source=github&utm_medium=meilisearch&utm_content=demos) — Ecommerce website using disjunctive [facets](https://www.meilisearch.com/docs/learn/fine_tuning_results/faceted_search?utm_campaign=oss&utm_source=github&utm_medium=meilisearch&utm_content=demos), range and rating filtering, and pagination.
 - [**Songs**](https://music.meilisearch.com/?utm_campaign=oss&utm_source=github&utm_medium=meilisearch&utm_content=demos) — Search through 47 million of songs.
 - [**SaaS**](https://saas.meilisearch.com/?utm_campaign=oss&utm_source=github&utm_medium=meilisearch&utm_content=demos) — Search for contacts, deals, and companies in this [multi-tenant](https://www.meilisearch.com/docs/learn/security/multitenancy_tenant_tokens?utm_campaign=oss&utm_source=github&utm_medium=meilisearch&utm_content=demos) CRM application.

crates/file-store/src/lib.rs

@@ -60,7 +60,7 @@ impl FileStore {
 
     /// Returns the file corresponding to the requested uuid.
     pub fn get_update(&self, uuid: Uuid) -> Result<StdFile> {
-        let path = self.get_update_path(uuid);
+        let path = self.update_path(uuid);
         let file = match StdFile::open(path) {
             Ok(file) => file,
             Err(e) => {
@@ -72,7 +72,7 @@ impl FileStore {
     }
 
     /// Returns the path that correspond to this uuid, the path could not exists.
-    pub fn get_update_path(&self, uuid: Uuid) -> PathBuf {
+    pub fn update_path(&self, uuid: Uuid) -> PathBuf {
         self.path.join(uuid.to_string())
     }
 

crates/index-scheduler/Cargo.toml

@@ -14,6 +14,7 @@ license.workspace = true
 anyhow = "1.0.98"
 bincode = "1.3.3"
 byte-unit = "5.1.6"
+bytes = "1.10.1"
 bumpalo = "3.18.1"
 bumparaw-collections = "0.1.4"
 convert_case = "0.8.0"
@@ -32,6 +33,7 @@ rayon = "1.10.0"
 roaring = { version = "0.10.12", features = ["serde"] }
 serde = { version = "1.0.219", features = ["derive"] }
 serde_json = { version = "1.0.140", features = ["preserve_order"] }
+tar = "0.4.44"
 synchronoise = "1.0.1"
 tempfile = "3.20.0"
 thiserror = "2.0.12"
@@ -45,6 +47,9 @@ tracing = "0.1.41"
 ureq = "2.12.1"
 uuid = { version = "1.17.0", features = ["serde", "v4"] }
 backoff = "0.4.0"
+reqwest = { version = "0.12.23", features = ["rustls-tls", "http2"], default-features = false }
+rusty-s3 = "0.8.1"
+tokio = { version = "1.47.1", features = ["full"] }
 
 [dev-dependencies]
 big_s = "1.0.2"

crates/index-scheduler/src/error.rs

@@ -5,6 +5,7 @@ use meilisearch_types::error::{Code, ErrorCode};
 use meilisearch_types::milli::index::RollbackOutcome;
 use meilisearch_types::tasks::{Kind, Status};
 use meilisearch_types::{heed, milli};
+use reqwest::StatusCode;
 use thiserror::Error;
 
 use crate::TaskId;
@@ -127,6 +128,14 @@ pub enum Error {
     #[error("Aborted task")]
     AbortedTask,
 
+    #[error("S3 error: status: {status}, body: {body}")]
+    S3Error { status: StatusCode, body: String },
+    #[error("S3 HTTP error: {0}")]
+    S3HttpError(reqwest::Error),
+    #[error("S3 XML error: {0}")]
+    S3XmlError(Box<dyn std::error::Error + Send + Sync>),
+    #[error("S3 bucket error: {0}")]
+    S3BucketError(rusty_s3::BucketError),
     #[error(transparent)]
     Dump(#[from] dump::Error),
     #[error(transparent)]
@@ -226,6 +235,10 @@ impl Error {
             | Error::TaskCancelationWithEmptyQuery
             | Error::FromRemoteWhenExporting { .. }
             | Error::AbortedTask
+            | Error::S3Error { .. }
+            | Error::S3HttpError(_)
+            | Error::S3XmlError(_)
+            | Error::S3BucketError(_)
             | Error::Dump(_)
             | Error::Heed(_)
             | Error::Milli { .. }
@@ -293,8 +306,14 @@ impl ErrorCode for Error {
             Error::BatchNotFound(_) => Code::BatchNotFound,
             Error::TaskDeletionWithEmptyQuery => Code::MissingTaskFilters,
             Error::TaskCancelationWithEmptyQuery => Code::MissingTaskFilters,
-            // TODO: not sure of the Code to use
             Error::NoSpaceLeftInTaskQueue => Code::NoSpaceLeftOnDevice,
+            Error::S3Error { status, .. } if status.is_client_error() => {
+                Code::InvalidS3SnapshotRequest
+            }
+            Error::S3Error { .. } => Code::S3SnapshotServerError,
+            Error::S3HttpError(_) => Code::S3SnapshotServerError,
+            Error::S3XmlError(_) => Code::S3SnapshotServerError,
+            Error::S3BucketError(_) => Code::InvalidS3SnapshotParameters,
             Error::Dump(e) => e.error_code(),
             Error::Milli { error, .. } => error.error_code(),
             Error::ProcessBatchPanicked(_) => Code::Internal,

crates/index-scheduler/src/index_mapper/mod.rs

@@ -199,7 +199,7 @@ impl IndexMapper {
                 let uuid = Uuid::new_v4();
                 self.index_mapping.put(&mut wtxn, name, &uuid)?;
 
-                let index_path = self.base_path.join(uuid.to_string());
+                let index_path = self.index_path(uuid);
                 fs::create_dir_all(&index_path)?;
 
                 // Error if the UUIDv4 somehow already exists in the map, since it should be fresh.
@@ -286,7 +286,7 @@ impl IndexMapper {
         };
 
         let index_map = self.index_map.clone();
-        let index_path = self.base_path.join(uuid.to_string());
+        let index_path = self.index_path(uuid);
         let index_name = name.to_string();
         thread::Builder::new()
             .name(String::from("index_deleter"))
@@ -408,7 +408,7 @@ impl IndexMapper {
                     } else {
                         continue;
                     };
-                    let index_path = self.base_path.join(uuid.to_string());
+                    let index_path = self.index_path(uuid);
                     // take the lock to reopen the environment.
                     reopen
                         .reopen(&mut self.index_map.write().unwrap(), &index_path)
@@ -425,7 +425,7 @@ impl IndexMapper {
                     // if it's not already there.
                     match index_map.get(&uuid) {
                         Missing => {
-                            let index_path = self.base_path.join(uuid.to_string());
+                            let index_path = self.index_path(uuid);
 
                             break index_map
                                 .create(
@@ -452,6 +452,14 @@ impl IndexMapper {
         Ok(index)
     }
 
+    /// Returns the path of the index.
+    ///
+    /// The folder located at this path is containing the data.mdb,
+    /// the lock.mdb and an optional data.mdb.cpy file.
+    pub fn index_path(&self, uuid: Uuid) -> PathBuf {
+        self.base_path.join(uuid.to_string())
+    }
+
     pub fn rollback_index(
         &self,
         rtxn: &RoTxn,
@@ -492,7 +500,7 @@ impl IndexMapper {
             };
         }
 
-        let index_path = self.base_path.join(uuid.to_string());
+        let index_path = self.index_path(uuid);
         Index::rollback(milli::heed::EnvOpenOptions::new().read_txn_without_tls(), index_path, to)
             .map_err(|err| crate::Error::from_milli(err, Some(name.to_string())))
     }

crates/index-scheduler/src/insta_snapshot.rs

@@ -36,6 +36,7 @@ pub fn snapshot_index_scheduler(scheduler: &IndexScheduler) -> String {
         run_loop_iteration: _,
         embedders: _,
         chat_settings: _,
+        runtime: _,
     } = scheduler;
 
     let rtxn = env.read_txn().unwrap();

crates/index-scheduler/src/lib.rs

@@ -216,6 +216,9 @@ pub struct IndexScheduler {
     /// A counter that is incremented before every call to [`tick`](IndexScheduler::tick)
     #[cfg(test)]
     run_loop_iteration: Arc<RwLock<usize>>,
+
+    /// The tokio runtime used for asynchronous tasks.
+    runtime: Option<tokio::runtime::Handle>,
 }
 
 impl IndexScheduler {
@@ -242,6 +245,7 @@ impl IndexScheduler {
             run_loop_iteration: self.run_loop_iteration.clone(),
             features: self.features.clone(),
             chat_settings: self.chat_settings,
+            runtime: self.runtime.clone(),
         }
     }
 
@@ -255,13 +259,23 @@ impl IndexScheduler {
     }
 
     /// Create an index scheduler and start its run loop.
-    #[allow(private_interfaces)] // because test_utils is private
     pub fn new(
         options: IndexSchedulerOptions,
         auth_env: Env<WithoutTls>,
         from_db_version: (u32, u32, u32),
-        #[cfg(test)] test_breakpoint_sdr: crossbeam_channel::Sender<(test_utils::Breakpoint, bool)>,
-        #[cfg(test)] planned_failures: Vec<(usize, test_utils::FailureLocation)>,
+        runtime: Option<tokio::runtime::Handle>,
+    ) -> Result<Self> {
+        let this = Self::new_without_run(options, auth_env, from_db_version, runtime)?;
+
+        this.run();
+        Ok(this)
+    }
+
+    fn new_without_run(
+        options: IndexSchedulerOptions,
+        auth_env: Env<WithoutTls>,
+        from_db_version: (u32, u32, u32),
+        runtime: Option<tokio::runtime::Handle>,
     ) -> Result<Self> {
         std::fs::create_dir_all(&options.tasks_path)?;
         std::fs::create_dir_all(&options.update_file_path)?;
@@ -316,8 +330,7 @@ impl IndexScheduler {
 
         wtxn.commit()?;
 
-        // allow unreachable_code to get rids of the warning in the case of a test build.
-        let this = Self {
+        Ok(Self {
             processing_tasks: Arc::new(RwLock::new(ProcessingTasks::new())),
             version,
             queue,
@@ -333,15 +346,32 @@ impl IndexScheduler {
             webhooks: Arc::new(webhooks),
             embedders: Default::default(),
 
-            #[cfg(test)]
-            test_breakpoint_sdr,
-            #[cfg(test)]
-            planned_failures,
+            #[cfg(test)] // Will be replaced in `new_tests` in test environments
+            test_breakpoint_sdr: crossbeam_channel::bounded(0).0,
+            #[cfg(test)] // Will be replaced in `new_tests` in test environments
+            planned_failures: Default::default(),
             #[cfg(test)]
             run_loop_iteration: Arc::new(RwLock::new(0)),
             features,
             chat_settings,
-        };
+            runtime,
+        })
+    }
+
+    /// Create an index scheduler and start its run loop.
+    #[cfg(test)]
+    fn new_test(
+        options: IndexSchedulerOptions,
+        auth_env: Env<WithoutTls>,
+        from_db_version: (u32, u32, u32),
+        runtime: Option<tokio::runtime::Handle>,
+        test_breakpoint_sdr: crossbeam_channel::Sender<(test_utils::Breakpoint, bool)>,
+        planned_failures: Vec<(usize, test_utils::FailureLocation)>,
+    ) -> Result<Self> {
+        let mut this = Self::new_without_run(options, auth_env, from_db_version, runtime)?;
+
+        this.test_breakpoint_sdr = test_breakpoint_sdr;
+        this.planned_failures = planned_failures;
 
         this.run();
         Ok(this)

crates/index-scheduler/src/processing.rs

@@ -75,6 +75,7 @@ make_enum_progress! {
     pub enum TaskCancelationProgress {
         RetrievingTasks,
         CancelingUpgrade,
+        CleaningCompactionLeftover,
         UpdatingTasks,
     }
 }

crates/index-scheduler/src/scheduler/autobatcher.rs

@@ -25,7 +25,6 @@ enum AutobatchKind {
     IndexDeletion,
     IndexUpdate,
     IndexSwap,
-    IndexCompaction,
 }
 
 impl AutobatchKind {
@@ -69,14 +68,14 @@ impl From<KindWithContent> for AutobatchKind {
             KindWithContent::IndexCreation { .. } => AutobatchKind::IndexCreation,
             KindWithContent::IndexUpdate { .. } => AutobatchKind::IndexUpdate,
             KindWithContent::IndexSwap { .. } => AutobatchKind::IndexSwap,
-            KindWithContent::IndexCompaction { .. } => AutobatchKind::IndexCompaction,
-            KindWithContent::TaskCancelation { .. }
+            KindWithContent::IndexCompaction { .. }
+            | KindWithContent::TaskCancelation { .. }
             | KindWithContent::TaskDeletion { .. }
             | KindWithContent::DumpCreation { .. }
             | KindWithContent::Export { .. }
             | KindWithContent::UpgradeDatabase { .. }
             | KindWithContent::SnapshotCreation => {
-                panic!("The autobatcher should never be called with tasks that don't apply to an index.")
+                panic!("The autobatcher should never be called with tasks with special priority or that don't apply to an index.")
             }
         }
     }
@@ -120,9 +119,6 @@ pub enum BatchKind {
     IndexSwap {
         id: TaskId,
     },
-    IndexCompaction {
-        id: TaskId,
-    },
 }
 
 impl BatchKind {
@@ -188,13 +184,6 @@ impl BatchKind {
                 )),
                 false,
             ),
-            K::IndexCompaction => (
-                Break((
-                    BatchKind::IndexCompaction { id: task_id },
-                    BatchStopReason::TaskCannotBeBatched { kind, id: task_id },
-                )),
-                false,
-            ),
             K::DocumentClear => (Continue(BatchKind::DocumentClear { ids: vec![task_id] }), false),
             K::DocumentImport { allow_index_creation, primary_key: pk }
                 if primary_key.is_none() || pk.is_none() || primary_key == pk.as_deref() =>
@@ -300,7 +289,7 @@ impl BatchKind {
 
         match (self, autobatch_kind) {
             // We don't batch any of these operations
-            (this, K::IndexCreation | K::IndexUpdate | K::IndexSwap | K::DocumentEdition | K::IndexCompaction) => {
+            (this, K::IndexCreation | K::IndexUpdate | K::IndexSwap | K::DocumentEdition) => {
                 Break((this, BatchStopReason::TaskCannotBeBatched { kind, id }))
             },
             // We must not batch tasks that don't have the same index creation rights if the index doesn't already exists.
@@ -497,7 +486,6 @@ impl BatchKind {
                 | BatchKind::IndexDeletion { .. }
                 | BatchKind::IndexUpdate { .. }
                 | BatchKind::IndexSwap { .. }
-                | BatchKind::IndexCompaction { .. }
                 | BatchKind::DocumentEdition { .. },
                 _,
             ) => {

crates/index-scheduler/src/scheduler/create_batch.rs

@@ -437,12 +437,6 @@ impl IndexScheduler {
                 current_batch.processing(Some(&mut task));
                 Ok(Some(Batch::IndexSwap { task }))
             }
-            BatchKind::IndexCompaction { id } => {
-                let mut task =
-                    self.queue.tasks.get_task(rtxn, id)?.ok_or(Error::CorruptedTaskQueue)?;
-                current_batch.processing(Some(&mut task));
-                Ok(Some(Batch::IndexCompaction { index_uid, task }))
-            }
         }
     }
 
@@ -525,17 +519,33 @@ impl IndexScheduler {
             return Ok(Some((Batch::TaskDeletions(tasks), current_batch)));
         }
 
-        // 3. we batch the export.
+        // 3. we get the next task to compact
+        let to_compact = self.queue.tasks.get_kind(rtxn, Kind::IndexCompaction)? & enqueued;
+        if let Some(task_id) = to_compact.min() {
+            let mut task =
+                self.queue.tasks.get_task(rtxn, task_id)?.ok_or(Error::CorruptedTaskQueue)?;
+            current_batch.processing(Some(&mut task));
+            current_batch.reason(BatchStopReason::TaskCannotBeBatched {
+                kind: Kind::IndexCompaction,
+                id: task_id,
+            });
+            let index_uid =
+                task.index_uid().expect("Compaction task must have an index uid").to_owned();
+            return Ok(Some((Batch::IndexCompaction { index_uid, task }, current_batch)));
+        }
+
+        // 4. we batch the export.
         let to_export = self.queue.tasks.get_kind(rtxn, Kind::Export)? & enqueued;
         if !to_export.is_empty() {
             let task_id = to_export.iter().next().expect("There must be at least one export task");
             let mut task = self.queue.tasks.get_task(rtxn, task_id)?.unwrap();
             current_batch.processing([&mut task]);
-            current_batch.reason(BatchStopReason::TaskKindCannotBeBatched { kind: Kind::Export });
+            current_batch
+                .reason(BatchStopReason::TaskCannotBeBatched { kind: Kind::Export, id: task_id });
             return Ok(Some((Batch::Export { task }, current_batch)));
         }
 
-        // 4. we batch the snapshot.
+        // 5. we batch the snapshot.
         let to_snapshot = self.queue.tasks.get_kind(rtxn, Kind::SnapshotCreation)? & enqueued;
         if !to_snapshot.is_empty() {
             let mut tasks = self.queue.tasks.get_existing_tasks(rtxn, to_snapshot)?;
@@ -545,7 +555,7 @@ impl IndexScheduler {
             return Ok(Some((Batch::SnapshotCreation(tasks), current_batch)));
         }
 
-        // 5. we batch the dumps.
+        // 6. we batch the dumps.
         let to_dump = self.queue.tasks.get_kind(rtxn, Kind::DumpCreation)? & enqueued;
         if let Some(to_dump) = to_dump.min() {
             let mut task =
@@ -558,7 +568,7 @@ impl IndexScheduler {
             return Ok(Some((Batch::Dump(task), current_batch)));
         }
 
-        // 6. We make a batch from the unprioritised tasks. Start by taking the next enqueued task.
+        // 7. We make a batch from the unprioritised tasks. Start by taking the next enqueued task.
         let task_id = if let Some(task_id) = enqueued.min() { task_id } else { return Ok(None) };
         let mut task =
             self.queue.tasks.get_task(rtxn, task_id)?.ok_or(Error::CorruptedTaskQueue)?;

crates/index-scheduler/src/scheduler/mod.rs

@@ -25,6 +25,7 @@ use convert_case::{Case, Casing as _};
 use meilisearch_types::error::ResponseError;
 use meilisearch_types::heed::{Env, WithoutTls};
 use meilisearch_types::milli;
+use meilisearch_types::milli::update::S3SnapshotOptions;
 use meilisearch_types::tasks::Status;
 use process_batch::ProcessBatchInfo;
 use rayon::current_num_threads;
@@ -87,11 +88,14 @@ pub struct Scheduler {
 
     /// Snapshot compaction status.
     pub(crate) experimental_no_snapshot_compaction: bool,
+
+    /// S3 Snapshot options.
+    pub(crate) s3_snapshot_options: Option<S3SnapshotOptions>,
 }
 
 impl Scheduler {
-    pub(crate) fn private_clone(&self) -> Scheduler {
-        Scheduler {
+    pub(crate) fn private_clone(&self) -> Self {
+        Self {
             must_stop_processing: self.must_stop_processing.clone(),
             wake_up: self.wake_up.clone(),
             autobatching_enabled: self.autobatching_enabled,
@@ -103,23 +107,52 @@ impl Scheduler {
             version_file_path: self.version_file_path.clone(),
             embedding_cache_cap: self.embedding_cache_cap,
             experimental_no_snapshot_compaction: self.experimental_no_snapshot_compaction,
+            s3_snapshot_options: self.s3_snapshot_options.clone(),
         }
     }
 
     pub fn new(options: &IndexSchedulerOptions, auth_env: Env<WithoutTls>) -> Scheduler {
+        let IndexSchedulerOptions {
+            version_file_path,
+            auth_path: _,
+            tasks_path: _,
+            update_file_path: _,
+            indexes_path: _,
+            snapshots_path,
+            dumps_path,
+            cli_webhook_url: _,
+            cli_webhook_authorization: _,
+            task_db_size: _,
+            index_base_map_size: _,
+            enable_mdb_writemap: _,
+            index_growth_amount: _,
+            index_count: _,
+            indexer_config,
+            autobatching_enabled,
+            cleanup_enabled: _,
+            max_number_of_tasks: _,
+            max_number_of_batched_tasks,
+            batched_tasks_size_limit,
+            instance_features: _,
+            auto_upgrade: _,
+            embedding_cache_cap,
+            experimental_no_snapshot_compaction,
+        } = options;
+
         Scheduler {
             must_stop_processing: MustStopProcessing::default(),
             // we want to start the loop right away in case meilisearch was ctrl+Ced while processing things
             wake_up: Arc::new(SignalEvent::auto(true)),
-            autobatching_enabled: options.autobatching_enabled,
-            max_number_of_batched_tasks: options.max_number_of_batched_tasks,
-            batched_tasks_size_limit: options.batched_tasks_size_limit,
-            dumps_path: options.dumps_path.clone(),
-            snapshots_path: options.snapshots_path.clone(),
+            autobatching_enabled: *autobatching_enabled,
+            max_number_of_batched_tasks: *max_number_of_batched_tasks,
+            batched_tasks_size_limit: *batched_tasks_size_limit,
+            dumps_path: dumps_path.clone(),
+            snapshots_path: snapshots_path.clone(),
             auth_env,
-            version_file_path: options.version_file_path.clone(),
-            embedding_cache_cap: options.embedding_cache_cap,
-            experimental_no_snapshot_compaction: options.experimental_no_snapshot_compaction,
+            version_file_path: version_file_path.clone(),
+            embedding_cache_cap: *embedding_cache_cap,
+            experimental_no_snapshot_compaction: *experimental_no_snapshot_compaction,
+            s3_snapshot_options: indexer_config.s3_snapshot_options.clone(),
         }
     }
 }

crates/index-scheduler/src/scheduler/process_batch.rs

@@ -1,5 +1,6 @@
 use std::collections::{BTreeSet, HashMap, HashSet};
-use std::io::{Seek, SeekFrom};
+use std::fs::{remove_file, File};
+use std::io::{ErrorKind, Seek, SeekFrom};
 use std::panic::{catch_unwind, AssertUnwindSafe};
 use std::sync::atomic::Ordering;
 
@@ -13,7 +14,7 @@ use meilisearch_types::tasks::{Details, IndexSwap, Kind, KindWithContent, Status
 use meilisearch_types::versioning::{VERSION_MAJOR, VERSION_MINOR, VERSION_PATCH};
 use milli::update::Settings as MilliSettings;
 use roaring::RoaringBitmap;
-use tempfile::PersistError;
+use tempfile::{PersistError, TempPath};
 use time::OffsetDateTime;
 
 use super::create_batch::Batch;
@@ -28,6 +29,9 @@ use crate::utils::{
 };
 use crate::{Error, IndexScheduler, Result, TaskId};
 
+/// The name of the copy of the data.mdb file used during compaction.
+const DATA_MDB_COPY_NAME: &str = "data.mdb.cpy";
+
 #[derive(Debug, Default)]
 pub struct ProcessBatchInfo {
     /// The write channel congestion. None when unavailable: settings update.
@@ -558,11 +562,12 @@ impl IndexScheduler {
             .set_currently_updating_index(Some((index_uid.to_string(), index.clone())));
 
         progress.update_progress(IndexCompaction::CreateTemporaryFile);
-        let pre_size = std::fs::metadata(index.path().join("data.mdb"))?.len();
-        let mut file = tempfile::Builder::new()
-            .suffix("data.")
-            .prefix(".mdb.cpy")
-            .tempfile_in(index.path())?;
+        let src_path = index.path().join("data.mdb");
+        let pre_size = std::fs::metadata(&src_path)?.len();
+
+        let dst_path = TempPath::from_path(index.path().join(DATA_MDB_COPY_NAME));
+        let file = File::create(&dst_path)?;
+        let mut file = tempfile::NamedTempFile::from_parts(file, dst_path);
 
         // 3. We copy the index data to the temporary file
         progress.update_progress(IndexCompaction::CopyAndCompactTheIndex);
@@ -574,7 +579,7 @@ impl IndexScheduler {
 
         // 4. We replace the index data file with the temporary file
         progress.update_progress(IndexCompaction::PersistTheCompactedIndex);
-        match file.persist(index.path().join("data.mdb")) {
+        match file.persist(src_path) {
             Ok(file) => file.sync_all()?,
             // TODO see if we have a _resource busy_ error and probably handle this by:
             //      1. closing the index, 2. replacing and 3. reopening it
@@ -910,9 +915,10 @@ impl IndexScheduler {
 
         let enqueued_tasks = &self.queue.tasks.get_status(rtxn, Status::Enqueued)?;
 
-        // 0. Check if any upgrade task was matched.
+        // 0. Check if any upgrade or compaction tasks were matched.
         //    If so, we cancel all the failed or enqueued upgrade tasks.
         let upgrade_tasks = &self.queue.tasks.get_kind(rtxn, Kind::UpgradeDatabase)?;
+        let compaction_tasks = &self.queue.tasks.get_kind(rtxn, Kind::IndexCompaction)?;
         let is_canceling_upgrade = !matched_tasks.is_disjoint(upgrade_tasks);
         if is_canceling_upgrade {
             let failed_tasks = self.queue.tasks.get_status(rtxn, Status::Failed)?;
@@ -977,7 +983,33 @@ impl IndexScheduler {
             }
         }
 
-        // 3. We now have a list of tasks to cancel, cancel them
+        // 3. If we are cancelling a compaction task, remove the tempfiles after incomplete compactions
+        for compaction_task in &tasks_to_cancel & compaction_tasks {
+            progress.update_progress(TaskCancelationProgress::CleaningCompactionLeftover);
+            let task = self.queue.tasks.get_task(rtxn, compaction_task)?.unwrap();
+            let Some(Details::IndexCompaction {
+                index_uid,
+                pre_compaction_size: _,
+                post_compaction_size: _,
+            }) = task.details
+            else {
+                unreachable!("wrong details for compaction task {compaction_task}")
+            };
+
+            let index_path = match self.index_mapper.index_mapping.get(rtxn, &index_uid)? {
+                Some(index_uuid) => self.index_mapper.index_path(index_uuid),
+                None => continue,
+            };
+
+            if let Err(e) = remove_file(index_path.join(DATA_MDB_COPY_NAME)) {
+                match e.kind() {
+                    ErrorKind::NotFound => (),
+                    _ => return Err(Error::IoError(e)),
+                }
+            }
+        }
+
+        // 4. We now have a list of tasks to cancel, cancel them
         let (task_progress, progress_obj) = AtomicTaskStep::new(tasks_to_cancel.len() as u32);
         progress.update_progress(progress_obj);
 

crates/index-scheduler/src/scheduler/process_snapshot_creation.rs

@@ -12,6 +12,8 @@ use crate::processing::{AtomicUpdateFileStep, SnapshotCreationProgress};
 use crate::queue::TaskQueue;
 use crate::{Error, IndexScheduler, Result};
 
+const UPDATE_FILES_DIR_NAME: &str = "update_files";
+
 /// # Safety
 ///
 /// See [`EnvOpenOptions::open`].
@@ -78,10 +80,32 @@ impl IndexScheduler {
     pub(super) fn process_snapshot(
         &self,
         progress: Progress,
-        mut tasks: Vec<Task>,
+        tasks: Vec<Task>,
     ) -> Result<Vec<Task>> {
         progress.update_progress(SnapshotCreationProgress::StartTheSnapshotCreation);
 
+        match self.scheduler.s3_snapshot_options.clone() {
+            Some(options) => {
+                #[cfg(not(unix))]
+                {
+                    let _ = options;
+                    panic!("Non-unix platform does not support S3 snapshotting");
+                }
+                #[cfg(unix)]
+                self.runtime
+                    .as_ref()
+                    .expect("Runtime not initialized")
+                    .block_on(self.process_snapshot_to_s3(progress, options, tasks))
+            }
+            None => self.process_snapshots_to_disk(progress, tasks),
+        }
+    }
+
+    fn process_snapshots_to_disk(
+        &self,
+        progress: Progress,
+        mut tasks: Vec<Task>,
+    ) -> Result<Vec<Task>, Error> {
         fs::create_dir_all(&self.scheduler.snapshots_path)?;
         let temp_snapshot_dir = tempfile::tempdir()?;
 
@@ -128,7 +152,7 @@ impl IndexScheduler {
         let rtxn = self.env.read_txn()?;
 
         // 2.4 Create the update files directory
-        let update_files_dir = temp_snapshot_dir.path().join("update_files");
+        let update_files_dir = temp_snapshot_dir.path().join(UPDATE_FILES_DIR_NAME);
         fs::create_dir_all(&update_files_dir)?;
 
         // 2.5 Only copy the update files of the enqueued tasks
@@ -140,7 +164,7 @@ impl IndexScheduler {
             let task =
                 self.queue.tasks.get_task(&rtxn, task_id)?.ok_or(Error::CorruptedTaskQueue)?;
             if let Some(content_uuid) = task.content_uuid() {
-                let src = self.queue.file_store.get_update_path(content_uuid);
+                let src = self.queue.file_store.update_path(content_uuid);
                 let dst = update_files_dir.join(content_uuid.to_string());
                 fs::copy(src, dst)?;
             }
@@ -206,4 +230,403 @@ impl IndexScheduler {
 
         Ok(tasks)
     }
+
+    #[cfg(unix)]
+    pub(super) async fn process_snapshot_to_s3(
+        &self,
+        progress: Progress,
+        opts: meilisearch_types::milli::update::S3SnapshotOptions,
+        mut tasks: Vec<Task>,
+    ) -> Result<Vec<Task>> {
+        use meilisearch_types::milli::update::S3SnapshotOptions;
+
+        let S3SnapshotOptions {
+            s3_bucket_url,
+            s3_bucket_region,
+            s3_bucket_name,
+            s3_snapshot_prefix,
+            s3_access_key,
+            s3_secret_key,
+            s3_max_in_flight_parts,
+            s3_compression_level: level,
+            s3_signature_duration,
+            s3_multipart_part_size,
+        } = opts;
+
+        let must_stop_processing = self.scheduler.must_stop_processing.clone();
+        let retry_backoff = backoff::ExponentialBackoff::default();
+        let db_name = {
+            let mut base_path = self.env.path().to_owned();
+            base_path.pop();
+            base_path.file_name().and_then(OsStr::to_str).unwrap_or("data.ms").to_string()
+        };
+
+        let (reader, writer) = std::io::pipe()?;
+        let uploader_task = tokio::spawn(multipart_stream_to_s3(
+            s3_bucket_url,
+            s3_bucket_region,
+            s3_bucket_name,
+            s3_snapshot_prefix,
+            s3_access_key,
+            s3_secret_key,
+            s3_max_in_flight_parts,
+            s3_signature_duration,
+            s3_multipart_part_size,
+            must_stop_processing,
+            retry_backoff,
+            db_name,
+            reader,
+        ));
+
+        let index_scheduler = IndexScheduler::private_clone(self);
+        let builder_task = tokio::task::spawn_blocking(move || {
+            stream_tarball_into_pipe(progress, level, writer, index_scheduler)
+        });
+
+        let (uploader_result, builder_result) = tokio::join!(uploader_task, builder_task);
+
+        // Check uploader result first to early return on task abortion.
+        // safety: JoinHandle can return an error if the task was aborted, cancelled, or panicked.
+        uploader_result.unwrap()?;
+        builder_result.unwrap()?;
+
+        for task in &mut tasks {
+            task.status = Status::Succeeded;
+        }
+
+        Ok(tasks)
+    }
+}
+
+/// Streams a tarball of the database content into a pipe.
+#[cfg(unix)]
+fn stream_tarball_into_pipe(
+    progress: Progress,
+    level: u32,
+    writer: std::io::PipeWriter,
+    index_scheduler: IndexScheduler,
+) -> std::result::Result<(), Error> {
+    use std::io::Write as _;
+    use std::path::Path;
+
+    let writer = flate2::write::GzEncoder::new(writer, flate2::Compression::new(level));
+    let mut tarball = tar::Builder::new(writer);
+
+    // 1. Snapshot the version file
+    tarball
+        .append_path_with_name(&index_scheduler.scheduler.version_file_path, VERSION_FILE_NAME)?;
+
+    // 2. Snapshot the index scheduler LMDB env
+    progress.update_progress(SnapshotCreationProgress::SnapshotTheIndexScheduler);
+    let tasks_env_file = index_scheduler.env.try_clone_inner_file()?;
+    let path = Path::new("tasks").join("data.mdb");
+    append_file_to_tarball(&mut tarball, path, tasks_env_file)?;
+
+    // 2.3 Create a read transaction on the index-scheduler
+    let rtxn = index_scheduler.env.read_txn()?;
+
+    // 2.4 Create the update files directory
+    //     And only copy the update files of the enqueued tasks
+    progress.update_progress(SnapshotCreationProgress::SnapshotTheUpdateFiles);
+    let enqueued = index_scheduler.queue.tasks.get_status(&rtxn, Status::Enqueued)?;
+    let (atomic, update_file_progress) = AtomicUpdateFileStep::new(enqueued.len() as u32);
+    progress.update_progress(update_file_progress);
+
+    // We create the update_files directory so that it
+    // always exists even if there are no update files
+    let update_files_dir = Path::new(UPDATE_FILES_DIR_NAME);
+    let src_update_files_dir = {
+        let mut path = index_scheduler.env.path().to_path_buf();
+        path.pop();
+        path.join(UPDATE_FILES_DIR_NAME)
+    };
+    tarball.append_dir(update_files_dir, src_update_files_dir)?;
+
+    for task_id in enqueued {
+        let task = index_scheduler
+            .queue
+            .tasks
+            .get_task(&rtxn, task_id)?
+            .ok_or(Error::CorruptedTaskQueue)?;
+        if let Some(content_uuid) = task.content_uuid() {
+            use std::fs::File;
+
+            let src = index_scheduler.queue.file_store.update_path(content_uuid);
+            let mut update_file = File::open(src)?;
+            let path = update_files_dir.join(content_uuid.to_string());
+            tarball.append_file(path, &mut update_file)?;
+        }
+        atomic.fetch_add(1, Ordering::Relaxed);
+    }
+
+    // 3. Snapshot every indexes
+    progress.update_progress(SnapshotCreationProgress::SnapshotTheIndexes);
+    let index_mapping = index_scheduler.index_mapper.index_mapping;
+    let nb_indexes = index_mapping.len(&rtxn)? as u32;
+    let indexes_dir = Path::new("indexes");
+    let indexes_references: Vec<_> = index_scheduler
+        .index_mapper
+        .index_mapping
+        .iter(&rtxn)?
+        .map(|res| res.map_err(Error::from).map(|(name, uuid)| (name.to_string(), uuid)))
+        .collect::<Result<_, Error>>()?;
+
+    // It's prettier to use a for loop instead of the IndexMapper::try_for_each_index
+    // method, especially when we need to access the UUID, local path and index number.
+    for (i, (name, uuid)) in indexes_references.into_iter().enumerate() {
+        progress.update_progress(VariableNameStep::<SnapshotCreationProgress>::new(
+            &name, i as u32, nb_indexes,
+        ));
+        let path = indexes_dir.join(uuid.to_string()).join("data.mdb");
+        let index = index_scheduler.index_mapper.index(&rtxn, &name)?;
+        let index_file = index.try_clone_inner_file()?;
+        tracing::trace!("Appending index file for {name} in {}", path.display());
+        append_file_to_tarball(&mut tarball, path, index_file)?;
+    }
+
+    drop(rtxn);
+
+    // 4. Snapshot the auth LMDB env
+    progress.update_progress(SnapshotCreationProgress::SnapshotTheApiKeys);
+    let auth_env_file = index_scheduler.scheduler.auth_env.try_clone_inner_file()?;
+    let path = Path::new("auth").join("data.mdb");
+    append_file_to_tarball(&mut tarball, path, auth_env_file)?;
+
+    let mut gzencoder = tarball.into_inner()?;
+    gzencoder.flush()?;
+    gzencoder.try_finish()?;
+    let mut writer = gzencoder.finish()?;
+    writer.flush()?;
+
+    Result::<_, Error>::Ok(())
+}
+
+#[cfg(unix)]
+fn append_file_to_tarball<W, P>(
+    tarball: &mut tar::Builder<W>,
+    path: P,
+    mut auth_env_file: fs::File,
+) -> Result<(), Error>
+where
+    W: std::io::Write,
+    P: AsRef<std::path::Path>,
+{
+    use std::io::{Seek as _, SeekFrom};
+
+    // Note: A previous snapshot operation may have left the cursor
+    //       at the end of the file so we need to seek to the start.
+    auth_env_file.seek(SeekFrom::Start(0))?;
+    tarball.append_file(path, &mut auth_env_file)?;
+    Ok(())
+}
+
+/// Streams the content read from the given reader to S3.
+#[cfg(unix)]
+#[allow(clippy::too_many_arguments)]
+async fn multipart_stream_to_s3(
+    s3_bucket_url: String,
+    s3_bucket_region: String,
+    s3_bucket_name: String,
+    s3_snapshot_prefix: String,
+    s3_access_key: String,
+    s3_secret_key: String,
+    s3_max_in_flight_parts: std::num::NonZero<usize>,
+    s3_signature_duration: std::time::Duration,
+    s3_multipart_part_size: u64,
+    must_stop_processing: super::MustStopProcessing,
+    retry_backoff: backoff::exponential::ExponentialBackoff<backoff::SystemClock>,
+    db_name: String,
+    reader: std::io::PipeReader,
+) -> Result<(), Error> {
+    use std::{collections::VecDeque, os::fd::OwnedFd, path::PathBuf};
+
+    use bytes::{Bytes, BytesMut};
+    use reqwest::{Client, Response};
+    use rusty_s3::S3Action as _;
+    use rusty_s3::{actions::CreateMultipartUpload, Bucket, BucketError, Credentials, UrlStyle};
+    use tokio::task::JoinHandle;
+
+    let reader = OwnedFd::from(reader);
+    let reader = tokio::net::unix::pipe::Receiver::from_owned_fd(reader)?;
+    let s3_snapshot_prefix = PathBuf::from(s3_snapshot_prefix);
+    let url =
+        s3_bucket_url.parse().map_err(BucketError::ParseError).map_err(Error::S3BucketError)?;
+    let bucket = Bucket::new(url, UrlStyle::Path, s3_bucket_name, s3_bucket_region)
+        .map_err(Error::S3BucketError)?;
+    let credential = Credentials::new(s3_access_key, s3_secret_key);
+
+    // Note for the future (rust 1.91+): use with_added_extension, it's prettier
+    let object_path = s3_snapshot_prefix.join(format!("{db_name}.snapshot"));
+    // Note: It doesn't work on Windows and if a port to this platform is needed,
+    //       use the slash-path crate or similar to get the correct path separator.
+    let object = object_path.display().to_string();
+
+    let action = bucket.create_multipart_upload(Some(&credential), &object);
+    let url = action.sign(s3_signature_duration);
+
+    let client = Client::new();
+    let resp = client.post(url).send().await.map_err(Error::S3HttpError)?;
+    let status = resp.status();
+
+    let body = match resp.error_for_status_ref() {
+        Ok(_) => resp.text().await.map_err(Error::S3HttpError)?,
+        Err(_) => {
+            return Err(Error::S3Error { status, body: resp.text().await.unwrap_or_default() })
+        }
+    };
+
+    let multipart =
+        CreateMultipartUpload::parse_response(&body).map_err(|e| Error::S3XmlError(Box::new(e)))?;
+    tracing::debug!("Starting the upload of the snapshot to {object}");
+
+    // We use this bumpalo for etags strings.
+    let bump = bumpalo::Bump::new();
+    let mut etags = Vec::<&str>::new();
+    let mut in_flight = VecDeque::<(JoinHandle<reqwest::Result<Response>>, Bytes)>::with_capacity(
+        s3_max_in_flight_parts.get(),
+    );
+
+    // Part numbers start at 1 and cannot be larger than 10k
+    for part_number in 1u16.. {
+        if must_stop_processing.get() {
+            return Err(Error::AbortedTask);
+        }
+
+        let part_upload =
+            bucket.upload_part(Some(&credential), &object, part_number, multipart.upload_id());
+        let url = part_upload.sign(s3_signature_duration);
+
+        // Wait for a buffer to be ready if there are in-flight parts that landed
+        let mut buffer = if in_flight.len() >= s3_max_in_flight_parts.get() {
+            let (handle, buffer) = in_flight.pop_front().expect("At least one in flight request");
+            let resp = join_and_map_error(handle).await?;
+            extract_and_append_etag(&bump, &mut etags, resp.headers())?;
+
+            let mut buffer = match buffer.try_into_mut() {
+                Ok(buffer) => buffer,
+                Err(_) => unreachable!("All bytes references were consumed in the task"),
+            };
+            buffer.clear();
+            buffer
+        } else {
+            BytesMut::with_capacity(s3_multipart_part_size as usize)
+        };
+
+        // If we successfully read enough bytes,
+        // we can continue and send the buffer/part
+        while buffer.len() < (s3_multipart_part_size as usize / 2) {
+            // Wait for the pipe to be readable
+
+            use std::io;
+            reader.readable().await?;
+
+            match reader.try_read_buf(&mut buffer) {
+                Ok(0) => break,
+                // We read some bytes but maybe not enough
+                Ok(_) => continue,
+                // The readiness event is a false positive.
+                Err(ref e) if e.kind() == io::ErrorKind::WouldBlock => continue,
+                Err(e) => return Err(e.into()),
+            }
+        }
+
+        if buffer.is_empty() {
+            // Break the loop if the buffer is
+            // empty after we tried to read bytes
+            break;
+        }
+
+        let body = buffer.freeze();
+        tracing::trace!("Sending part {part_number}");
+        let task = tokio::spawn({
+            let client = client.clone();
+            let body = body.clone();
+            backoff::future::retry(retry_backoff.clone(), move || {
+                let client = client.clone();
+                let url = url.clone();
+                let body = body.clone();
+                async move {
+                    match client.put(url).body(body).send().await {
+                        Ok(resp) if resp.status().is_client_error() => {
+                            resp.error_for_status().map_err(backoff::Error::Permanent)
+                        }
+                        Ok(resp) => Ok(resp),
+                        Err(e) => Err(backoff::Error::transient(e)),
+                    }
+                }
+            })
+        });
+        in_flight.push_back((task, body));
+    }
+
+    for (handle, _buffer) in in_flight {
+        let resp = join_and_map_error(handle).await?;
+        extract_and_append_etag(&bump, &mut etags, resp.headers())?;
+    }
+
+    tracing::debug!("Finalizing the multipart upload");
+
+    let action = bucket.complete_multipart_upload(
+        Some(&credential),
+        &object,
+        multipart.upload_id(),
+        etags.iter().map(AsRef::as_ref),
+    );
+    let url = action.sign(s3_signature_duration);
+    let body = action.body();
+    let resp = backoff::future::retry(retry_backoff, move || {
+        let client = client.clone();
+        let url = url.clone();
+        let body = body.clone();
+        async move {
+            match client.post(url).body(body).send().await {
+                Ok(resp) if resp.status().is_client_error() => {
+                    resp.error_for_status().map_err(backoff::Error::Permanent)
+                }
+                Ok(resp) => Ok(resp),
+                Err(e) => Err(backoff::Error::transient(e)),
+            }
+        }
+    })
+    .await
+    .map_err(Error::S3HttpError)?;
+
+    let status = resp.status();
+    let body = resp.text().await.map_err(|e| Error::S3Error { status, body: e.to_string() })?;
+    if status.is_success() {
+        Ok(())
+    } else {
+        Err(Error::S3Error { status, body })
+    }
+}
+
+#[cfg(unix)]
+async fn join_and_map_error(
+    join_handle: tokio::task::JoinHandle<Result<reqwest::Response, reqwest::Error>>,
+) -> Result<reqwest::Response> {
+    // safety: Panic happens if the task (JoinHandle) was aborted, cancelled, or panicked
+    let request = join_handle.await.unwrap();
+    let resp = request.map_err(Error::S3HttpError)?;
+    match resp.error_for_status_ref() {
+        Ok(_) => Ok(resp),
+        Err(_) => Err(Error::S3Error {
+            status: resp.status(),
+            body: resp.text().await.unwrap_or_default(),
+        }),
+    }
+}
+
+#[cfg(unix)]
+fn extract_and_append_etag<'b>(
+    bump: &'b bumpalo::Bump,
+    etags: &mut Vec<&'b str>,
+    headers: &reqwest::header::HeaderMap,
+) -> Result<()> {
+    use reqwest::header::ETAG;
+
+    let etag = headers.get(ETAG).ok_or_else(|| Error::S3XmlError("Missing ETag header".into()))?;
+    let etag = etag.to_str().map_err(|e| Error::S3XmlError(Box::new(e)))?;
+    etags.push(bump.alloc_str(etag));
+
+    Ok(())
 }

crates/index-scheduler/src/scheduler/snapshots/test_failure.rs/upgrade_failure/after_processing_everything.snap

@@ -6,7 +6,7 @@ source: crates/index-scheduler/src/scheduler/test_failure.rs
 []
 ----------------------------------------------------------------------
 ### All Tasks:
-0 {uid: 0, batch_uid: 0, status: succeeded, details: { from: (1, 12, 0), to: (1, 23, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
+0 {uid: 0, batch_uid: 0, status: succeeded, details: { from: (1, 12, 0), to: (1, 24, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
 1 {uid: 1, batch_uid: 1, status: succeeded, details: { primary_key: Some("mouse"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "catto", primary_key: Some("mouse") }}
 2 {uid: 2, batch_uid: 2, status: succeeded, details: { primary_key: Some("bone"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "doggo", primary_key: Some("bone") }}
 3 {uid: 3, batch_uid: 3, status: failed, error: ResponseError { code: 200, message: "Index `doggo` already exists.", error_code: "index_already_exists", error_type: "invalid_request", error_link: "https://docs.meilisearch.com/errors#index_already_exists" }, details: { primary_key: Some("bone"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "doggo", primary_key: Some("bone") }}
@@ -57,7 +57,7 @@ girafo: { number_of_documents: 0, field_distribution: {} }
 [timestamp] [4,]
 ----------------------------------------------------------------------
 ### All Batches:
-0 {uid: 0, details: {"upgradeFrom":"v1.12.0","upgradeTo":"v1.23.0"}, stats: {"totalNbTasks":1,"status":{"succeeded":1},"types":{"upgradeDatabase":1},"indexUids":{}}, stop reason: "stopped after the last task of type `upgradeDatabase` because they cannot be batched with tasks of any other type.", }
+0 {uid: 0, details: {"upgradeFrom":"v1.12.0","upgradeTo":"v1.24.0"}, stats: {"totalNbTasks":1,"status":{"succeeded":1},"types":{"upgradeDatabase":1},"indexUids":{}}, stop reason: "stopped after the last task of type `upgradeDatabase` because they cannot be batched with tasks of any other type.", }
 1 {uid: 1, details: {"primaryKey":"mouse"}, stats: {"totalNbTasks":1,"status":{"succeeded":1},"types":{"indexCreation":1},"indexUids":{"catto":1}}, stop reason: "created batch containing only task with id 1 of type `indexCreation` that cannot be batched with any other task.", }
 2 {uid: 2, details: {"primaryKey":"bone"}, stats: {"totalNbTasks":1,"status":{"succeeded":1},"types":{"indexCreation":1},"indexUids":{"doggo":1}}, stop reason: "created batch containing only task with id 2 of type `indexCreation` that cannot be batched with any other task.", }
 3 {uid: 3, details: {"primaryKey":"bone"}, stats: {"totalNbTasks":1,"status":{"failed":1},"types":{"indexCreation":1},"indexUids":{"doggo":1}}, stop reason: "created batch containing only task with id 3 of type `indexCreation` that cannot be batched with any other task.", }

crates/index-scheduler/src/scheduler/snapshots/test_failure.rs/upgrade_failure/register_automatic_upgrade_task.snap

@@ -6,7 +6,7 @@ source: crates/index-scheduler/src/scheduler/test_failure.rs
 []
 ----------------------------------------------------------------------
 ### All Tasks:
-0 {uid: 0, status: enqueued, details: { from: (1, 12, 0), to: (1, 23, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
+0 {uid: 0, status: enqueued, details: { from: (1, 12, 0), to: (1, 24, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
 ----------------------------------------------------------------------
 ### Status:
 enqueued [0,]

crates/index-scheduler/src/scheduler/snapshots/test_failure.rs/upgrade_failure/registered_a_task_while_the_upgrade_task_is_enqueued.snap

@@ -6,7 +6,7 @@ source: crates/index-scheduler/src/scheduler/test_failure.rs
 []
 ----------------------------------------------------------------------
 ### All Tasks:
-0 {uid: 0, status: enqueued, details: { from: (1, 12, 0), to: (1, 23, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
+0 {uid: 0, status: enqueued, details: { from: (1, 12, 0), to: (1, 24, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
 1 {uid: 1, status: enqueued, details: { primary_key: Some("mouse"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "catto", primary_key: Some("mouse") }}
 ----------------------------------------------------------------------
 ### Status:

crates/index-scheduler/src/scheduler/snapshots/test_failure.rs/upgrade_failure/upgrade_task_failed.snap

@@ -6,7 +6,7 @@ source: crates/index-scheduler/src/scheduler/test_failure.rs
 []
 ----------------------------------------------------------------------
 ### All Tasks:
-0 {uid: 0, batch_uid: 0, status: failed, error: ResponseError { code: 200, message: "Planned failure for tests.", error_code: "internal", error_type: "internal", error_link: "https://docs.meilisearch.com/errors#internal" }, details: { from: (1, 12, 0), to: (1, 23, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
+0 {uid: 0, batch_uid: 0, status: failed, error: ResponseError { code: 200, message: "Planned failure for tests.", error_code: "internal", error_type: "internal", error_link: "https://docs.meilisearch.com/errors#internal" }, details: { from: (1, 12, 0), to: (1, 24, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
 1 {uid: 1, status: enqueued, details: { primary_key: Some("mouse"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "catto", primary_key: Some("mouse") }}
 ----------------------------------------------------------------------
 ### Status:
@@ -37,7 +37,7 @@ catto [1,]
 [timestamp] [0,]
 ----------------------------------------------------------------------
 ### All Batches:
-0 {uid: 0, details: {"upgradeFrom":"v1.12.0","upgradeTo":"v1.23.0"}, stats: {"totalNbTasks":1,"status":{"failed":1},"types":{"upgradeDatabase":1},"indexUids":{}}, stop reason: "stopped after the last task of type `upgradeDatabase` because they cannot be batched with tasks of any other type.", }
+0 {uid: 0, details: {"upgradeFrom":"v1.12.0","upgradeTo":"v1.24.0"}, stats: {"totalNbTasks":1,"status":{"failed":1},"types":{"upgradeDatabase":1},"indexUids":{}}, stop reason: "stopped after the last task of type `upgradeDatabase` because they cannot be batched with tasks of any other type.", }
 ----------------------------------------------------------------------
 ### Batch to tasks mapping:
 0 [0,]

crates/index-scheduler/src/scheduler/snapshots/test_failure.rs/upgrade_failure/upgrade_task_failed_again.snap

@@ -6,7 +6,7 @@ source: crates/index-scheduler/src/scheduler/test_failure.rs
 []
 ----------------------------------------------------------------------
 ### All Tasks:
-0 {uid: 0, batch_uid: 0, status: failed, error: ResponseError { code: 200, message: "Planned failure for tests.", error_code: "internal", error_type: "internal", error_link: "https://docs.meilisearch.com/errors#internal" }, details: { from: (1, 12, 0), to: (1, 23, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
+0 {uid: 0, batch_uid: 0, status: failed, error: ResponseError { code: 200, message: "Planned failure for tests.", error_code: "internal", error_type: "internal", error_link: "https://docs.meilisearch.com/errors#internal" }, details: { from: (1, 12, 0), to: (1, 24, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
 1 {uid: 1, status: enqueued, details: { primary_key: Some("mouse"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "catto", primary_key: Some("mouse") }}
 2 {uid: 2, status: enqueued, details: { primary_key: Some("bone"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "doggo", primary_key: Some("bone") }}
 ----------------------------------------------------------------------
@@ -40,7 +40,7 @@ doggo [2,]
 [timestamp] [0,]
 ----------------------------------------------------------------------
 ### All Batches:
-0 {uid: 0, details: {"upgradeFrom":"v1.12.0","upgradeTo":"v1.23.0"}, stats: {"totalNbTasks":1,"status":{"failed":1},"types":{"upgradeDatabase":1},"indexUids":{}}, stop reason: "stopped after the last task of type `upgradeDatabase` because they cannot be batched with tasks of any other type.", }
+0 {uid: 0, details: {"upgradeFrom":"v1.12.0","upgradeTo":"v1.24.0"}, stats: {"totalNbTasks":1,"status":{"failed":1},"types":{"upgradeDatabase":1},"indexUids":{}}, stop reason: "stopped after the last task of type `upgradeDatabase` because they cannot be batched with tasks of any other type.", }
 ----------------------------------------------------------------------
 ### Batch to tasks mapping:
 0 [0,]

crates/index-scheduler/src/scheduler/snapshots/test_failure.rs/upgrade_failure/upgrade_task_succeeded.snap

@@ -6,7 +6,7 @@ source: crates/index-scheduler/src/scheduler/test_failure.rs
 []
 ----------------------------------------------------------------------
 ### All Tasks:
-0 {uid: 0, batch_uid: 0, status: succeeded, details: { from: (1, 12, 0), to: (1, 23, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
+0 {uid: 0, batch_uid: 0, status: succeeded, details: { from: (1, 12, 0), to: (1, 24, 0) }, kind: UpgradeDatabase { from: (1, 12, 0) }}
 1 {uid: 1, status: enqueued, details: { primary_key: Some("mouse"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "catto", primary_key: Some("mouse") }}
 2 {uid: 2, status: enqueued, details: { primary_key: Some("bone"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "doggo", primary_key: Some("bone") }}
 3 {uid: 3, status: enqueued, details: { primary_key: Some("bone"), old_new_uid: None, new_index_uid: None }, kind: IndexCreation { index_uid: "doggo", primary_key: Some("bone") }}
@@ -43,7 +43,7 @@ doggo [2,3,]
 [timestamp] [0,]
 ----------------------------------------------------------------------
 ### All Batches:
-0 {uid: 0, details: {"upgradeFrom":"v1.12.0","upgradeTo":"v1.23.0"}, stats: {"totalNbTasks":1,"status":{"succeeded":1},"types":{"upgradeDatabase":1},"indexUids":{}}, stop reason: "stopped after the last task of type `upgradeDatabase` because they cannot be batched with tasks of any other type.", }
+0 {uid: 0, details: {"upgradeFrom":"v1.12.0","upgradeTo":"v1.24.0"}, stats: {"totalNbTasks":1,"status":{"succeeded":1},"types":{"upgradeDatabase":1},"indexUids":{}}, stop reason: "stopped after the last task of type `upgradeDatabase` because they cannot be batched with tasks of any other type.", }
 ----------------------------------------------------------------------
 ### Batch to tasks mapping:
 0 [0,]

crates/index-scheduler/src/test_utils.rs

@@ -126,7 +126,7 @@ impl IndexScheduler {
         std::fs::create_dir_all(&options.auth_path).unwrap();
         let auth_env = open_auth_store_env(&options.auth_path).unwrap();
         let index_scheduler =
-            Self::new(options, auth_env, version, sender, planned_failures).unwrap();
+            Self::new_test(options, auth_env, version, None, sender, planned_failures).unwrap();
 
         // To be 100% consistent between all test we're going to start the scheduler right now
         // and ensure it's in the expected starting state.

crates/index-scheduler/src/upgrade/mod.rs

@@ -47,6 +47,7 @@ pub fn upgrade_index_scheduler(
         (1, 21, _) => 0,
         (1, 22, _) => 0,
         (1, 23, _) => 0,
+        (1, 24, _) => 0,
         (major, minor, patch) => {
             if major > current_major
                 || (major == current_major && minor > current_minor)

crates/meilisearch-types/src/error.rs

@@ -258,6 +258,7 @@ InvalidIndexUid                                , InvalidRequest       , BAD_REQU
 InvalidMultiSearchFacets                       , InvalidRequest       , BAD_REQUEST ;
 InvalidMultiSearchFacetsByIndex                , InvalidRequest       , BAD_REQUEST ;
 InvalidMultiSearchFacetOrder                   , InvalidRequest       , BAD_REQUEST ;
+InvalidMultiSearchQueryPersonalization         , InvalidRequest       , BAD_REQUEST ;
 InvalidMultiSearchFederated                    , InvalidRequest       , BAD_REQUEST ;
 InvalidMultiSearchFederationOptions            , InvalidRequest       , BAD_REQUEST ;
 InvalidMultiSearchMaxValuesPerFacet            , InvalidRequest       , BAD_REQUEST ;
@@ -315,6 +316,8 @@ InvalidSearchShowRankingScoreDetails           , InvalidRequest       , BAD_REQU
 InvalidSimilarShowRankingScoreDetails          , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchSort                              , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchDistinct                          , InvalidRequest       , BAD_REQUEST ;
+InvalidSearchPersonalize                       , InvalidRequest       , BAD_REQUEST ;
+InvalidSearchPersonalizeUserContext            , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchMediaAndVector                    , InvalidRequest       , BAD_REQUEST ;
 InvalidSettingsDisplayedAttributes             , InvalidRequest       , BAD_REQUEST ;
 InvalidSettingsDistinctAttribute               , InvalidRequest       , BAD_REQUEST ;
@@ -390,6 +393,9 @@ TooManyVectors                                 , InvalidRequest       , BAD_REQU
 UnretrievableDocument                          , Internal             , BAD_REQUEST ;
 UnretrievableErrorCode                         , InvalidRequest       , BAD_REQUEST ;
 UnsupportedMediaType                           , InvalidRequest       , UNSUPPORTED_MEDIA_TYPE ;
+InvalidS3SnapshotRequest                       , Internal             , BAD_REQUEST ;
+InvalidS3SnapshotParameters                    , Internal             , BAD_REQUEST ;
+S3SnapshotServerError                          , Internal             , BAD_GATEWAY ;
 
 // Experimental features
 VectorEmbeddingError                           , InvalidRequest       , BAD_REQUEST ;
@@ -679,6 +685,18 @@ impl fmt::Display for deserr_codes::InvalidNetworkSearchApiKey {
     }
 }
 
+impl fmt::Display for deserr_codes::InvalidSearchPersonalize {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "the value of `personalize` is invalid, expected a JSON object with `userContext` string.")
+    }
+}
+
+impl fmt::Display for deserr_codes::InvalidSearchPersonalizeUserContext {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "the value of `userContext` is invalid, expected a string.")
+    }
+}
+
 #[macro_export]
 macro_rules! internal_error {
     ($target:ty : $($other:path), *) => {

crates/meilisearch-types/src/settings.rs

@@ -346,24 +346,26 @@ impl<T> Settings<T> {
                 continue;
             };
 
-            Self::hide_secret(api_key);
+            hide_secret(api_key, 0);
         }
     }
+}
 
-    fn hide_secret(secret: &mut String) {
-        match secret.len() {
-            x if x < 10 => {
-                secret.replace_range(.., "XXX...");
-            }
-            x if x < 20 => {
-                secret.replace_range(2.., "XXXX...");
-            }
-            x if x < 30 => {
-                secret.replace_range(3.., "XXXXX...");
-            }
-            _x => {
-                secret.replace_range(5.., "XXXXXX...");
-            }
+/// Redact a secret string, starting from the `secret_offset`th byte.
+pub fn hide_secret(secret: &mut String, secret_offset: usize) {
+    match secret.len().checked_sub(secret_offset) {
+        None => (),
+        Some(x) if x < 10 => {
+            secret.replace_range(secret_offset.., "XXX...");
+        }
+        Some(x) if x < 20 => {
+            secret.replace_range((secret_offset + 2).., "XXXX...");
+        }
+        Some(x) if x < 30 => {
+            secret.replace_range((secret_offset + 3).., "XXXXX...");
+        }
+        Some(_x) => {
+            secret.replace_range((secret_offset + 5).., "XXXXXX...");
         }
     }
 }

crates/meilisearch-types/src/webhooks.rs

@@ -11,6 +11,24 @@ pub struct Webhook {
     pub headers: BTreeMap<String, String>,
 }
 
+impl Webhook {
+    pub fn redact_authorization_header(&mut self) {
+        // headers are case insensitive, so to make the redaction robust we iterate over qualifying headers
+        // rather than getting one canonical `Authorization` header.
+        for value in self
+            .headers
+            .iter_mut()
+            .filter_map(|(name, value)| name.eq_ignore_ascii_case("authorization").then_some(value))
+        {
+            if value.starts_with("Bearer ") {
+                crate::settings::hide_secret(value, "Bearer ".len());
+            } else {
+                crate::settings::hide_secret(value, 0);
+            }
+        }
+    }
+}
+
 #[derive(Debug, Serialize, Default, Clone, PartialEq)]
 #[serde(rename_all = "camelCase")]
 pub struct WebhooksView {

crates/meilisearch/src/analytics/segment_analytics.rs

@@ -208,6 +208,7 @@ struct Infos {
     experimental_no_edition_2024_for_prefix_post_processing: bool,
     experimental_no_edition_2024_for_facet_post_processing: bool,
     experimental_vector_store_setting: bool,
+    experimental_personalization: bool,
     gpu_enabled: bool,
     db_path: bool,
     import_dump: bool,
@@ -217,6 +218,7 @@ struct Infos {
     import_snapshot: bool,
     schedule_snapshot: Option<u64>,
     snapshot_dir: bool,
+    uses_s3_snapshots: bool,
     ignore_missing_snapshot: bool,
     ignore_snapshot_if_db_exists: bool,
     http_addr: bool,
@@ -285,6 +287,8 @@ impl Infos {
             indexer_options,
             config_file_path,
             no_analytics: _,
+            experimental_personalization_api_key,
+            s3_snapshot_options,
         } = options;
 
         let schedule_snapshot = match schedule_snapshot {
@@ -348,6 +352,7 @@ impl Infos {
             import_snapshot: import_snapshot.is_some(),
             schedule_snapshot,
             snapshot_dir: snapshot_dir != PathBuf::from("snapshots/"),
+            uses_s3_snapshots: s3_snapshot_options.is_some(),
             ignore_missing_snapshot,
             ignore_snapshot_if_db_exists,
             http_addr: http_addr != default_http_addr(),
@@ -371,6 +376,7 @@ impl Infos {
             experimental_no_edition_2024_for_settings,
             experimental_no_edition_2024_for_prefix_post_processing,
             experimental_no_edition_2024_for_facet_post_processing,
+            experimental_personalization: experimental_personalization_api_key.is_some(),
         }
     }
 }

crates/meilisearch/src/error.rs

@@ -38,6 +38,8 @@ pub enum MeilisearchHttpError {
     PaginationInFederatedQuery(usize, &'static str),
     #[error("Inside `.queries[{0}]`: Using facet options is not allowed in federated queries.\n - Hint: remove `facets` from query #{0} or remove `federation` from the request\n - Hint: pass `federation.facetsByIndex.{1}: {2:?}` for facets in federated search")]
     FacetsInFederatedQuery(usize, String, Vec<String>),
+    #[error("Inside `.queries[{0}]`: Using `.personalize` is not allowed in federated queries.\n - Hint: remove `personalize` from query #{0} or remove `federation` from the request")]
+    PersonalizationInFederatedQuery(usize),
     #[error("Inconsistent order for values in facet `{facet}`: index `{previous_uid}` orders {previous_facet_order}, but index `{current_uid}` orders {index_facet_order}.\n - Hint: Remove `federation.mergeFacets` or change `faceting.sortFacetValuesBy` to be consistent in settings.")]
     InconsistentFacetOrder {
         facet: String,
@@ -137,6 +139,9 @@ impl ErrorCode for MeilisearchHttpError {
             MeilisearchHttpError::InconsistentFacetOrder { .. } => {
                 Code::InvalidMultiSearchFacetOrder
             }
+            MeilisearchHttpError::PersonalizationInFederatedQuery(_) => {
+                Code::InvalidMultiSearchQueryPersonalization
+            }
             MeilisearchHttpError::InconsistentOriginHeaders { .. } => {
                 Code::InconsistentDocumentChangeHeaders
             }

crates/meilisearch/src/lib.rs

@@ -11,6 +11,7 @@ pub mod middleware;
 pub mod option;
 #[cfg(test)]
 mod option_test;
+pub mod personalization;
 pub mod routes;
 pub mod search;
 pub mod search_queue;
@@ -58,6 +59,7 @@ use tracing::{error, info_span};
 use tracing_subscriber::filter::Targets;
 
 use crate::error::MeilisearchHttpError;
+use crate::personalization::PersonalizationService;
 
 /// Default number of simultaneously opened indexes.
 ///
@@ -128,12 +130,8 @@ pub type LogStderrType = tracing_subscriber::filter::Filtered<
 >;
 
 pub fn create_app(
-    index_scheduler: Data<IndexScheduler>,
-    auth_controller: Data<AuthController>,
-    search_queue: Data<SearchQueue>,
+    services: ServicesData,
     opt: Opt,
-    logs: (LogRouteHandle, LogStderrHandle),
-    analytics: Data<Analytics>,
     enable_dashboard: bool,
 ) -> actix_web::App<
     impl ServiceFactory<
@@ -145,17 +143,7 @@ pub fn create_app(
     >,
 > {
     let app = actix_web::App::new()
-        .configure(|s| {
-            configure_data(
-                s,
-                index_scheduler.clone(),
-                auth_controller.clone(),
-                search_queue.clone(),
-                &opt,
-                logs,
-                analytics.clone(),
-            )
-        })
+        .configure(|s| configure_data(s, services, &opt))
         .configure(routes::configure)
         .configure(|s| dashboard(s, enable_dashboard));
 
@@ -216,7 +204,10 @@ enum OnFailure {
     KeepDb,
 }
 
-pub fn setup_meilisearch(opt: &Opt) -> anyhow::Result<(Arc<IndexScheduler>, Arc<AuthController>)> {
+pub fn setup_meilisearch(
+    opt: &Opt,
+    handle: tokio::runtime::Handle,
+) -> anyhow::Result<(Arc<IndexScheduler>, Arc<AuthController>)> {
     let index_scheduler_opt = IndexSchedulerOptions {
         version_file_path: opt.db_path.join(VERSION_FILE_NAME),
         auth_path: opt.db_path.join("auth"),
@@ -230,7 +221,11 @@ pub fn setup_meilisearch(opt: &Opt) -> anyhow::Result<(Arc<IndexScheduler>, Arc<
         task_db_size: opt.max_task_db_size.as_u64() as usize,
         index_base_map_size: opt.max_index_size.as_u64() as usize,
         enable_mdb_writemap: opt.experimental_reduce_indexing_memory_usage,
-        indexer_config: Arc::new((&opt.indexer_options).try_into()?),
+        indexer_config: Arc::new({
+            let s3_snapshot_options =
+                opt.s3_snapshot_options.clone().map(|opt| opt.try_into()).transpose()?;
+            IndexerConfig { s3_snapshot_options, ..(&opt.indexer_options).try_into()? }
+        }),
         autobatching_enabled: true,
         cleanup_enabled: !opt.experimental_replication_parameters,
         max_number_of_tasks: 1_000_000,
@@ -256,6 +251,7 @@ pub fn setup_meilisearch(opt: &Opt) -> anyhow::Result<(Arc<IndexScheduler>, Arc<
                     index_scheduler_opt,
                     OnFailure::RemoveDb,
                     binary_version, // the db is empty
+                    handle,
                 )?,
                 Err(e) => {
                     std::fs::remove_dir_all(&opt.db_path)?;
@@ -273,7 +269,7 @@ pub fn setup_meilisearch(opt: &Opt) -> anyhow::Result<(Arc<IndexScheduler>, Arc<
             bail!("snapshot doesn't exist at {}", snapshot_path.display())
         // the snapshot and the db exist, and we can ignore the snapshot because of the ignore_snapshot_if_db_exists flag
         } else {
-            open_or_create_database(opt, index_scheduler_opt, empty_db, binary_version)?
+            open_or_create_database(opt, index_scheduler_opt, empty_db, binary_version, handle)?
         }
     } else if let Some(ref path) = opt.import_dump {
         let src_path_exists = path.exists();
@@ -284,6 +280,7 @@ pub fn setup_meilisearch(opt: &Opt) -> anyhow::Result<(Arc<IndexScheduler>, Arc<
                 index_scheduler_opt,
                 OnFailure::RemoveDb,
                 binary_version, // the db is empty
+                handle,
             )?;
             match import_dump(&opt.db_path, path, &mut index_scheduler, &mut auth_controller) {
                 Ok(()) => (index_scheduler, auth_controller),
@@ -304,10 +301,10 @@ pub fn setup_meilisearch(opt: &Opt) -> anyhow::Result<(Arc<IndexScheduler>, Arc<
         // the dump and the db exist and we can ignore the dump because of the ignore_dump_if_db_exists flag
         // or, the dump is missing but we can ignore that because of the ignore_missing_dump flag
         } else {
-            open_or_create_database(opt, index_scheduler_opt, empty_db, binary_version)?
+            open_or_create_database(opt, index_scheduler_opt, empty_db, binary_version, handle)?
         }
     } else {
-        open_or_create_database(opt, index_scheduler_opt, empty_db, binary_version)?
+        open_or_create_database(opt, index_scheduler_opt, empty_db, binary_version, handle)?
     };
 
     // We create a loop in a thread that registers snapshotCreation tasks
@@ -338,6 +335,7 @@ fn open_or_create_database_unchecked(
     index_scheduler_opt: IndexSchedulerOptions,
     on_failure: OnFailure,
     version: (u32, u32, u32),
+    handle: tokio::runtime::Handle,
 ) -> anyhow::Result<(IndexScheduler, AuthController)> {
     // we don't want to create anything in the data.ms yet, thus we
     // wrap our two builders in a closure that'll be executed later.
@@ -345,7 +343,7 @@ fn open_or_create_database_unchecked(
     let auth_env = open_auth_store_env(&index_scheduler_opt.auth_path).unwrap();
     let auth_controller = AuthController::new(auth_env.clone(), &opt.master_key);
     let index_scheduler_builder = || -> anyhow::Result<_> {
-        Ok(IndexScheduler::new(index_scheduler_opt, auth_env, version)?)
+        Ok(IndexScheduler::new(index_scheduler_opt, auth_env, version, Some(handle))?)
     };
 
     match (
@@ -452,6 +450,7 @@ fn open_or_create_database(
     index_scheduler_opt: IndexSchedulerOptions,
     empty_db: bool,
     binary_version: (u32, u32, u32),
+    handle: tokio::runtime::Handle,
 ) -> anyhow::Result<(IndexScheduler, AuthController)> {
     let version = if !empty_db {
         check_version(opt, &index_scheduler_opt, binary_version)?
@@ -459,7 +458,7 @@ fn open_or_create_database(
         binary_version
     };
 
-    open_or_create_database_unchecked(opt, index_scheduler_opt, OnFailure::KeepDb, version)
+    open_or_create_database_unchecked(opt, index_scheduler_opt, OnFailure::KeepDb, version, handle)
 }
 
 fn import_dump(
@@ -527,7 +526,11 @@ fn import_dump(
     let indexer_config = if base_config.max_threads.is_none() {
         let (thread_pool, _) = default_thread_pool_and_threads();
 
-        let _config = IndexerConfig { thread_pool, ..*base_config };
+        let _config = IndexerConfig {
+            thread_pool,
+            s3_snapshot_options: base_config.s3_snapshot_options.clone(),
+            ..*base_config
+        };
         backup_config = _config;
         &backup_config
     } else {
@@ -675,23 +678,26 @@ fn import_dump(
     Ok(index_scheduler_dump.finish()?)
 }
 
-pub fn configure_data(
-    config: &mut web::ServiceConfig,
-    index_scheduler: Data<IndexScheduler>,
-    auth: Data<AuthController>,
-    search_queue: Data<SearchQueue>,
-    opt: &Opt,
-    (logs_route, logs_stderr): (LogRouteHandle, LogStderrHandle),
-    analytics: Data<Analytics>,
-) {
+pub fn configure_data(config: &mut web::ServiceConfig, services: ServicesData, opt: &Opt) {
+    let ServicesData {
+        index_scheduler,
+        auth,
+        search_queue,
+        personalization_service,
+        logs_route_handle,
+        logs_stderr_handle,
+        analytics,
+    } = services;
+
     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::new(logs_route))
-        .app_data(web::Data::new(logs_stderr))
+        .app_data(personalization_service)
+        .app_data(logs_route_handle)
+        .app_data(logs_stderr_handle)
         .app_data(web::Data::new(opt.clone()))
         .app_data(
             web::JsonConfig::default()
@@ -752,3 +758,14 @@ pub fn dashboard(config: &mut web::ServiceConfig, enable_frontend: bool) {
 pub fn dashboard(config: &mut web::ServiceConfig, _enable_frontend: bool) {
     config.service(web::resource("/").route(web::get().to(routes::running)));
 }
+
+#[derive(Clone)]
+pub struct ServicesData {
+    pub index_scheduler: Data<IndexScheduler>,
+    pub auth: Data<AuthController>,
+    pub search_queue: Data<SearchQueue>,
+    pub personalization_service: Data<PersonalizationService>,
+    pub logs_route_handle: Data<LogRouteHandle>,
+    pub logs_stderr_handle: Data<LogStderrHandle>,
+    pub analytics: Data<Analytics>,
+}

crates/meilisearch/src/main.rs

@@ -14,10 +14,11 @@ use index_scheduler::IndexScheduler;
 use is_terminal::IsTerminal;
 use meilisearch::analytics::Analytics;
 use meilisearch::option::LogMode;
+use meilisearch::personalization::PersonalizationService;
 use meilisearch::search_queue::SearchQueue;
 use meilisearch::{
     analytics, create_app, setup_meilisearch, LogRouteHandle, LogRouteType, LogStderrHandle,
-    LogStderrType, Opt, SubscriberForSecondLayer,
+    LogStderrType, Opt, ServicesData, SubscriberForSecondLayer,
 };
 use meilisearch_auth::{generate_master_key, AuthController, MASTER_KEY_MIN_SIZE};
 use termcolor::{Color, ColorChoice, ColorSpec, StandardStream, WriteColor};
@@ -76,7 +77,10 @@ fn on_panic(info: &std::panic::PanicHookInfo) {
 
 #[actix_web::main]
 async fn main() -> anyhow::Result<()> {
-    try_main().await.inspect_err(|error| {
+    // won't panic inside of tokio::main
+    let runtime = tokio::runtime::Handle::current();
+
+    try_main(runtime).await.inspect_err(|error| {
         tracing::error!(%error);
         let mut current = error.source();
         let mut depth = 0;
@@ -88,7 +92,7 @@ async fn main() -> anyhow::Result<()> {
     })
 }
 
-async fn try_main() -> anyhow::Result<()> {
+async fn try_main(runtime: tokio::runtime::Handle) -> anyhow::Result<()> {
     let (opt, config_read_from) = Opt::try_build()?;
 
     std::panic::set_hook(Box::new(on_panic));
@@ -122,7 +126,7 @@ async fn try_main() -> anyhow::Result<()> {
         _ => (),
     }
 
-    let (index_scheduler, auth_controller) = setup_meilisearch(&opt)?;
+    let (index_scheduler, auth_controller) = setup_meilisearch(&opt, runtime)?;
 
     let analytics =
         analytics::Analytics::new(&opt, index_scheduler.clone(), auth_controller.clone()).await;
@@ -149,8 +153,15 @@ async fn run_http(
     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 auth = Data::from(auth_controller);
     let analytics = Data::from(analytics);
+    // Create personalization service with API key from options
+    let personalization_service = Data::new(
+        opt.experimental_personalization_api_key
+            .clone()
+            .map(PersonalizationService::cohere)
+            .unwrap_or_else(PersonalizationService::disabled),
+    );
     let search_queue = SearchQueue::new(
         opt.experimental_search_queue_size,
         available_parallelism()
@@ -162,21 +173,25 @@ async fn run_http(
         usize::from(opt.experimental_drop_search_after) as u64
     ));
     let search_queue = Data::new(search_queue);
+    let (logs_route_handle, logs_stderr_handle) = logs;
+    let logs_route_handle = Data::new(logs_route_handle);
+    let logs_stderr_handle = Data::new(logs_stderr_handle);
 
-    let http_server = HttpServer::new(move || {
-        create_app(
-            index_scheduler.clone(),
-            auth_controller.clone(),
-            search_queue.clone(),
-            opt.clone(),
-            logs.clone(),
-            analytics.clone(),
-            enable_dashboard,
-        )
-    })
-    // Disable signals allows the server to terminate immediately when a user enter CTRL-C
-    .disable_signals()
-    .keep_alive(KeepAlive::Os);
+    let services = ServicesData {
+        index_scheduler,
+        auth,
+        search_queue,
+        personalization_service,
+        logs_route_handle,
+        logs_stderr_handle,
+        analytics,
+    };
+
+    let http_server =
+        HttpServer::new(move || create_app(services.clone(), opt.clone(), enable_dashboard))
+            // Disable signals allows the server to terminate immediately when a user enter CTRL-C
+            .disable_signals()
+            .keep_alive(KeepAlive::Os);
 
     if let Some(config) = opt_clone.get_ssl_config()? {
         http_server.bind_rustls_0_23(opt_clone.http_addr, config)?.run().await?;

crates/meilisearch/src/metrics.rs

@@ -114,4 +114,9 @@ lazy_static! {
             "Meilisearch Task Queue Size Until Stop Registering",
         ))
         .expect("Can't create a metric");
+    pub static ref MEILISEARCH_PERSONALIZED_SEARCH_REQUESTS: IntGauge = register_int_gauge!(opts!(
+        "meilisearch_personalized_search_requests",
+        "Meilisearch number of search requests with personalization"
+    ))
+    .expect("Can't create a metric");
 }

crates/meilisearch/src/option.rs

@@ -7,12 +7,13 @@ use std::ops::Deref;
 use std::path::PathBuf;
 use std::str::FromStr;
 use std::sync::Arc;
+use std::time::Duration;
 use std::{env, fmt, fs};
 
 use byte_unit::{Byte, ParseError, UnitType};
 use clap::Parser;
 use meilisearch_types::features::InstanceTogglableFeatures;
-use meilisearch_types::milli::update::IndexerConfig;
+use meilisearch_types::milli::update::{IndexerConfig, S3SnapshotOptions};
 use meilisearch_types::milli::ThreadPoolNoAbortBuilder;
 use rustls::server::{ServerSessionMemoryCache, WebPkiClientVerifier};
 use rustls::RootCertStore;
@@ -74,6 +75,22 @@ const MEILI_EXPERIMENTAL_EMBEDDING_CACHE_ENTRIES: &str =
 const MEILI_EXPERIMENTAL_NO_SNAPSHOT_COMPACTION: &str = "MEILI_EXPERIMENTAL_NO_SNAPSHOT_COMPACTION";
 const MEILI_EXPERIMENTAL_NO_EDITION_2024_FOR_DUMPS: &str =
     "MEILI_EXPERIMENTAL_NO_EDITION_2024_FOR_DUMPS";
+const MEILI_EXPERIMENTAL_PERSONALIZATION_API_KEY: &str =
+    "MEILI_EXPERIMENTAL_PERSONALIZATION_API_KEY";
+
+// Related to S3 snapshots
+const MEILI_S3_BUCKET_URL: &str = "MEILI_S3_BUCKET_URL";
+const MEILI_S3_BUCKET_REGION: &str = "MEILI_S3_BUCKET_REGION";
+const MEILI_S3_BUCKET_NAME: &str = "MEILI_S3_BUCKET_NAME";
+const MEILI_S3_SNAPSHOT_PREFIX: &str = "MEILI_S3_SNAPSHOT_PREFIX";
+const MEILI_S3_ACCESS_KEY: &str = "MEILI_S3_ACCESS_KEY";
+const MEILI_S3_SECRET_KEY: &str = "MEILI_S3_SECRET_KEY";
+const MEILI_EXPERIMENTAL_S3_MAX_IN_FLIGHT_PARTS: &str = "MEILI_EXPERIMENTAL_S3_MAX_IN_FLIGHT_PARTS";
+const MEILI_EXPERIMENTAL_S3_COMPRESSION_LEVEL: &str = "MEILI_EXPERIMENTAL_S3_COMPRESSION_LEVEL";
+const MEILI_EXPERIMENTAL_S3_SIGNATURE_DURATION_SECONDS: &str =
+    "MEILI_EXPERIMENTAL_S3_SIGNATURE_DURATION_SECONDS";
+const MEILI_EXPERIMENTAL_S3_MULTIPART_PART_SIZE: &str = "MEILI_EXPERIMENTAL_S3_MULTIPART_PART_SIZE";
+
 const DEFAULT_CONFIG_FILE_PATH: &str = "./config.toml";
 const DEFAULT_DB_PATH: &str = "./data.ms";
 const DEFAULT_HTTP_ADDR: &str = "localhost:7700";
@@ -83,6 +100,10 @@ const DEFAULT_SNAPSHOT_DIR: &str = "snapshots/";
 const DEFAULT_SNAPSHOT_INTERVAL_SEC: u64 = 86400;
 const DEFAULT_SNAPSHOT_INTERVAL_SEC_STR: &str = "86400";
 const DEFAULT_DUMP_DIR: &str = "dumps/";
+const DEFAULT_S3_SNAPSHOT_MAX_IN_FLIGHT_PARTS: NonZeroUsize = NonZeroUsize::new(10).unwrap();
+const DEFAULT_S3_SNAPSHOT_COMPRESSION_LEVEL: u32 = 0;
+const DEFAULT_S3_SNAPSHOT_SIGNATURE_DURATION_SECONDS: u64 = 8 * 3600; // 8 hours
+const DEFAULT_S3_SNAPSHOT_MULTIPART_PART_SIZE: Byte = Byte::from_u64(375 * 1024 * 1024); // 375 MiB
 
 const MEILI_MAX_INDEXING_MEMORY: &str = "MEILI_MAX_INDEXING_MEMORY";
 const MEILI_MAX_INDEXING_THREADS: &str = "MEILI_MAX_INDEXING_THREADS";
@@ -475,10 +496,20 @@ pub struct Opt {
     #[serde(default)]
     pub experimental_no_snapshot_compaction: bool,
 
+    /// Experimental personalization API key feature.
+    ///
+    /// Sets the API key for personalization features.
+    #[clap(long, env = MEILI_EXPERIMENTAL_PERSONALIZATION_API_KEY)]
+    pub experimental_personalization_api_key: Option<String>,
+
     #[serde(flatten)]
     #[clap(flatten)]
     pub indexer_options: IndexerOpts,
 
+    #[serde(flatten)]
+    #[clap(flatten)]
+    pub s3_snapshot_options: Option<S3SnapshotOpts>,
+
     /// Set the path to a configuration file that should be used to setup the engine.
     /// Format must be TOML.
     #[clap(long)]
@@ -580,6 +611,8 @@ impl Opt {
             experimental_limit_batched_tasks_total_size,
             experimental_embedding_cache_entries,
             experimental_no_snapshot_compaction,
+            experimental_personalization_api_key,
+            s3_snapshot_options,
         } = self;
         export_to_env_if_not_present(MEILI_DB_PATH, db_path);
         export_to_env_if_not_present(MEILI_HTTP_ADDR, http_addr);
@@ -680,7 +713,22 @@ impl Opt {
             MEILI_EXPERIMENTAL_NO_SNAPSHOT_COMPACTION,
             experimental_no_snapshot_compaction.to_string(),
         );
+        if let Some(experimental_personalization_api_key) = experimental_personalization_api_key {
+            export_to_env_if_not_present(
+                MEILI_EXPERIMENTAL_PERSONALIZATION_API_KEY,
+                experimental_personalization_api_key,
+            );
+        }
         indexer_options.export_to_env();
+        if let Some(s3_snapshot_options) = s3_snapshot_options {
+            #[cfg(not(unix))]
+            {
+                let _ = s3_snapshot_options;
+                panic!("S3 snapshot options are not supported on Windows");
+            }
+            #[cfg(unix)]
+            s3_snapshot_options.export_to_env();
+        }
     }
 
     pub fn get_ssl_config(&self) -> anyhow::Result<Option<rustls::ServerConfig>> {
@@ -849,6 +897,16 @@ impl TryFrom<&IndexerOpts> for IndexerConfig {
     type Error = anyhow::Error;
 
     fn try_from(other: &IndexerOpts) -> Result<Self, Self::Error> {
+        let IndexerOpts {
+            max_indexing_memory,
+            max_indexing_threads,
+            skip_index_budget,
+            experimental_no_edition_2024_for_settings,
+            experimental_no_edition_2024_for_dumps,
+            experimental_no_edition_2024_for_prefix_post_processing,
+            experimental_no_edition_2024_for_facet_post_processing,
+        } = other;
+
         let thread_pool = ThreadPoolNoAbortBuilder::new_for_indexing()
             .num_threads(other.max_indexing_threads.unwrap_or_else(|| num_cpus::get() / 2))
             .build()?;
@@ -856,21 +914,163 @@ impl TryFrom<&IndexerOpts> for IndexerConfig {
         Ok(Self {
             thread_pool,
             log_every_n: Some(DEFAULT_LOG_EVERY_N),
-            max_memory: other.max_indexing_memory.map(|b| b.as_u64() as usize),
-            max_threads: *other.max_indexing_threads,
+            max_memory: max_indexing_memory.map(|b| b.as_u64() as usize),
+            max_threads: max_indexing_threads.0,
             max_positions_per_attributes: None,
-            skip_index_budget: other.skip_index_budget,
-            experimental_no_edition_2024_for_settings: other
-                .experimental_no_edition_2024_for_settings,
-            experimental_no_edition_2024_for_dumps: other.experimental_no_edition_2024_for_dumps,
+            skip_index_budget: *skip_index_budget,
+            experimental_no_edition_2024_for_settings: *experimental_no_edition_2024_for_settings,
+            experimental_no_edition_2024_for_dumps: *experimental_no_edition_2024_for_dumps,
             chunk_compression_type: Default::default(),
             chunk_compression_level: Default::default(),
             documents_chunk_size: Default::default(),
             max_nb_chunks: Default::default(),
-            experimental_no_edition_2024_for_prefix_post_processing: other
-                .experimental_no_edition_2024_for_prefix_post_processing,
-            experimental_no_edition_2024_for_facet_post_processing: other
-                .experimental_no_edition_2024_for_facet_post_processing,
+            experimental_no_edition_2024_for_prefix_post_processing:
+                *experimental_no_edition_2024_for_prefix_post_processing,
+            experimental_no_edition_2024_for_facet_post_processing:
+                *experimental_no_edition_2024_for_facet_post_processing,
+            s3_snapshot_options: None,
+        })
+    }
+}
+
+#[derive(Debug, Clone, Parser, Deserialize)]
+// This group is a bit tricky but makes it possible to require all listed fields if one of them
+// is specified. It lets us keep an Option for the S3SnapshotOpts configuration.
+// <https://github.com/clap-rs/clap/issues/5092#issuecomment-2616986075>
+#[group(requires_all = ["s3_bucket_url", "s3_bucket_region", "s3_bucket_name", "s3_snapshot_prefix", "s3_access_key", "s3_secret_key"])]
+pub struct S3SnapshotOpts {
+    /// The S3 bucket URL in the format https://s3.<region>.amazonaws.com.
+    #[clap(long, env = MEILI_S3_BUCKET_URL, required = false)]
+    #[serde(default)]
+    pub s3_bucket_url: String,
+
+    /// The region in the format us-east-1.
+    #[clap(long, env = MEILI_S3_BUCKET_REGION, required = false)]
+    #[serde(default)]
+    pub s3_bucket_region: String,
+
+    /// The bucket name.
+    #[clap(long, env = MEILI_S3_BUCKET_NAME, required = false)]
+    #[serde(default)]
+    pub s3_bucket_name: String,
+
+    /// The prefix path where to put the snapshot, uses normal slashes (/).
+    #[clap(long, env = MEILI_S3_SNAPSHOT_PREFIX, required = false)]
+    #[serde(default)]
+    pub s3_snapshot_prefix: String,
+
+    /// The S3 access key.
+    #[clap(long, env = MEILI_S3_ACCESS_KEY, required = false)]
+    #[serde(default)]
+    pub s3_access_key: String,
+
+    /// The S3 secret key.
+    #[clap(long, env = MEILI_S3_SECRET_KEY, required = false)]
+    #[serde(default)]
+    pub s3_secret_key: String,
+
+    /// The maximum number of parts that can be uploaded in parallel.
+    ///
+    /// For more information, see <https://github.com/orgs/meilisearch/discussions/869>.
+    #[clap(long, env = MEILI_EXPERIMENTAL_S3_MAX_IN_FLIGHT_PARTS, default_value_t = default_experimental_s3_snapshot_max_in_flight_parts())]
+    #[serde(default = "default_experimental_s3_snapshot_max_in_flight_parts")]
+    pub experimental_s3_max_in_flight_parts: NonZeroUsize,
+
+    /// The compression level. Defaults to no compression (0).
+    ///
+    /// For more information, see <https://github.com/orgs/meilisearch/discussions/869>.
+    #[clap(long, env = MEILI_EXPERIMENTAL_S3_COMPRESSION_LEVEL, default_value_t = default_experimental_s3_snapshot_compression_level())]
+    #[serde(default = "default_experimental_s3_snapshot_compression_level")]
+    pub experimental_s3_compression_level: u32,
+
+    /// The signature duration for the multipart upload.
+    ///
+    /// For more information, see <https://github.com/orgs/meilisearch/discussions/869>.
+    #[clap(long, env = MEILI_EXPERIMENTAL_S3_SIGNATURE_DURATION_SECONDS, default_value_t = default_experimental_s3_snapshot_signature_duration_seconds())]
+    #[serde(default = "default_experimental_s3_snapshot_signature_duration_seconds")]
+    pub experimental_s3_signature_duration_seconds: u64,
+
+    /// The size of the the multipart parts.
+    ///
+    /// Must not be less than 10MiB and larger than 8GiB. Yes,
+    /// twice the boundaries of the AWS S3 multipart upload
+    /// because we use it a bit differently internally.
+    ///
+    /// For more information, see <https://github.com/orgs/meilisearch/discussions/869>.
+    #[clap(long, env = MEILI_EXPERIMENTAL_S3_MULTIPART_PART_SIZE, default_value_t = default_experimental_s3_snapshot_multipart_part_size())]
+    #[serde(default = "default_experimental_s3_snapshot_multipart_part_size")]
+    pub experimental_s3_multipart_part_size: Byte,
+}
+
+impl S3SnapshotOpts {
+    /// Exports the values to their corresponding env vars if they are not set.
+    pub fn export_to_env(self) {
+        let S3SnapshotOpts {
+            s3_bucket_url,
+            s3_bucket_region,
+            s3_bucket_name,
+            s3_snapshot_prefix,
+            s3_access_key,
+            s3_secret_key,
+            experimental_s3_max_in_flight_parts,
+            experimental_s3_compression_level,
+            experimental_s3_signature_duration_seconds,
+            experimental_s3_multipart_part_size,
+        } = self;
+
+        export_to_env_if_not_present(MEILI_S3_BUCKET_URL, s3_bucket_url);
+        export_to_env_if_not_present(MEILI_S3_BUCKET_REGION, s3_bucket_region);
+        export_to_env_if_not_present(MEILI_S3_BUCKET_NAME, s3_bucket_name);
+        export_to_env_if_not_present(MEILI_S3_SNAPSHOT_PREFIX, s3_snapshot_prefix);
+        export_to_env_if_not_present(MEILI_S3_ACCESS_KEY, s3_access_key);
+        export_to_env_if_not_present(MEILI_S3_SECRET_KEY, s3_secret_key);
+        export_to_env_if_not_present(
+            MEILI_EXPERIMENTAL_S3_MAX_IN_FLIGHT_PARTS,
+            experimental_s3_max_in_flight_parts.to_string(),
+        );
+        export_to_env_if_not_present(
+            MEILI_EXPERIMENTAL_S3_COMPRESSION_LEVEL,
+            experimental_s3_compression_level.to_string(),
+        );
+        export_to_env_if_not_present(
+            MEILI_EXPERIMENTAL_S3_SIGNATURE_DURATION_SECONDS,
+            experimental_s3_signature_duration_seconds.to_string(),
+        );
+        export_to_env_if_not_present(
+            MEILI_EXPERIMENTAL_S3_MULTIPART_PART_SIZE,
+            experimental_s3_multipart_part_size.to_string(),
+        );
+    }
+}
+
+impl TryFrom<S3SnapshotOpts> for S3SnapshotOptions {
+    type Error = anyhow::Error;
+
+    fn try_from(other: S3SnapshotOpts) -> Result<Self, Self::Error> {
+        let S3SnapshotOpts {
+            s3_bucket_url,
+            s3_bucket_region,
+            s3_bucket_name,
+            s3_snapshot_prefix,
+            s3_access_key,
+            s3_secret_key,
+            experimental_s3_max_in_flight_parts,
+            experimental_s3_compression_level,
+            experimental_s3_signature_duration_seconds,
+            experimental_s3_multipart_part_size,
+        } = other;
+
+        Ok(S3SnapshotOptions {
+            s3_bucket_url,
+            s3_bucket_region,
+            s3_bucket_name,
+            s3_snapshot_prefix,
+            s3_access_key,
+            s3_secret_key,
+            s3_max_in_flight_parts: experimental_s3_max_in_flight_parts,
+            s3_compression_level: experimental_s3_compression_level,
+            s3_signature_duration: Duration::from_secs(experimental_s3_signature_duration_seconds),
+            s3_multipart_part_size: experimental_s3_multipart_part_size.as_u64(),
         })
     }
 }
@@ -1089,6 +1289,22 @@ fn default_snapshot_interval_sec() -> &'static str {
     DEFAULT_SNAPSHOT_INTERVAL_SEC_STR
 }
 
+fn default_experimental_s3_snapshot_max_in_flight_parts() -> NonZeroUsize {
+    DEFAULT_S3_SNAPSHOT_MAX_IN_FLIGHT_PARTS
+}
+
+fn default_experimental_s3_snapshot_compression_level() -> u32 {
+    DEFAULT_S3_SNAPSHOT_COMPRESSION_LEVEL
+}
+
+fn default_experimental_s3_snapshot_signature_duration_seconds() -> u64 {
+    DEFAULT_S3_SNAPSHOT_SIGNATURE_DURATION_SECONDS
+}
+
+fn default_experimental_s3_snapshot_multipart_part_size() -> Byte {
+    DEFAULT_S3_SNAPSHOT_MULTIPART_PART_SIZE
+}
+
 fn default_dump_dir() -> PathBuf {
     PathBuf::from(DEFAULT_DUMP_DIR)
 }

crates/meilisearch/src/personalization/mod.rs

@@ -0,0 +1,366 @@
+use crate::search::{Personalize, SearchResult};
+use meilisearch_types::{
+    error::{Code, ErrorCode, ResponseError},
+    milli::TimeBudget,
+};
+use rand::Rng;
+use reqwest::Client;
+use serde::{Deserialize, Serialize};
+use std::time::Duration;
+use tracing::{debug, info, warn};
+
+const COHERE_API_URL: &str = "https://api.cohere.ai/v1/rerank";
+const MAX_RETRIES: u32 = 10;
+
+#[derive(Debug, thiserror::Error)]
+enum PersonalizationError {
+    #[error("Personalization service: HTTP request failed: {0}")]
+    Request(#[from] reqwest::Error),
+    #[error("Personalization service: Failed to parse response: {0}")]
+    Parse(String),
+    #[error("Personalization service: Cohere API error: {0}")]
+    Api(String),
+    #[error("Personalization service: Unauthorized: invalid API key")]
+    Unauthorized,
+    #[error("Personalization service: Rate limited: too many requests")]
+    RateLimited,
+    #[error("Personalization service: Bad request: {0}")]
+    BadRequest(String),
+    #[error("Personalization service: Internal server error: {0}")]
+    InternalServerError(String),
+    #[error("Personalization service: Network error: {0}")]
+    Network(String),
+    #[error("Personalization service: Deadline exceeded")]
+    DeadlineExceeded,
+    #[error(transparent)]
+    FeatureNotEnabled(#[from] index_scheduler::error::FeatureNotEnabledError),
+}
+
+impl ErrorCode for PersonalizationError {
+    fn error_code(&self) -> Code {
+        match self {
+            PersonalizationError::FeatureNotEnabled { .. } => Code::FeatureNotEnabled,
+            PersonalizationError::Unauthorized => Code::RemoteInvalidApiKey,
+            PersonalizationError::RateLimited => Code::TooManySearchRequests,
+            PersonalizationError::BadRequest(_) => Code::RemoteBadRequest,
+            PersonalizationError::InternalServerError(_) => Code::RemoteRemoteError,
+            PersonalizationError::Network(_) | PersonalizationError::Request(_) => {
+                Code::RemoteCouldNotSendRequest
+            }
+            PersonalizationError::Parse(_) | PersonalizationError::Api(_) => {
+                Code::RemoteBadResponse
+            }
+            PersonalizationError::DeadlineExceeded => Code::Internal, // should not be returned to the client
+        }
+    }
+}
+
+pub struct CohereService {
+    client: Client,
+    api_key: String,
+}
+
+impl CohereService {
+    pub fn new(api_key: String) -> Self {
+        info!("Personalization service initialized with Cohere API");
+        let client = Client::builder()
+            .timeout(Duration::from_secs(30))
+            .build()
+            .expect("Failed to create HTTP client");
+        Self { client, api_key }
+    }
+
+    pub async fn rerank_search_results(
+        &self,
+        search_result: SearchResult,
+        personalize: &Personalize,
+        query: Option<&str>,
+        time_budget: TimeBudget,
+    ) -> Result<SearchResult, ResponseError> {
+        if time_budget.exceeded() {
+            warn!("Could not rerank due to deadline");
+            // If the deadline is exceeded, return the original search result instead of an error
+            return Ok(search_result);
+        }
+
+        // Extract user context from personalization
+        let user_context = personalize.user_context.as_str();
+
+        // Build the prompt by merging query and user context
+        let prompt = match query {
+            Some(q) => format!("User Context: {user_context}\nQuery: {q}"),
+            None => format!("User Context: {user_context}"),
+        };
+
+        // Extract documents for reranking
+        let documents: Vec<String> = search_result
+            .hits
+            .iter()
+            .map(|hit| {
+                // Convert the document to a string representation for reranking
+                serde_json::to_string(&hit.document).unwrap_or_else(|_| "{}".to_string())
+            })
+            .collect();
+
+        if documents.is_empty() {
+            return Ok(search_result);
+        }
+
+        // Call Cohere's rerank API with retry logic
+        let reranked_indices =
+            match self.call_rerank_with_retry(&prompt, &documents, time_budget).await {
+                Ok(indices) => indices,
+                Err(PersonalizationError::DeadlineExceeded) => {
+                    // If the deadline is exceeded, return the original search result instead of an error
+                    return Ok(search_result);
+                }
+                Err(e) => return Err(e.into()),
+            };
+
+        debug!("Cohere rerank successful, reordering {} results", search_result.hits.len());
+
+        // Reorder the hits based on Cohere's reranking
+        let mut reranked_hits = Vec::new();
+        for index in reranked_indices.iter() {
+            if let Some(hit) = search_result.hits.get(*index) {
+                reranked_hits.push(hit.clone());
+            }
+        }
+
+        Ok(SearchResult { hits: reranked_hits, ..search_result })
+    }
+
+    async fn call_rerank_with_retry(
+        &self,
+        query: &str,
+        documents: &[String],
+        time_budget: TimeBudget,
+    ) -> Result<Vec<usize>, PersonalizationError> {
+        let request_body = CohereRerankRequest {
+            query: query.to_string(),
+            documents: documents.to_vec(),
+            model: "rerank-english-v3.0".to_string(),
+        };
+
+        // Retry loop similar to vector extraction
+        for attempt in 0..MAX_RETRIES {
+            let response_result = self.send_rerank_request(&request_body).await;
+
+            let retry_duration = match self.handle_response(response_result).await {
+                Ok(indices) => return Ok(indices),
+                Err(retry) => {
+                    warn!("Cohere rerank attempt #{} failed: {}", attempt, retry.error);
+
+                    if time_budget.exceeded() {
+                        warn!("Could not rerank due to deadline");
+                        return Err(PersonalizationError::DeadlineExceeded);
+                    } else {
+                        match retry.into_duration(attempt) {
+                            Ok(d) => d,
+                            Err(error) => return Err(error),
+                        }
+                    }
+                }
+            };
+
+            // randomly up to double the retry duration
+            let retry_duration = retry_duration
+                + rand::thread_rng().gen_range(std::time::Duration::ZERO..retry_duration);
+
+            warn!("Retrying after {}ms", retry_duration.as_millis());
+            tokio::time::sleep(retry_duration).await;
+        }
+
+        // Final attempt without retry
+        let response_result = self.send_rerank_request(&request_body).await;
+
+        match self.handle_response(response_result).await {
+            Ok(indices) => Ok(indices),
+            Err(retry) => Err(retry.into_error()),
+        }
+    }
+
+    async fn send_rerank_request(
+        &self,
+        request_body: &CohereRerankRequest,
+    ) -> Result<reqwest::Response, reqwest::Error> {
+        self.client
+            .post(COHERE_API_URL)
+            .header("Authorization", format!("Bearer {}", self.api_key))
+            .header("Content-Type", "application/json")
+            .json(request_body)
+            .send()
+            .await
+    }
+
+    async fn handle_response(
+        &self,
+        response_result: Result<reqwest::Response, reqwest::Error>,
+    ) -> Result<Vec<usize>, Retry> {
+        let response = match response_result {
+            Ok(r) => r,
+            Err(e) if e.is_timeout() => {
+                return Err(Retry::retry_later(PersonalizationError::Network(format!(
+                    "Request timeout: {}",
+                    e
+                ))));
+            }
+            Err(e) => {
+                return Err(Retry::retry_later(PersonalizationError::Network(format!(
+                    "Network error: {}",
+                    e
+                ))));
+            }
+        };
+
+        let status = response.status();
+        let status_code = status.as_u16();
+
+        if status.is_success() {
+            let rerank_response: CohereRerankResponse = match response.json().await {
+                Ok(r) => r,
+                Err(e) => {
+                    return Err(Retry::retry_later(PersonalizationError::Parse(format!(
+                        "Failed to parse response: {}",
+                        e
+                    ))));
+                }
+            };
+
+            // Extract indices from rerank results
+            let indices: Vec<usize> =
+                rerank_response.results.iter().map(|result| result.index as usize).collect();
+
+            return Ok(indices);
+        }
+
+        // Handle error status codes
+        let error_body = response.text().await.unwrap_or_else(|_| "Unknown error".to_string());
+
+        let retry = match status_code {
+            401 => Retry::give_up(PersonalizationError::Unauthorized),
+            429 => Retry::rate_limited(PersonalizationError::RateLimited),
+            400 => Retry::give_up(PersonalizationError::BadRequest(error_body)),
+            500..=599 => Retry::retry_later(PersonalizationError::InternalServerError(format!(
+                "Status {}: {}",
+                status_code, error_body
+            ))),
+            402..=499 => Retry::give_up(PersonalizationError::Api(format!(
+                "Status {}: {}",
+                status_code, error_body
+            ))),
+            _ => Retry::retry_later(PersonalizationError::Api(format!(
+                "Unexpected status {}: {}",
+                status_code, error_body
+            ))),
+        };
+
+        Err(retry)
+    }
+}
+
+#[derive(Serialize)]
+struct CohereRerankRequest {
+    query: String,
+    documents: Vec<String>,
+    model: String,
+}
+
+#[derive(Deserialize)]
+struct CohereRerankResponse {
+    results: Vec<CohereRerankResult>,
+}
+
+#[derive(Deserialize)]
+struct CohereRerankResult {
+    index: u32,
+}
+
+// Retry strategy similar to vector extraction
+struct Retry {
+    error: PersonalizationError,
+    strategy: RetryStrategy,
+}
+
+enum RetryStrategy {
+    GiveUp,
+    Retry,
+    RetryAfterRateLimit,
+}
+
+impl Retry {
+    fn give_up(error: PersonalizationError) -> Self {
+        Self { error, strategy: RetryStrategy::GiveUp }
+    }
+
+    fn retry_later(error: PersonalizationError) -> Self {
+        Self { error, strategy: RetryStrategy::Retry }
+    }
+
+    fn rate_limited(error: PersonalizationError) -> Self {
+        Self { error, strategy: RetryStrategy::RetryAfterRateLimit }
+    }
+
+    fn into_duration(self, attempt: u32) -> Result<Duration, PersonalizationError> {
+        match self.strategy {
+            RetryStrategy::GiveUp => Err(self.error),
+            RetryStrategy::Retry => {
+                // Exponential backoff: 10^attempt milliseconds
+                Ok(Duration::from_millis((10u64).pow(attempt)))
+            }
+            RetryStrategy::RetryAfterRateLimit => {
+                // Longer backoff for rate limits: 100ms + exponential
+                Ok(Duration::from_millis(100 + (10u64).pow(attempt)))
+            }
+        }
+    }
+
+    fn into_error(self) -> PersonalizationError {
+        self.error
+    }
+}
+
+pub enum PersonalizationService {
+    Cohere(CohereService),
+    Disabled,
+}
+
+impl PersonalizationService {
+    pub fn cohere(api_key: String) -> Self {
+        // If the API key is empty, consider the personalization service as disabled
+        if api_key.trim().is_empty() {
+            Self::disabled()
+        } else {
+            Self::Cohere(CohereService::new(api_key))
+        }
+    }
+
+    pub fn disabled() -> Self {
+        debug!("Personalization service disabled");
+        Self::Disabled
+    }
+
+    pub async fn rerank_search_results(
+        &self,
+        search_result: SearchResult,
+        personalize: &Personalize,
+        query: Option<&str>,
+        time_budget: TimeBudget,
+    ) -> Result<SearchResult, ResponseError> {
+        match self {
+            Self::Cohere(cohere_service) => {
+                cohere_service
+                    .rerank_search_results(search_result, personalize, query, time_budget)
+                    .await
+            }
+            Self::Disabled => Err(PersonalizationError::FeatureNotEnabled(
+                index_scheduler::error::FeatureNotEnabledError {
+                    disabled_action: "reranking search results",
+                    feature: "personalization",
+                    issue_link: "https://github.com/orgs/meilisearch/discussions/866",
+                },
+            )
+            .into()),
+        }
+    }
+}

crates/meilisearch/src/routes/indexes/facet_search.rs

@@ -343,6 +343,7 @@ impl From<FacetSearchQuery> for SearchQuery {
             hybrid,
             ranking_score_threshold,
             locales,
+            personalize: None,
         }
     }
 }

crates/meilisearch/src/routes/indexes/search.rs

@@ -22,11 +22,12 @@ 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::routes::parse_include_metadata_header;
 use crate::search::{
-    add_search_rules, perform_search, HybridQuery, MatchingStrategy, RankingScoreThreshold,
-    RetrieveVectors, SearchKind, SearchQuery, SearchResult, SemanticRatio, DEFAULT_CROP_LENGTH,
-    DEFAULT_CROP_MARKER, DEFAULT_HIGHLIGHT_POST_TAG, DEFAULT_HIGHLIGHT_PRE_TAG,
-    DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET, DEFAULT_SEMANTIC_RATIO,
+    add_search_rules, perform_search, HybridQuery, MatchingStrategy, Personalize,
+    RankingScoreThreshold, RetrieveVectors, SearchKind, SearchParams, SearchQuery, SearchResult,
+    SemanticRatio, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER, DEFAULT_HIGHLIGHT_POST_TAG,
+    DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET, DEFAULT_SEMANTIC_RATIO,
 };
 use crate::search_queue::SearchQueue;
 
@@ -133,6 +134,8 @@ pub struct SearchQueryGet {
     #[deserr(default, error = DeserrQueryParamError<InvalidSearchLocales>)]
     #[param(value_type = Vec<Locale>, explode = false)]
     pub locales: Option<CS<Locale>>,
+    #[deserr(default, error = DeserrQueryParamError<InvalidSearchPersonalizeUserContext>)]
+    pub personalize_user_context: Option<String>,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, deserr::Deserr)]
@@ -204,6 +207,9 @@ impl TryFrom<SearchQueryGet> for SearchQuery {
             ));
         }
 
+        let personalize =
+            other.personalize_user_context.map(|user_context| Personalize { user_context });
+
         Ok(Self {
             q: other.q,
             // `media` not supported for `GET`
@@ -233,6 +239,7 @@ impl TryFrom<SearchQueryGet> for SearchQuery {
             hybrid,
             ranking_score_threshold: other.ranking_score_threshold.map(|o| o.0),
             locales: other.locales.map(|o| o.into_iter().collect()),
+            personalize,
         })
     }
 }
@@ -321,6 +328,7 @@ pub fn fix_sort_query_parameters(sort_query: &str) -> Vec<String> {
 pub async fn search_with_url_query(
     index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
     search_queue: web::Data<SearchQueue>,
+    personalization_service: web::Data<crate::personalization::PersonalizationService>,
     index_uid: web::Path<String>,
     params: AwebQueryParameter<SearchQueryGet, DeserrQueryParamError>,
     req: HttpRequest,
@@ -341,30 +349,54 @@ pub async fn search_with_url_query(
 
     let index = index_scheduler.index(&index_uid)?;
 
+    // Extract personalization and query string before moving query
+    let personalize = query.personalize.take();
+
     let search_kind =
         search_kind(&query, index_scheduler.get_ref(), index_uid.to_string(), &index)?;
     let retrieve_vector = RetrieveVectors::new(query.retrieve_vectors);
+
+    // Save the query string for personalization if requested
+    let personalize_query = personalize.is_some().then(|| query.q.clone()).flatten();
+
     let permit = search_queue.try_get_search_permit().await?;
+    let include_metadata = parse_include_metadata_header(&req);
+
     let search_result = tokio::task::spawn_blocking(move || {
         perform_search(
-            index_uid.to_string(),
+            SearchParams {
+                index_uid: index_uid.to_string(),
+                query,
+                search_kind,
+                retrieve_vectors: retrieve_vector,
+                features: index_scheduler.features(),
+                request_uid,
+                include_metadata,
+            },
             &index,
-            query,
-            search_kind,
-            retrieve_vector,
-            index_scheduler.features(),
-            request_uid,
         )
     })
     .await;
     permit.drop().await;
     let search_result = search_result?;
-    if let Ok(ref search_result) = search_result {
+    if let Ok((search_result, _)) = search_result.as_ref() {
         aggregate.succeed(search_result);
     }
     analytics.publish(aggregate, &req);
 
-    let search_result = search_result?;
+    let (mut search_result, time_budget) = search_result?;
+
+    // Apply personalization if requested
+    if let Some(personalize) = personalize.as_ref() {
+        search_result = personalization_service
+            .rerank_search_results(
+                search_result,
+                personalize,
+                personalize_query.as_deref(),
+                time_budget,
+            )
+            .await?;
+    }
 
     debug!(request_uid = ?request_uid, returns = ?search_result, "Search get");
     Ok(HttpResponse::Ok().json(search_result))
@@ -429,6 +461,7 @@ pub async fn search_with_url_query(
 pub async fn search_with_post(
     index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
     search_queue: web::Data<SearchQueue>,
+    personalization_service: web::Data<crate::personalization::PersonalizationService>,
     index_uid: web::Path<String>,
     params: AwebJson<SearchQuery, DeserrJsonError>,
     req: HttpRequest,
@@ -449,26 +482,37 @@ pub async fn search_with_post(
 
     let index = index_scheduler.index(&index_uid)?;
 
+    // Extract personalization and query string before moving query
+    let personalize = query.personalize.take();
+
     let search_kind =
         search_kind(&query, index_scheduler.get_ref(), index_uid.to_string(), &index)?;
     let retrieve_vectors = RetrieveVectors::new(query.retrieve_vectors);
 
+    let include_metadata = parse_include_metadata_header(&req);
+
+    // Save the query string for personalization if requested
+    let personalize_query = personalize.is_some().then(|| query.q.clone()).flatten();
+
     let permit = search_queue.try_get_search_permit().await?;
     let search_result = tokio::task::spawn_blocking(move || {
         perform_search(
-            index_uid.to_string(),
+            SearchParams {
+                index_uid: index_uid.to_string(),
+                query,
+                search_kind,
+                retrieve_vectors,
+                features: index_scheduler.features(),
+                request_uid,
+                include_metadata,
+            },
             &index,
-            query,
-            search_kind,
-            retrieve_vectors,
-            index_scheduler.features(),
-            request_uid,
         )
     })
     .await;
     permit.drop().await;
     let search_result = search_result?;
-    if let Ok(ref search_result) = search_result {
+    if let Ok((ref search_result, _)) = search_result {
         aggregate.succeed(search_result);
         if search_result.degraded {
             MEILISEARCH_DEGRADED_SEARCH_REQUESTS.inc();
@@ -476,7 +520,19 @@ pub async fn search_with_post(
     }
     analytics.publish(aggregate, &req);
 
-    let search_result = search_result?;
+    let (mut search_result, time_budget) = search_result?;
+
+    // Apply personalization if requested
+    if let Some(personalize) = personalize.as_ref() {
+        search_result = personalization_service
+            .rerank_search_results(
+                search_result,
+                personalize,
+                personalize_query.as_deref(),
+                time_budget,
+            )
+            .await?;
+    }
 
     debug!(request_uid = ?request_uid, returns = ?search_result, "Search post");
     Ok(HttpResponse::Ok().json(search_result))

crates/meilisearch/src/routes/indexes/search_analytics.rs

@@ -7,6 +7,7 @@ use serde_json::{json, Value};
 
 use crate::aggregate_methods;
 use crate::analytics::{Aggregate, AggregateMethod};
+use crate::metrics::MEILISEARCH_PERSONALIZED_SEARCH_REQUESTS;
 use crate::search::{
     SearchQuery, SearchResult, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER,
     DEFAULT_HIGHLIGHT_POST_TAG, DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT,
@@ -95,6 +96,9 @@ pub struct SearchAggregator<Method: AggregateMethod> {
     show_ranking_score_details: bool,
     ranking_score_threshold: bool,
 
+    // personalization
+    total_personalized: usize,
+
     marker: std::marker::PhantomData<Method>,
 }
 
@@ -129,6 +133,7 @@ impl<Method: AggregateMethod> SearchAggregator<Method> {
             hybrid,
             ranking_score_threshold,
             locales,
+            personalize,
         } = query;
 
         let mut ret = Self::default();
@@ -204,6 +209,12 @@ impl<Method: AggregateMethod> SearchAggregator<Method> {
             ret.locales = locales.iter().copied().collect();
         }
 
+        // personalization
+        if personalize.is_some() {
+            ret.total_personalized = 1;
+            MEILISEARCH_PERSONALIZED_SEARCH_REQUESTS.inc();
+        }
+
         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();
@@ -235,6 +246,7 @@ impl<Method: AggregateMethod> SearchAggregator<Method> {
             degraded,
             used_negative_operator,
             request_uid: _,
+            metadata: _,
         } = result;
 
         self.total_succeeded = self.total_succeeded.saturating_add(1);
@@ -295,6 +307,7 @@ impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
             total_used_negative_operator,
             ranking_score_threshold,
             mut locales,
+            total_personalized,
             marker: _,
         } = *new;
 
@@ -380,6 +393,9 @@ impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
         // locales
         self.locales.append(&mut locales);
 
+        // personalization
+        self.total_personalized = self.total_personalized.saturating_add(total_personalized);
+
         self
     }
 
@@ -425,6 +441,7 @@ impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
             total_used_negative_operator,
             ranking_score_threshold,
             locales,
+            total_personalized,
             marker: _,
         } = *self;
 
@@ -498,6 +515,9 @@ impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
                 "show_ranking_score_details": show_ranking_score_details,
                 "ranking_score_threshold": ranking_score_threshold,
             },
+            "personalization": {
+                "total_personalized": total_personalized,
+            },
         })
     }
 }

crates/meilisearch/src/routes/mod.rs

@@ -41,10 +41,13 @@ use crate::routes::indexes::IndexView;
 use crate::routes::multi_search::SearchResults;
 use crate::routes::network::{Network, Remote};
 use crate::routes::swap_indexes::SwapIndexesPayload;
-use crate::routes::webhooks::{WebhookResults, WebhookSettings, WebhookWithMetadata};
+use crate::routes::webhooks::{
+    WebhookResults, WebhookSettings, WebhookWithMetadataRedactedAuthorization,
+};
 use crate::search::{
     FederatedSearch, FederatedSearchResult, Federation, FederationOptions, MergeFacets,
     SearchQueryWithIndex, SearchResultWithIndex, SimilarQuery, SimilarResult,
+    INCLUDE_METADATA_HEADER,
 };
 use crate::search_queue::SearchQueue;
 use crate::Opt;
@@ -102,7 +105,7 @@ mod webhooks;
         url = "/",
         description = "Local server",
     )),
-    components(schemas(PaginationView<KeyView>, PaginationView<IndexView>, IndexView, DocumentDeletionByFilter, AllBatches, BatchStats, ProgressStepView, ProgressView, BatchView, RuntimeTogglableFeatures, SwapIndexesPayload, DocumentEditionByFunction, MergeFacets, FederationOptions, SearchQueryWithIndex, Federation, FederatedSearch, FederatedSearchResult, SearchResults, SearchResultWithIndex, SimilarQuery, SimilarResult, PaginationView<serde_json::Value>, BrowseQuery, UpdateIndexRequest, IndexUid, IndexCreateRequest, KeyView, Action, CreateApiKey, UpdateStderrLogs, LogMode, GetLogs, IndexStats, Stats, HealthStatus, HealthResponse, VersionResponse, Code, ErrorType, AllTasks, TaskView, Status, DetailsView, ResponseError, Settings<Unchecked>, Settings<Checked>, TypoSettings, MinWordSizeTyposSetting, FacetingSettings, PaginationSettings, SummarizedTaskView, Kind, Network, Remote, FilterableAttributesRule, FilterableAttributesPatterns, AttributePatterns, FilterableAttributesFeatures, FilterFeatures, Export, WebhookSettings, WebhookResults, WebhookWithMetadata, meilisearch_types::milli::vector::VectorStoreBackend))
+    components(schemas(PaginationView<KeyView>, PaginationView<IndexView>, IndexView, DocumentDeletionByFilter, AllBatches, BatchStats, ProgressStepView, ProgressView, BatchView, RuntimeTogglableFeatures, SwapIndexesPayload, DocumentEditionByFunction, MergeFacets, FederationOptions, SearchQueryWithIndex, Federation, FederatedSearch, FederatedSearchResult, SearchResults, SearchResultWithIndex, SimilarQuery, SimilarResult, PaginationView<serde_json::Value>, BrowseQuery, UpdateIndexRequest, IndexUid, IndexCreateRequest, KeyView, Action, CreateApiKey, UpdateStderrLogs, LogMode, GetLogs, IndexStats, Stats, HealthStatus, HealthResponse, VersionResponse, Code, ErrorType, AllTasks, TaskView, Status, DetailsView, ResponseError, Settings<Unchecked>, Settings<Checked>, TypoSettings, MinWordSizeTyposSetting, FacetingSettings, PaginationSettings, SummarizedTaskView, Kind, Network, Remote, FilterableAttributesRule, FilterableAttributesPatterns, AttributePatterns, FilterableAttributesFeatures, FilterFeatures, Export, WebhookSettings, WebhookResults, WebhookWithMetadataRedactedAuthorization, meilisearch_types::milli::vector::VectorStoreBackend))
 )]
 pub struct MeilisearchApi;
 
@@ -184,6 +187,18 @@ pub fn is_dry_run(req: &HttpRequest, opt: &Opt) -> Result<bool, ResponseError> {
         .is_some_and(|s| s.to_lowercase() == "true"))
 }
 
+/// Parse the `Meili-Include-Metadata` header from an HTTP request.
+///
+/// Returns `true` if the header is present and set to "true" or "1" (case-insensitive).
+/// Returns `false` if the header is not present or has any other value.
+pub fn parse_include_metadata_header(req: &HttpRequest) -> bool {
+    req.headers()
+        .get(INCLUDE_METADATA_HEADER)
+        .and_then(|h| h.to_str().ok())
+        .map(|v| matches!(v.to_lowercase().as_str(), "true" | "1"))
+        .unwrap_or(false)
+}
+
 #[derive(Debug, Serialize, Deserialize, ToSchema)]
 #[serde(rename_all = "camelCase")]
 pub struct SummarizedTaskView {

crates/meilisearch/src/routes/multi_search.rs

@@ -18,10 +18,11 @@ use crate::extractors::authentication::policies::ActionPolicy;
 use crate::extractors::authentication::{AuthenticationError, GuardedData};
 use crate::extractors::sequential_extractor::SeqHandler;
 use crate::routes::indexes::search::search_kind;
+use crate::routes::parse_include_metadata_header;
 use crate::search::{
     add_search_rules, perform_federated_search, perform_search, FederatedSearch,
-    FederatedSearchResult, RetrieveVectors, SearchQueryWithIndex, SearchResultWithIndex,
-    PROXY_SEARCH_HEADER, PROXY_SEARCH_HEADER_VALUE,
+    FederatedSearchResult, RetrieveVectors, SearchParams, SearchQueryWithIndex,
+    SearchResultWithIndex, PROXY_SEARCH_HEADER, PROXY_SEARCH_HEADER_VALUE,
 };
 use crate::search_queue::SearchQueue;
 
@@ -145,6 +146,7 @@ pub struct SearchResults {
 pub async fn multi_search_with_post(
     index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
     search_queue: Data<SearchQueue>,
+    personalization_service: web::Data<crate::personalization::PersonalizationService>,
     params: AwebJson<FederatedSearch, DeserrJsonError>,
     req: HttpRequest,
     analytics: web::Data<Analytics>,
@@ -188,6 +190,7 @@ pub async fn multi_search_with_post(
         err
     })?;
 
+    let include_metadata = parse_include_metadata_header(&req);
     let response = match federation {
         Some(federation) => {
             debug!(
@@ -209,6 +212,7 @@ pub async fn multi_search_with_post(
                 features,
                 is_proxy,
                 request_uid,
+                include_metadata,
             )
             .await;
             permit.drop().await;
@@ -233,7 +237,7 @@ pub async fn multi_search_with_post(
             // changes.
             let search_results: Result<_, (ResponseError, usize)> = async {
                 let mut search_results = Vec::with_capacity(queries.len());
-                for (query_index, (index_uid, query, federation_options)) in queries
+                for (query_index, (index_uid, mut query, federation_options)) in queries
                     .into_iter()
                     .map(SearchQueryWithIndex::into_index_query_federation)
                     .enumerate()
@@ -266,6 +270,13 @@ pub async fn multi_search_with_post(
                         })
                         .with_index(query_index)?;
 
+                    // Extract personalization and query string before moving query
+                    let personalize = query.personalize.take();
+
+                    // Save the query string for personalization if requested
+                    let personalize_query =
+                        personalize.is_some().then(|| query.q.clone()).flatten();
+
                     let index_uid_str = index_uid.to_string();
 
                     let search_kind = search_kind(
@@ -277,23 +288,40 @@ pub async fn multi_search_with_post(
                     .with_index(query_index)?;
                     let retrieve_vector = RetrieveVectors::new(query.retrieve_vectors);
 
-                    let search_result = tokio::task::spawn_blocking(move || {
+                    let (mut search_result, time_budget) = tokio::task::spawn_blocking(move || {
                         perform_search(
-                            index_uid_str.clone(),
+                            SearchParams {
+                                index_uid: index_uid_str.clone(),
+                                query,
+                                search_kind,
+                                retrieve_vectors: retrieve_vector,
+                                features,
+                                request_uid,
+                                include_metadata,
+                            },
                             &index,
-                            query,
-                            search_kind,
-                            retrieve_vector,
-                            features,
-                            request_uid,
                         )
                     })
                     .await
+                    .with_index(query_index)?
                     .with_index(query_index)?;
 
+                    // Apply personalization if requested
+                    if let Some(personalize) = personalize.as_ref() {
+                        search_result = personalization_service
+                            .rerank_search_results(
+                                search_result,
+                                personalize,
+                                personalize_query.as_deref(),
+                                time_budget,
+                            )
+                            .await
+                            .with_index(query_index)?;
+                    }
+
                     search_results.push(SearchResultWithIndex {
                         index_uid: index_uid.into_inner(),
-                        result: search_result.with_index(query_index)?,
+                        result: search_result,
                     });
                 }
                 Ok(search_results)

crates/meilisearch/src/routes/multi_search_analytics.rs

@@ -67,6 +67,7 @@ impl MultiSearchAggregator {
             hybrid: _,
             ranking_score_threshold: _,
             locales: _,
+            personalize: _,
         } in &federated_search.queries
         {
             if let Some(federation_options) = federation_options {

crates/meilisearch/src/routes/webhooks.rs

@@ -90,7 +90,7 @@ fn deny_immutable_fields_webhook(
 #[derive(Debug, Serialize, ToSchema)]
 #[serde(rename_all = "camelCase")]
 #[schema(rename_all = "camelCase")]
-pub(super) struct WebhookWithMetadata {
+pub(super) struct WebhookWithMetadataRedactedAuthorization {
     uuid: Uuid,
     is_editable: bool,
     #[schema(value_type = WebhookSettings)]
@@ -98,8 +98,9 @@ pub(super) struct WebhookWithMetadata {
     webhook: Webhook,
 }
 
-impl WebhookWithMetadata {
-    pub fn from(uuid: Uuid, webhook: Webhook) -> Self {
+impl WebhookWithMetadataRedactedAuthorization {
+    pub fn from(uuid: Uuid, mut webhook: Webhook) -> Self {
+        webhook.redact_authorization_header();
         Self { uuid, is_editable: uuid != Uuid::nil(), webhook }
     }
 }
@@ -107,7 +108,7 @@ impl WebhookWithMetadata {
 #[derive(Debug, Serialize, ToSchema)]
 #[serde(rename_all = "camelCase")]
 pub(super) struct WebhookResults {
-    results: Vec<WebhookWithMetadata>,
+    results: Vec<WebhookWithMetadataRedactedAuthorization>,
 }
 
 #[utoipa::path(
@@ -150,7 +151,7 @@ async fn get_webhooks(
     let results = webhooks
         .webhooks
         .into_iter()
-        .map(|(uuid, webhook)| WebhookWithMetadata::from(uuid, webhook))
+        .map(|(uuid, webhook)| WebhookWithMetadataRedactedAuthorization::from(uuid, webhook))
         .collect::<Vec<_>>();
     let results = WebhookResults { results };
 
@@ -301,7 +302,7 @@ fn check_changed(uuid: Uuid, webhook: &Webhook) -> Result<(), WebhooksError> {
     tag = "Webhooks",
     security(("Bearer" = ["webhooks.get", "webhooks.*", "*.get", "*"])),
     responses(
-        (status = 200, description = "Webhook found", body = WebhookWithMetadata, content_type = "application/json", example = json!({
+        (status = 200, description = "Webhook found", body = WebhookWithMetadataRedactedAuthorization, content_type = "application/json", example = json!({
             "uuid": "550e8400-e29b-41d4-a716-446655440000",
             "url": "https://your.site/on-tasks-completed",
             "headers": {
@@ -324,7 +325,7 @@ async fn get_webhook(
     let mut webhooks = index_scheduler.webhooks_view();
 
     let webhook = webhooks.webhooks.remove(&uuid).ok_or(WebhookNotFound(uuid))?;
-    let webhook = WebhookWithMetadata::from(uuid, webhook);
+    let webhook = WebhookWithMetadataRedactedAuthorization::from(uuid, webhook);
 
     debug!(returns = ?webhook, "Get webhook");
     Ok(HttpResponse::Ok().json(webhook))
@@ -337,7 +338,7 @@ async fn get_webhook(
     request_body = WebhookSettings,
     security(("Bearer" = ["webhooks.create", "webhooks.*", "*"])),
     responses(
-        (status = 201, description = "Webhook created successfully", body = WebhookWithMetadata, content_type = "application/json", example = json!({
+        (status = 201, description = "Webhook created successfully", body = WebhookWithMetadataRedactedAuthorization, content_type = "application/json", example = json!({
             "uuid": "550e8400-e29b-41d4-a716-446655440000",
             "url": "https://your.site/on-tasks-completed",
             "headers": {
@@ -383,7 +384,7 @@ async fn post_webhook(
 
     analytics.publish(PostWebhooksAnalytics, &req);
 
-    let response = WebhookWithMetadata::from(uuid, webhook);
+    let response = WebhookWithMetadataRedactedAuthorization::from(uuid, webhook);
     debug!(returns = ?response, "Post webhook");
     Ok(HttpResponse::Created().json(response))
 }
@@ -395,7 +396,7 @@ async fn post_webhook(
     request_body = WebhookSettings,
     security(("Bearer" = ["webhooks.update", "webhooks.*", "*"])),
     responses(
-        (status = 200, description = "Webhook updated successfully", body = WebhookWithMetadata, content_type = "application/json", example = json!({
+        (status = 200, description = "Webhook updated successfully", body = WebhookWithMetadataRedactedAuthorization, content_type = "application/json", example = json!({
             "uuid": "550e8400-e29b-41d4-a716-446655440000",
             "url": "https://your.site/on-tasks-completed",
             "headers": {
@@ -435,7 +436,7 @@ async fn patch_webhook(
 
     analytics.publish(PatchWebhooksAnalytics, &req);
 
-    let response = WebhookWithMetadata::from(uuid, webhook);
+    let response = WebhookWithMetadataRedactedAuthorization::from(uuid, webhook);
     debug!(returns = ?response, "Patch webhook");
     Ok(HttpResponse::Ok().json(response))
 }

crates/meilisearch/src/search/federated/perform.rs

@@ -22,7 +22,8 @@ use uuid::Uuid;
 use super::super::ranking_rules::{self, RankingRules};
 use super::super::{
     compute_facet_distribution_stats, prepare_search, AttributesFormat, ComputedFacets, HitMaker,
-    HitsInfo, RetrieveVectors, SearchHit, SearchKind, SearchQuery, SearchQueryWithIndex,
+    HitsInfo, RetrieveVectors, SearchHit, SearchKind, SearchMetadata, SearchQuery,
+    SearchQueryWithIndex,
 };
 use super::proxy::{proxy_search, ProxySearchError, ProxySearchParams};
 use super::types::{
@@ -41,6 +42,7 @@ pub async fn perform_federated_search(
     features: RoFeatures,
     is_proxy: bool,
     request_uid: Uuid,
+    include_metadata: bool,
 ) -> Result<FederatedSearchResult, ResponseError> {
     if is_proxy {
         features.check_network("Performing a remote federated search")?;
@@ -59,20 +61,38 @@ pub async fn perform_federated_search(
 
     let network = index_scheduler.network();
 
+    // Preconstruct metadata keeping the original queries order for later metadata building
+    let precomputed_query_metadata: Option<Vec<_>> = include_metadata.then(|| {
+        queries
+            .iter()
+            .map(|q| {
+                (
+                    q.index_uid.to_string(),
+                    q.federation_options.as_ref().and_then(|o| o.remote.clone()),
+                )
+            })
+            .collect()
+    });
+
     // this implementation partition the queries by index to guarantee an important property:
     // - all the queries to a particular index use the same read transaction.
     // This is an important property, otherwise we cannot guarantee the self-consistency of the results.
 
     // 1. partition queries by host and index
     let mut partitioned_queries = PartitionedQueries::new();
+
     for (query_index, federated_query) in queries.into_iter().enumerate() {
         partitioned_queries.partition(federated_query, query_index, &network, features)?
     }
 
     // 2. perform queries, merge and make hits index by index
     // 2.1. start remote queries
-    let remote_search =
-        RemoteSearch::start(partitioned_queries.remote_queries_by_host, &federation, deadline);
+    let remote_search = RemoteSearch::start(
+        partitioned_queries.remote_queries_by_host,
+        &federation,
+        deadline,
+        include_metadata,
+    );
 
     // 2.2. concurrently execute local queries
     let params = SearchByIndexParams {
@@ -114,11 +134,25 @@ pub async fn perform_federated_search(
     let after_waiting_remote_results = std::time::Instant::now();
 
     // 3. merge hits and metadata across indexes and hosts
-    // 3.1. merge metadata
+
+    // 3.1. Build metadata in the same order as the original queries
+    let query_metadata = precomputed_query_metadata.map(|precomputed_query_metadata| {
+        // If a remote is present, set the local remote name
+        let local_remote_name = network.local.clone().filter(|_| partitioned_queries.has_remote);
+
+        build_query_metadata(
+            precomputed_query_metadata,
+            local_remote_name,
+            &remote_results,
+            &results_by_index,
+        )
+    });
+
+    // 3.2. merge federation metadata
     let (estimated_total_hits, degraded, used_negative_operator, facets, max_remote_duration) =
         merge_metadata(&mut results_by_index, &remote_results);
 
-    // 3.2. merge hits
+    // 3.3. merge hits
     let merged_hits: Vec<_> = merge_index_global_results(results_by_index, &mut remote_results)
         .skip(federation.offset)
         .take(federation.limit)
@@ -133,7 +167,7 @@ pub async fn perform_federated_search(
         .map(|hit| hit.hit())
         .collect();
 
-    // 3.3. merge query vectors
+    // 3.4. merge query vectors
     let query_vectors = if retrieve_vectors {
         for remote_results in remote_results.iter_mut() {
             if let Some(remote_vectors) = remote_results.query_vectors.take() {
@@ -152,7 +186,7 @@ pub async fn perform_federated_search(
         None
     };
 
-    // 3.4. merge facets
+    // 3.5. merge facets
     let (facet_distribution, facet_stats, facets_by_index) =
         facet_order.merge(federation.merge_facets, remote_results, facets);
 
@@ -179,6 +213,7 @@ pub async fn perform_federated_search(
         facets_by_index,
         remote_errors: partitioned_queries.has_remote.then_some(remote_errors),
         request_uid: Some(request_uid),
+        metadata: query_metadata,
     })
 }
 
@@ -402,6 +437,7 @@ struct SearchHitByIndex {
 
 struct SearchResultByIndex {
     index: String,
+    primary_key: Option<String>,
     hits: Vec<SearchHitByIndex>,
     estimated_total_hits: usize,
     degraded: bool,
@@ -409,6 +445,61 @@ struct SearchResultByIndex {
     facets: Option<ComputedFacets>,
 }
 
+/// Builds query metadata for federated search results.
+///
+/// This function creates metadata for each query in the same order as the original queries,
+/// combining information from both local and remote search results. It handles the mapping
+/// of primary keys to their respective indexes and remotes to prevent collisions when
+/// multiple remotes have the same index_uid but different primary keys.
+fn build_query_metadata(
+    precomputed_query_metadata: Vec<(String, Option<String>)>,
+    local_remote_name: Option<String>,
+    remote_results: &[FederatedSearchResult],
+    results_by_index: &[SearchResultByIndex],
+) -> Vec<SearchMetadata> {
+    // Create a map of (remote, index_uid) -> primary_key for quick lookup
+    // This prevents collisions when multiple remotes have the same index_uid but different primary keys
+    let mut primary_key_per_index = std::collections::HashMap::new();
+
+    // Build metadata for remote results
+    for remote_result in remote_results {
+        if let Some(remote_metadata) = &remote_result.metadata {
+            for remote_meta in remote_metadata {
+                if let SearchMetadata {
+                    remote: Some(remote_name),
+                    index_uid,
+                    primary_key: Some(primary_key),
+                    ..
+                } = remote_meta
+                {
+                    let key = (Some(remote_name), index_uid);
+                    primary_key_per_index.insert(key, primary_key);
+                }
+            }
+        }
+    }
+
+    // Build metadata for local results
+    for local_meta in results_by_index {
+        if let SearchResultByIndex { index, primary_key: Some(primary_key), .. } = local_meta {
+            let key = (None, index);
+            primary_key_per_index.insert(key, primary_key);
+        }
+    }
+
+    // Build metadata in the same order as the original queries
+    let mut query_metadata = Vec::new();
+    for (index_uid, remote) in precomputed_query_metadata {
+        let primary_key =
+            primary_key_per_index.get(&(remote.as_ref(), &index_uid)).map(|pk| pk.to_string());
+        let query_uid = Uuid::now_v7();
+        // if the remote is not set, use the local remote name
+        let remote = remote.or_else(|| local_remote_name.clone());
+        query_metadata.push(SearchMetadata { query_uid, primary_key, index_uid, remote });
+    }
+    query_metadata
+}
+
 fn merge_metadata(
     results_by_index: &mut Vec<SearchResultByIndex>,
     remote_results: &Vec<FederatedSearchResult>,
@@ -420,6 +511,7 @@ fn merge_metadata(
     let mut max_remote_duration = Duration::ZERO;
     for SearchResultByIndex {
         index,
+        primary_key: _,
         hits: _,
         estimated_total_hits: estimated_total_hits_by_index,
         facets: facets_by_index,
@@ -448,6 +540,7 @@ fn merge_metadata(
         degraded: degraded_for_host,
         used_negative_operator: host_used_negative_operator,
         remote_errors: _,
+        metadata: _,
         request_uid: _,
     } in remote_results
     {
@@ -508,6 +601,10 @@ impl PartitionedQueries {
             .into());
         }
 
+        if federated_query.has_personalize() {
+            return Err(MeilisearchHttpError::PersonalizationInFederatedQuery(query_index).into());
+        }
+
         let (index_uid, query, federation_options) = federated_query.into_index_query_federation();
 
         let federation_options = federation_options.unwrap_or_default();
@@ -576,7 +673,12 @@ struct RemoteSearch {
 }
 
 impl RemoteSearch {
-    fn start(queries: RemoteQueriesByHost, federation: &Federation, deadline: Instant) -> Self {
+    fn start(
+        queries: RemoteQueriesByHost,
+        federation: &Federation,
+        deadline: Instant,
+        include_metadata: bool,
+    ) -> Self {
         let mut in_flight_remote_queries = BTreeMap::new();
         let client = reqwest::ClientBuilder::new()
             .connect_timeout(std::time::Duration::from_millis(200))
@@ -596,7 +698,10 @@ impl RemoteSearch {
                     // never merge distant facets
                     proxy_federation.merge_facets = None;
                     let params = params.clone();
-                    async move { proxy_search(&node, queries, proxy_federation, &params).await }
+                    async move {
+                        proxy_search(&node, queries, proxy_federation, &params, include_metadata)
+                            .await
+                    }
                 }),
             );
         }
@@ -640,6 +745,13 @@ impl RemoteSearch {
                             continue 'remote_queries;
                         }
 
+                        // Add remote name to metadata
+                        if let Some(metadata) = res.metadata.as_mut() {
+                            for meta in metadata {
+                                meta.remote = Some(node_name.clone());
+                            }
+                        }
+
                         federation.insert(
                             FEDERATION_REMOTE.to_string(),
                             serde_json::Value::String(node_name.clone()),
@@ -735,6 +847,7 @@ impl SearchByIndex {
             }
         };
         let rtxn = index.read_txn()?;
+        let primary_key = index.primary_key(&rtxn)?.map(|pk| pk.to_string());
         let criteria = index.criteria(&rtxn)?;
         let dictionary = index.dictionary(&rtxn)?;
         let dictionary: Option<Vec<_>> =
@@ -761,6 +874,12 @@ impl SearchByIndex {
             return Err(error);
         }
         let mut results_by_query = Vec::with_capacity(queries.len());
+
+        // all queries for an index share the same budget
+        let time_budget = match cutoff {
+            Some(cutoff) => TimeBudget::new(Duration::from_millis(cutoff)),
+            None => TimeBudget::default(),
+        };
         for QueryByIndex { query, weight, query_index } in queries {
             // use an immediately invoked lambda to capture the result without returning from the function
 
@@ -830,17 +949,13 @@ impl SearchByIndex {
 
                 let retrieve_vectors = RetrieveVectors::new(query.retrieve_vectors);
 
-                let time_budget = match cutoff {
-                    Some(cutoff) => TimeBudget::new(Duration::from_millis(cutoff)),
-                    None => TimeBudget::default(),
-                };
-
                 let (mut search, _is_finite_pagination, _max_total_hits, _offset) = prepare_search(
                     &index,
                     &rtxn,
                     &query,
                     &search_kind,
-                    time_budget,
+                    // clones of `TimeBudget` share the budget rather than restart it
+                    time_budget.clone(),
                     params.features,
                 )?;
 
@@ -987,6 +1102,7 @@ impl SearchByIndex {
             })?;
         self.results_by_index.push(SearchResultByIndex {
             index: index_uid,
+            primary_key,
             hits: merged_result,
             estimated_total_hits,
             degraded,

crates/meilisearch/src/search/federated/proxy.rs

@@ -7,7 +7,7 @@ use serde::de::DeserializeOwned;
 use serde_json::Value;
 
 use super::types::{FederatedSearch, FederatedSearchResult, Federation};
-use crate::search::SearchQueryWithIndex;
+use crate::search::{SearchQueryWithIndex, INCLUDE_METADATA_HEADER};
 
 pub const PROXY_SEARCH_HEADER: &str = "Meili-Proxy-Search";
 pub const PROXY_SEARCH_HEADER_VALUE: &str = "true";
@@ -98,6 +98,7 @@ pub async fn proxy_search(
     queries: Vec<SearchQueryWithIndex>,
     federation: Federation,
     params: &ProxySearchParams,
+    include_metadata: bool,
 ) -> Result<FederatedSearchResult, ProxySearchError> {
     let url = format!("{}/multi-search", node.url);
 
@@ -119,7 +120,16 @@ pub async fn proxy_search(
     };
 
     for i in 0..params.try_count {
-        match try_proxy_search(&url, search_api_key, &federated, &params.client, deadline).await {
+        match try_proxy_search(
+            &url,
+            search_api_key,
+            &federated,
+            &params.client,
+            deadline,
+            include_metadata,
+        )
+        .await
+        {
             Ok(response) => return Ok(response),
             Err(retry) => {
                 let duration = retry.into_duration(i)?;
@@ -127,7 +137,7 @@ pub async fn proxy_search(
             }
         }
     }
-    try_proxy_search(&url, search_api_key, &federated, &params.client, deadline)
+    try_proxy_search(&url, search_api_key, &federated, &params.client, deadline, include_metadata)
         .await
         .map_err(Retry::into_error)
 }
@@ -138,6 +148,7 @@ async fn try_proxy_search(
     federated: &FederatedSearch,
     client: &Client,
     deadline: std::time::Instant,
+    include_metadata: bool,
 ) -> Result<FederatedSearchResult, Retry> {
     let timeout = deadline.saturating_duration_since(std::time::Instant::now());
 
@@ -148,6 +159,8 @@ async fn try_proxy_search(
         request
     };
     let request = request.header(PROXY_SEARCH_HEADER, PROXY_SEARCH_HEADER_VALUE);
+    let request =
+        if include_metadata { request.header(INCLUDE_METADATA_HEADER, "true") } else { request };
 
     let response = request.send().await;
     let response = match response {

crates/meilisearch/src/search/federated/types.rs

@@ -18,6 +18,8 @@ use serde::{Deserialize, Serialize};
 use utoipa::ToSchema;
 use uuid::Uuid;
 
+use crate::search::SearchMetadata;
+
 use super::super::{ComputedFacets, FacetStats, HitsInfo, SearchHit, SearchQueryWithIndex};
 use crate::milli::vector::Embedding;
 
@@ -134,6 +136,8 @@ pub struct FederatedSearchResult {
     pub facets_by_index: FederatedFacets,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub request_uid: Option<Uuid>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub metadata: Option<Vec<SearchMetadata>>,
 
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub remote_errors: Option<BTreeMap<String, ResponseError>>,
@@ -160,6 +164,7 @@ impl fmt::Debug for FederatedSearchResult {
             facets_by_index,
             remote_errors,
             request_uid,
+            metadata,
         } = self;
 
         let mut debug = f.debug_struct("SearchResult");
@@ -195,6 +200,9 @@ impl fmt::Debug for FederatedSearchResult {
         if let Some(request_uid) = request_uid {
             debug.field("request_uid", &request_uid);
         }
+        if let Some(metadata) = metadata {
+            debug.field("metadata", &metadata);
+        }
 
         debug.finish()
     }

crates/meilisearch/src/search/mod.rs

@@ -57,6 +57,14 @@ pub const DEFAULT_CROP_MARKER: fn() -> String = || "…".to_string();
 pub const DEFAULT_HIGHLIGHT_PRE_TAG: fn() -> String = || "<em>".to_string();
 pub const DEFAULT_HIGHLIGHT_POST_TAG: fn() -> String = || "</em>".to_string();
 pub const DEFAULT_SEMANTIC_RATIO: fn() -> SemanticRatio = || SemanticRatio(0.5);
+pub const INCLUDE_METADATA_HEADER: &str = "Meili-Include-Metadata";
+
+#[derive(Clone, Default, PartialEq, Deserr, ToSchema, Debug)]
+#[deserr(error = DeserrJsonError<InvalidSearchPersonalize>, rename_all = camelCase, deny_unknown_fields)]
+pub struct Personalize {
+    #[deserr(error = DeserrJsonError<InvalidSearchPersonalizeUserContext>)]
+    pub user_context: String,
+}
 
 #[derive(Clone, Default, PartialEq, Deserr, ToSchema)]
 #[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
@@ -121,6 +129,8 @@ pub struct SearchQuery {
     pub ranking_score_threshold: Option<RankingScoreThreshold>,
     #[deserr(default, error = DeserrJsonError<InvalidSearchLocales>)]
     pub locales: Option<Vec<Locale>>,
+    #[deserr(default, error = DeserrJsonError<InvalidSearchPersonalize>, default)]
+    pub personalize: Option<Personalize>,
 }
 
 impl From<SearchParameters> for SearchQuery {
@@ -168,6 +178,7 @@ impl From<SearchParameters> for SearchQuery {
             highlight_post_tag: DEFAULT_HIGHLIGHT_POST_TAG(),
             crop_marker: DEFAULT_CROP_MARKER(),
             locales: None,
+            personalize: None,
         }
     }
 }
@@ -249,6 +260,7 @@ impl fmt::Debug for SearchQuery {
             attributes_to_search_on,
             ranking_score_threshold,
             locales,
+            personalize,
         } = self;
 
         let mut debug = f.debug_struct("SearchQuery");
@@ -337,6 +349,10 @@ impl fmt::Debug for SearchQuery {
             debug.field("locales", &locales);
         }
 
+        if let Some(personalize) = personalize {
+            debug.field("personalize", &personalize);
+        }
+
         debug.finish()
     }
 }
@@ -542,6 +558,9 @@ pub struct SearchQueryWithIndex {
     pub ranking_score_threshold: Option<RankingScoreThreshold>,
     #[deserr(default, error = DeserrJsonError<InvalidSearchLocales>, default)]
     pub locales: Option<Vec<Locale>>,
+    #[deserr(default, error = DeserrJsonError<InvalidSearchPersonalize>, default)]
+    #[serde(skip)]
+    pub personalize: Option<Personalize>,
 
     #[deserr(default)]
     pub federation_options: Option<FederationOptions>,
@@ -566,6 +585,10 @@ impl SearchQueryWithIndex {
         self.facets.as_deref().filter(|v| !v.is_empty())
     }
 
+    pub fn has_personalize(&self) -> bool {
+        self.personalize.is_some()
+    }
+
     pub fn from_index_query_federation(
         index_uid: IndexUid,
         query: SearchQuery,
@@ -599,6 +622,7 @@ impl SearchQueryWithIndex {
             attributes_to_search_on,
             ranking_score_threshold,
             locales,
+            personalize,
         } = query;
 
         SearchQueryWithIndex {
@@ -630,6 +654,7 @@ impl SearchQueryWithIndex {
             attributes_to_search_on,
             ranking_score_threshold,
             locales,
+            personalize,
             federation_options,
         }
     }
@@ -665,6 +690,7 @@ impl SearchQueryWithIndex {
             hybrid,
             ranking_score_threshold,
             locales,
+            personalize,
         } = self;
         (
             index_uid,
@@ -696,6 +722,7 @@ impl SearchQueryWithIndex {
                 hybrid,
                 ranking_score_threshold,
                 locales,
+                personalize,
                 // do not use ..Default::default() here,
                 // rather add any missing field from `SearchQuery` to `SearchQueryWithIndex`
             },
@@ -836,6 +863,18 @@ pub struct SearchHit {
     pub ranking_score_details: Option<serde_json::Map<String, serde_json::Value>>,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, ToSchema)]
+#[serde(rename_all = "camelCase")]
+#[schema(rename_all = "camelCase")]
+pub struct SearchMetadata {
+    pub query_uid: Uuid,
+    pub index_uid: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub primary_key: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub remote: Option<String>,
+}
+
 #[derive(Serialize, Clone, PartialEq, ToSchema)]
 #[serde(rename_all = "camelCase")]
 #[schema(rename_all = "camelCase")]
@@ -854,6 +893,8 @@ pub struct SearchResult {
     pub facet_stats: Option<BTreeMap<String, FacetStats>>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub request_uid: Option<Uuid>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub metadata: Option<SearchMetadata>,
 
     #[serde(skip_serializing_if = "Option::is_none")]
     pub semantic_hit_count: Option<u32>,
@@ -876,6 +917,7 @@ impl fmt::Debug for SearchResult {
             facet_distribution,
             facet_stats,
             request_uid,
+            metadata,
             semantic_hit_count,
             degraded,
             used_negative_operator,
@@ -908,6 +950,9 @@ impl fmt::Debug for SearchResult {
         if let Some(request_uid) = request_uid {
             debug.field("request_uid", &request_uid);
         }
+        if let Some(metadata) = metadata {
+            debug.field("metadata", &metadata);
+        }
 
         debug.finish()
     }
@@ -1120,16 +1165,31 @@ pub fn prepare_search<'t>(
     Ok((search, is_finite_pagination, max_total_hits, offset))
 }
 
+pub struct SearchParams {
+    pub index_uid: String,
+    pub query: SearchQuery,
+    pub search_kind: SearchKind,
+    pub retrieve_vectors: RetrieveVectors,
+    pub features: RoFeatures,
+    pub request_uid: Uuid,
+    pub include_metadata: bool,
+}
+
 pub fn perform_search(
-    index_uid: String,
+    params: SearchParams,
     index: &Index,
-    query: SearchQuery,
-    search_kind: SearchKind,
-    retrieve_vectors: RetrieveVectors,
-    features: RoFeatures,
-    request_uid: Uuid,
-) -> Result<SearchResult, ResponseError> {
+) -> Result<(SearchResult, TimeBudget), ResponseError> {
+    let SearchParams {
+        index_uid,
+        query,
+        search_kind,
+        retrieve_vectors,
+        features,
+        request_uid,
+        include_metadata,
+    } = params;
     let before_search = Instant::now();
+    let index_uid_for_metadata = index_uid.clone();
     let rtxn = index.read_txn()?;
     let time_budget = match index.search_cutoff(&rtxn)? {
         Some(cutoff) => TimeBudget::new(Duration::from_millis(cutoff)),
@@ -1137,7 +1197,7 @@ pub fn perform_search(
     };
 
     let (search, is_finite_pagination, max_total_hits, offset) =
-        prepare_search(index, &rtxn, &query, &search_kind, time_budget, features)?;
+        prepare_search(index, &rtxn, &query, &search_kind, time_budget.clone(), features)?;
 
     let (
         milli::SearchResult {
@@ -1150,7 +1210,20 @@ pub fn perform_search(
             query_vector,
         },
         semantic_hit_count,
-    ) = search_from_kind(index_uid, search_kind, search)?;
+    ) = search_from_kind(index_uid.clone(), search_kind, search)?;
+
+    let metadata = if include_metadata {
+        let query_uid = Uuid::now_v7();
+        let primary_key = index.primary_key(&rtxn)?.map(|pk| pk.to_string());
+        Some(SearchMetadata {
+            query_uid,
+            index_uid: index_uid_for_metadata,
+            primary_key,
+            remote: None, // Local searches don't have a remote
+        })
+    } else {
+        None
+    };
 
     let SearchQuery {
         q,
@@ -1182,6 +1255,7 @@ pub fn perform_search(
         attributes_to_search_on: _,
         filter: _,
         distinct: _,
+        personalize: _,
     } = query;
 
     let format = AttributesFormat {
@@ -1233,7 +1307,6 @@ pub fn perform_search(
         .transpose()?
         .map(|ComputedFacets { distribution, stats }| (distribution, stats))
         .unzip();
-
     let result = SearchResult {
         hits: documents,
         hits_info,
@@ -1246,8 +1319,9 @@ pub fn perform_search(
         used_negative_operator,
         semantic_hit_count,
         request_uid: Some(request_uid),
+        metadata,
     };
-    Ok(result)
+    Ok((result, time_budget))
 }
 
 #[derive(Debug, Clone, Default, Serialize, Deserialize, ToSchema)]

crates/meilisearch/tests/common/index.rs

@@ -516,6 +516,18 @@ impl<State> Index<'_, State> {
         self.service.post_encoded(url, query, self.encoder).await
     }
 
+    pub async fn search_with_headers(
+        &self,
+        query: Value,
+        headers: Vec<(&str, &str)>,
+    ) -> (Value, StatusCode) {
+        let url = format!("/indexes/{}/search", urlencode(self.uid.as_ref()));
+        let body = serde_json::to_string(&query).unwrap();
+        let mut all_headers = vec![("content-type", "application/json")];
+        all_headers.extend(headers);
+        self.service.post_str(url, body, all_headers).await
+    }
+
     pub async fn search_get(&self, query: &str) -> (Value, StatusCode) {
         let url = format!("/indexes/{}/search{}", urlencode(self.uid.as_ref()), query);
         self.service.get(url).await

crates/meilisearch/tests/common/server.rs

@@ -49,8 +49,8 @@ impl Server<Owned> {
         }
 
         let options = default_settings(dir.path());
-
-        let (index_scheduler, auth) = setup_meilisearch(&options).unwrap();
+        let handle = tokio::runtime::Handle::current();
+        let (index_scheduler, auth) = setup_meilisearch(&options, handle).unwrap();
         let service = Service { index_scheduler, auth, options, api_key: None };
 
         Server { service, _dir: Some(dir), _marker: PhantomData }
@@ -65,7 +65,9 @@ impl Server<Owned> {
 
         options.master_key = Some("MASTER_KEY".to_string());
 
-        let (index_scheduler, auth) = setup_meilisearch(&options).unwrap();
+        let handle = tokio::runtime::Handle::current();
+
+        let (index_scheduler, auth) = setup_meilisearch(&options, handle).unwrap();
         let service = Service { index_scheduler, auth, options, api_key: None };
 
         Server { service, _dir: Some(dir), _marker: PhantomData }
@@ -78,7 +80,9 @@ impl Server<Owned> {
     }
 
     pub async fn new_with_options(options: Opt) -> Result<Self, anyhow::Error> {
-        let (index_scheduler, auth) = setup_meilisearch(&options)?;
+        let handle = tokio::runtime::Handle::current();
+
+        let (index_scheduler, auth) = setup_meilisearch(&options, handle)?;
         let service = Service { index_scheduler, auth, options, api_key: None };
 
         Ok(Server { service, _dir: None, _marker: PhantomData })
@@ -217,8 +221,9 @@ impl Server<Shared> {
         }
 
         let options = default_settings(dir.path());
+        let handle = tokio::runtime::Handle::current();
 
-        let (index_scheduler, auth) = setup_meilisearch(&options).unwrap();
+        let (index_scheduler, auth) = setup_meilisearch(&options, handle).unwrap();
         let service = Service { index_scheduler, auth, api_key: None, options };
 
         Server { service, _dir: Some(dir), _marker: PhantomData }
@@ -390,6 +395,17 @@ impl<State> Server<State> {
         self.service.post("/multi-search", queries).await
     }
 
+    pub async fn multi_search_with_headers(
+        &self,
+        queries: Value,
+        headers: Vec<(&str, &str)>,
+    ) -> (Value, StatusCode) {
+        let body = serde_json::to_string(&queries).unwrap();
+        let mut all_headers = vec![("content-type", "application/json")];
+        all_headers.extend(headers);
+        self.service.post_str("/multi-search", body, all_headers).await
+    }
+
     pub async fn list_indexes_raw(&self, parameters: &str) -> (Value, StatusCode) {
         self.service.get(format!("/indexes{parameters}")).await
     }

crates/meilisearch/tests/common/service.rs

@@ -10,8 +10,9 @@ use actix_web::test::TestRequest;
 use actix_web::web::Data;
 use index_scheduler::IndexScheduler;
 use meilisearch::analytics::Analytics;
+use meilisearch::personalization::PersonalizationService;
 use meilisearch::search_queue::SearchQueue;
-use meilisearch::{create_app, Opt, SubscriberForSecondLayer};
+use meilisearch::{create_app, Opt, ServicesData, SubscriberForSecondLayer};
 use meilisearch_auth::AuthController;
 use tracing::level_filters::LevelFilter;
 use tracing_subscriber::Layer;
@@ -135,14 +136,24 @@ impl Service {
             self.options.experimental_search_queue_size,
             NonZeroUsize::new(1).unwrap(),
         );
+        let personalization_service = self
+            .options
+            .experimental_personalization_api_key
+            .clone()
+            .map(PersonalizationService::cohere)
+            .unwrap_or_else(PersonalizationService::disabled);
 
         actix_web::test::init_service(create_app(
-            self.index_scheduler.clone().into(),
-            self.auth.clone().into(),
-            Data::new(search_queue),
+            ServicesData {
+                index_scheduler: self.index_scheduler.clone().into(),
+                auth: self.auth.clone().into(),
+                search_queue: Data::new(search_queue),
+                personalization_service: Data::new(personalization_service),
+                logs_route_handle: Data::new(route_layer_handle),
+                logs_stderr_handle: Data::new(stderr_layer_handle),
+                analytics: Data::new(Analytics::no_analytics()),
+            },
             self.options.clone(),
-            (route_layer_handle, stderr_layer_handle),
-            Data::new(Analytics::no_analytics()),
             true,
         ))
         .await

crates/meilisearch/tests/features/mod.rs

@@ -207,3 +207,118 @@ async fn errors() {
     }
     "###);
 }
+
+#[actix_rt::test]
+async fn search_with_personalization_without_enabling_the_feature() {
+    let server = Server::new().await;
+    let index = server.unique_index();
+
+    // Create the index and add some documents
+    let (task, _code) = index.create(None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    let (task, _code) = index
+        .add_documents(
+            json!([
+                {"id": 1, "title": "The Dark Knight", "genre": "Action"},
+                {"id": 2, "title": "Inception", "genre": "Sci-Fi"},
+                {"id": 3, "title": "The Matrix", "genre": "Sci-Fi"}
+            ]),
+            None,
+        )
+        .await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Try to search with personalization - should return feature_not_enabled error
+    let (response, code) = index
+        .search_post(json!({
+            "q": "movie",
+            "personalize": {
+                "userContext": "I love science fiction movies"
+            }
+        }))
+        .await;
+
+    meili_snap::snapshot!(code, @"400 Bad Request");
+    meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
+    {
+      "message": "reranking search results requires enabling the `personalization` experimental feature. See https://github.com/orgs/meilisearch/discussions/866",
+      "code": "feature_not_enabled",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+    }
+    "###);
+}
+
+#[actix_rt::test]
+async fn multi_search_with_personalization_without_enabling_the_feature() {
+    let server = Server::new().await;
+    let index = server.unique_index();
+
+    // Create the index and add some documents
+    let (task, _code) = index.create(None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    let (task, _code) = index
+        .add_documents(
+            json!([
+                {"id": 1, "title": "The Dark Knight", "genre": "Action"},
+                {"id": 2, "title": "Inception", "genre": "Sci-Fi"},
+                {"id": 3, "title": "The Matrix", "genre": "Sci-Fi"}
+            ]),
+            None,
+        )
+        .await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Try to multi-search with personalization - should return feature_not_enabled error
+    let (response, code) = server
+        .multi_search(json!({
+            "queries": [
+                {
+                    "indexUid": index.uid,
+                    "q": "movie",
+                    "personalize": {
+                        "userContext": "I love science fiction movies"
+                    }
+                }
+            ]
+        }))
+        .await;
+
+    meili_snap::snapshot!(code, @"400 Bad Request");
+    meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
+    {
+      "message": "Inside `.queries[0]`: reranking search results requires enabling the `personalization` experimental feature. See https://github.com/orgs/meilisearch/discussions/866",
+      "code": "feature_not_enabled",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+    }
+    "###);
+
+    // Try to federated search with personalization - should return feature_not_enabled error
+    let (response, code) = server
+        .multi_search(json!({
+          "federation": {},
+            "queries": [
+                {
+                    "indexUid": index.uid,
+                    "q": "movie",
+                    "personalize": {
+                        "userContext": "I love science fiction movies"
+                    }
+                }
+            ]
+        }))
+        .await;
+
+    meili_snap::snapshot!(code, @"400 Bad Request");
+    meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
+    {
+      "message": "Inside `.queries[0]`: Using `.personalize` is not allowed in federated queries.\n - Hint: remove `personalize` from query #0 or remove `federation` from the request",
+      "code": "invalid_multi_search_query_personalization",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_multi_search_query_personalization"
+    }
+    "###);
+}

crates/meilisearch/tests/logs/mod.rs

@@ -8,8 +8,9 @@ use actix_web::http::header::ContentType;
 use actix_web::web::Data;
 use meili_snap::snapshot;
 use meilisearch::analytics::Analytics;
+use meilisearch::personalization::PersonalizationService;
 use meilisearch::search_queue::SearchQueue;
-use meilisearch::{create_app, Opt, SubscriberForSecondLayer};
+use meilisearch::{create_app, Opt, ServicesData, SubscriberForSecondLayer};
 use tracing::level_filters::LevelFilter;
 use tracing_subscriber::layer::SubscriberExt;
 use tracing_subscriber::Layer;
@@ -50,12 +51,16 @@ async fn basic_test_log_stream_route() {
     );
 
     let app = actix_web::test::init_service(create_app(
-        server.service.index_scheduler.clone().into(),
-        server.service.auth.clone().into(),
-        Data::new(search_queue),
+        ServicesData {
+            index_scheduler: server.service.index_scheduler.clone().into(),
+            auth: server.service.auth.clone().into(),
+            search_queue: Data::new(search_queue),
+            personalization_service: Data::new(PersonalizationService::disabled()),
+            logs_route_handle: Data::new(route_layer_handle),
+            logs_stderr_handle: Data::new(stderr_layer_handle),
+            analytics: Data::new(Analytics::no_analytics()),
+        },
         server.service.options.clone(),
-        (route_layer_handle, stderr_layer_handle),
-        Data::new(Analytics::no_analytics()),
         true,
     ))
     .await;

crates/meilisearch/tests/search/errors.rs

@@ -1,7 +1,11 @@
 use meili_snap::*;
+use meilisearch::Opt;
+use tempfile::TempDir;
 
 use super::test_settings_documents_indexing_swapping_and_search;
-use crate::common::{shared_does_not_exists_index, Server, DOCUMENTS, NESTED_DOCUMENTS};
+use crate::common::{
+    default_settings, shared_does_not_exists_index, Server, DOCUMENTS, NESTED_DOCUMENTS,
+};
 use crate::json;
 
 #[actix_rt::test]
@@ -1320,3 +1324,98 @@ async fn search_with_contains_without_enabling_the_feature() {
     }
     "#);
 }
+
+#[actix_rt::test]
+#[ignore]
+async fn search_with_personalization_invalid_api_key() {
+    // Create a server with a fake personalization API key
+    let dir = TempDir::new().unwrap();
+    let options = Opt {
+        experimental_personalization_api_key: Some("fake-api-key-12345".to_string()),
+        ..default_settings(dir.path())
+    };
+    let server = Server::new_with_options(options).await.unwrap();
+    let index = server.unique_index();
+
+    // Create the index and add some documents
+    let (task, _code) = index.create(None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    let (task, _code) = index
+        .add_documents(
+            json!([
+                {"id": 1, "title": "The Dark Knight", "genre": "Action"},
+                {"id": 2, "title": "Inception", "genre": "Sci-Fi"},
+                {"id": 3, "title": "The Matrix", "genre": "Sci-Fi"}
+            ]),
+            None,
+        )
+        .await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Try to search with personalization - should return remote_invalid_api_key error
+    let (response, code) = index
+        .search_post(json!({
+            "q": "the",
+            "personalize": {
+                "userContext": "I love science fiction movies"
+            }
+        }))
+        .await;
+
+    snapshot!(code, @"403 Forbidden");
+    snapshot!(json_string!(response), @r#"
+    {
+      "message": "Personalization service: Unauthorized: invalid API key",
+      "code": "remote_invalid_api_key",
+      "type": "auth",
+      "link": "https://docs.meilisearch.com/errors#remote_invalid_api_key"
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn search_with_personalization_no_user_context() {
+    // Create a server with a fake personalization API key
+    let dir = TempDir::new().unwrap();
+    let options = Opt {
+        experimental_personalization_api_key: Some("fake-api-key-12345".to_string()),
+        ..default_settings(dir.path())
+    };
+    let server = Server::new_with_options(options).await.unwrap();
+    let index = server.unique_index();
+
+    // Create the index and add some documents
+    let (task, _code) = index.create(None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    let (task, _code) = index
+        .add_documents(
+            json!([
+                {"id": 1, "title": "The Dark Knight", "genre": "Action"},
+                {"id": 2, "title": "Inception", "genre": "Sci-Fi"},
+                {"id": 3, "title": "The Matrix", "genre": "Sci-Fi"}
+            ]),
+            None,
+        )
+        .await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Try to search with personalization - should return remote_invalid_api_key error
+    let (response, code) = index
+        .search_post(json!({
+            "q": "the",
+            "personalize": {}
+        }))
+        .await;
+
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Missing field `userContext` inside `.personalize`",
+      "code": "invalid_search_personalize",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_personalize"
+    }
+    "###);
+}

crates/meilisearch/tests/search/metadata.rs

@@ -0,0 +1,387 @@
+use meili_snap::{json_string, snapshot};
+
+use crate::common::{shared_index_with_documents, Server, DOCUMENTS};
+use crate::json;
+
+#[actix_rt::test]
+async fn search_without_metadata_header() {
+    let index = shared_index_with_documents().await;
+
+    // Test that metadata is not included by default
+    index
+        .search(json!({"q": "glass"}), |response, code| {
+            snapshot!(code, @"200 OK");
+            snapshot!(json_string!(response, { ".processingTimeMs" => "[duration]", ".requestUid" => "[uuid]" }), @r###"
+            {
+              "hits": [
+                {
+                  "title": "Gläss",
+                  "id": "450465",
+                  "color": [
+                    "blue",
+                    "red"
+                  ]
+                }
+              ],
+              "query": "glass",
+              "processingTimeMs": "[duration]",
+              "limit": 20,
+              "offset": 0,
+              "estimatedTotalHits": 1,
+              "requestUid": "[uuid]"
+            }
+            "###);
+        })
+        .await;
+}
+
+#[actix_rt::test]
+async fn search_with_metadata_header() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    let documents = DOCUMENTS.clone();
+    let (task, _code) = index.add_documents(documents, None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Test with Meili-Include-Metadata header
+    let (response, code) = index
+        .search_with_headers(json!({"q": "glass"}), vec![("Meili-Include-Metadata", "true")])
+        .await;
+
+    snapshot!(code, @"200 OK");
+    snapshot!(json_string!(response, { ".processingTimeMs" => "[duration]", ".requestUid" => "[uuid]", ".metadata.queryUid" => "[uuid]" }), @r###"
+    {
+      "hits": [
+        {
+          "title": "Gläss",
+          "id": "450465",
+          "color": [
+            "blue",
+            "red"
+          ]
+        }
+      ],
+      "query": "glass",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 1,
+      "requestUid": "[uuid]",
+      "metadata": {
+        "queryUid": "[uuid]",
+        "indexUid": "[uuid]",
+        "primaryKey": "id"
+      }
+    }
+    "###);
+}
+
+#[actix_rt::test]
+async fn search_with_metadata_header_and_primary_key() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    let documents = DOCUMENTS.clone();
+    let (task, _code) = index.add_documents(documents, Some("id")).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Test with Meili-Include-Metadata header
+    let (response, code) = index
+        .search_with_headers(json!({"q": "glass"}), vec![("Meili-Include-Metadata", "true")])
+        .await;
+
+    snapshot!(code, @"200 OK");
+    snapshot!(json_string!(response, { ".processingTimeMs" => "[duration]", ".requestUid" => "[uuid]", ".metadata.queryUid" => "[uuid]" }), @r###"
+    {
+      "hits": [
+        {
+          "id": "450465",
+          "title": "Gläss",
+          "color": [
+            "blue",
+            "red"
+          ]
+        }
+      ],
+      "query": "glass",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 1,
+      "requestUid": "[uuid]",
+      "metadata": {
+        "queryUid": "[uuid]",
+        "indexUid": "[uuid]",
+        "primaryKey": "id"
+      }
+    }
+    "###);
+}
+
+#[actix_rt::test]
+async fn multi_search_without_metadata_header() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    let documents = DOCUMENTS.clone();
+    let (task, _code) = index.add_documents(documents, None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Test multi-search without metadata header
+    let (response, code) = server
+        .multi_search(json!({
+            "queries": [
+                {"indexUid": index.uid, "q": "glass"},
+                {"indexUid": index.uid, "q": "dragon"}
+            ]
+        }))
+        .await;
+
+    snapshot!(code, @"200 OK");
+    snapshot!(json_string!(response, { ".results[0].processingTimeMs" => "[duration]", ".results[0].requestUid" => "[uuid]", ".results[1].processingTimeMs" => "[duration]", ".results[1].requestUid" => "[uuid]" }), @r###"
+    {
+      "results": [
+        {
+          "indexUid": "[uuid]",
+          "hits": [
+            {
+              "title": "Gläss",
+              "id": "450465",
+              "color": [
+                "blue",
+                "red"
+              ]
+            }
+          ],
+          "query": "glass",
+          "processingTimeMs": "[duration]",
+          "limit": 20,
+          "offset": 0,
+          "estimatedTotalHits": 1,
+          "requestUid": "[uuid]"
+        },
+        {
+          "indexUid": "[uuid]",
+          "hits": [
+            {
+              "title": "How to Train Your Dragon: The Hidden World",
+              "id": "166428",
+              "color": [
+                "green",
+                "red"
+              ]
+            }
+          ],
+          "query": "dragon",
+          "processingTimeMs": "[duration]",
+          "limit": 20,
+          "offset": 0,
+          "estimatedTotalHits": 1,
+          "requestUid": "[uuid]"
+        }
+      ]
+    }
+    "###);
+}
+
+#[actix_rt::test]
+async fn multi_search_with_metadata_header() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    let documents = DOCUMENTS.clone();
+    let (task, _code) = index.add_documents(documents, Some("id")).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Test multi-search with metadata header
+    let (response, code) = server
+        .multi_search_with_headers(
+            json!({
+                "queries": [
+                    {"indexUid": index.uid, "q": "glass"},
+                    {"indexUid": index.uid, "q": "dragon"}
+                ]
+            }),
+            vec![("Meili-Include-Metadata", "true")],
+        )
+        .await;
+
+    snapshot!(code, @"200 OK");
+    snapshot!(json_string!(response, { ".results[0].processingTimeMs" => "[duration]", ".results[0].requestUid" => "[uuid]", ".results[0].metadata.queryUid" => "[uuid]", ".results[1].processingTimeMs" => "[duration]", ".results[1].requestUid" => "[uuid]", ".results[1].metadata.queryUid" => "[uuid]" }), @r###"
+    {
+      "results": [
+        {
+          "indexUid": "[uuid]",
+          "hits": [
+            {
+              "id": "450465",
+              "title": "Gläss",
+              "color": [
+                "blue",
+                "red"
+              ]
+            }
+          ],
+          "query": "glass",
+          "processingTimeMs": "[duration]",
+          "limit": 20,
+          "offset": 0,
+          "estimatedTotalHits": 1,
+          "requestUid": "[uuid]",
+          "metadata": {
+            "queryUid": "[uuid]",
+            "indexUid": "[uuid]",
+            "primaryKey": "id"
+          }
+        },
+        {
+          "indexUid": "[uuid]",
+          "hits": [
+            {
+              "id": "166428",
+              "title": "How to Train Your Dragon: The Hidden World",
+              "color": [
+                "green",
+                "red"
+              ]
+            }
+          ],
+          "query": "dragon",
+          "processingTimeMs": "[duration]",
+          "limit": 20,
+          "offset": 0,
+          "estimatedTotalHits": 1,
+          "requestUid": "[uuid]",
+          "metadata": {
+            "queryUid": "[uuid]",
+            "indexUid": "[uuid]",
+            "primaryKey": "id"
+          }
+        }
+      ]
+    }
+    "###);
+}
+
+#[actix_rt::test]
+async fn search_metadata_header_false_value() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    let documents = DOCUMENTS.clone();
+    let (task, _code) = index.add_documents(documents, None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Test with header set to false
+    let (response, code) = index
+        .search_with_headers(json!({"q": "glass"}), vec![("Meili-Include-Metadata", "false")])
+        .await;
+
+    snapshot!(code, @"200 OK");
+    snapshot!(json_string!(response, { ".processingTimeMs" => "[duration]", ".requestUid" => "[uuid]" }), @r###"
+    {
+      "hits": [
+        {
+          "title": "Gläss",
+          "id": "450465",
+          "color": [
+            "blue",
+            "red"
+          ]
+        }
+      ],
+      "query": "glass",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 1,
+      "requestUid": "[uuid]"
+    }
+    "###);
+}
+
+#[actix_rt::test]
+async fn search_metadata_uuid_format() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    let documents = DOCUMENTS.clone();
+    let (task, _code) = index.add_documents(documents, None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    let (response, code) = index
+        .search_with_headers(json!({"q": "glass"}), vec![("Meili-Include-Metadata", "true")])
+        .await;
+
+    snapshot!(code, @"200 OK");
+    snapshot!(json_string!(response, { ".processingTimeMs" => "[duration]", ".requestUid" => "[uuid]", ".metadata.queryUid" => "[uuid]" }), @r###"
+    {
+      "hits": [
+        {
+          "title": "Gläss",
+          "id": "450465",
+          "color": [
+            "blue",
+            "red"
+          ]
+        }
+      ],
+      "query": "glass",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 1,
+      "requestUid": "[uuid]",
+      "metadata": {
+        "queryUid": "[uuid]",
+        "indexUid": "[uuid]",
+        "primaryKey": "id"
+      }
+    }
+    "###);
+}
+
+#[actix_rt::test]
+async fn search_metadata_consistency_across_requests() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    let documents = DOCUMENTS.clone();
+    let (task, _code) = index.add_documents(documents, Some("id")).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Make multiple requests and check that metadata is consistent
+    for _i in 0..3 {
+        let (response, code) = index
+            .search_with_headers(json!({"q": "glass"}), vec![("Meili-Include-Metadata", "true")])
+            .await;
+
+        snapshot!(code, @"200 OK");
+        snapshot!(json_string!(response, { ".processingTimeMs" => "[duration]", ".requestUid" => "[uuid]", ".metadata.queryUid" => "[uuid]" }), @r###"
+        {
+          "hits": [
+            {
+              "id": "450465",
+              "title": "Gläss",
+              "color": [
+                "blue",
+                "red"
+              ]
+            }
+          ],
+          "query": "glass",
+          "processingTimeMs": "[duration]",
+          "limit": 20,
+          "offset": 0,
+          "estimatedTotalHits": 1,
+          "requestUid": "[uuid]",
+          "metadata": {
+            "queryUid": "[uuid]",
+            "indexUid": "[uuid]",
+            "primaryKey": "id"
+          }
+        }
+        "###);
+    }
+}

crates/meilisearch/tests/search/mod.rs

@@ -11,6 +11,7 @@ mod hybrid;
 #[cfg(not(feature = "chinese-pinyin"))]
 mod locales;
 mod matching_strategy;
+mod metadata;
 mod multi;
 mod pagination;
 mod restrict_searchable;

crates/meilisearch/tests/tasks/webhook.rs

@@ -82,7 +82,7 @@ async fn cli_only() {
 
     let (webhooks, code) = server.get_webhooks().await;
     snapshot!(code, @"200 OK");
-    snapshot!(webhooks, @r#"
+    snapshot!(webhooks, @r###"
     {
       "results": [
         {
@@ -90,12 +90,12 @@ async fn cli_only() {
           "isEditable": false,
           "url": "https://example-cli.com/",
           "headers": {
-            "Authorization": "Bearer a-secret-token"
+            "Authorization": "Bearer a-XXXX..."
           }
         }
       ]
     }
-    "#);
+    "###);
 }
 
 #[actix_web::test]
@@ -233,7 +233,7 @@ async fn cli_with_dumps() {
 
     let (webhooks, code) = server.get_webhooks().await;
     snapshot!(code, @"200 OK");
-    snapshot!(webhooks, @r#"
+    snapshot!(webhooks, @r###"
     {
       "results": [
         {
@@ -241,7 +241,7 @@ async fn cli_with_dumps() {
           "isEditable": false,
           "url": "http://defined-in-test-cli.com/",
           "headers": {
-            "Authorization": "Bearer a-secret-token-defined-in-test-cli"
+            "Authorization": "Bearer a-secXXXXXX..."
           }
         },
         {
@@ -255,7 +255,7 @@ async fn cli_with_dumps() {
           "isEditable": true,
           "url": "https://example.com/hook",
           "headers": {
-            "authorization": "TOKEN"
+            "authorization": "XXX..."
           }
         },
         {
@@ -266,7 +266,7 @@ async fn cli_with_dumps() {
         }
       ]
     }
-    "#);
+    "###);
 }
 
 #[actix_web::test]
@@ -367,30 +367,30 @@ async fn post_get_delete() {
         }))
         .await;
     snapshot!(code, @"201 Created");
-    snapshot!(json_string!(value, { ".uuid" => "[uuid]" }), @r#"
+    snapshot!(json_string!(value, { ".uuid" => "[uuid]" }), @r###"
     {
       "uuid": "[uuid]",
       "isEditable": true,
       "url": "https://example.com/hook",
       "headers": {
-        "authorization": "TOKEN"
+        "authorization": "XXX..."
       }
     }
-    "#);
+    "###);
 
     let uuid = value.get("uuid").unwrap().as_str().unwrap();
     let (value, code) = server.get_webhook(uuid).await;
     snapshot!(code, @"200 OK");
-    snapshot!(json_string!(value, { ".uuid" => "[uuid]" }), @r#"
+    snapshot!(json_string!(value, { ".uuid" => "[uuid]" }), @r###"
     {
       "uuid": "[uuid]",
       "isEditable": true,
       "url": "https://example.com/hook",
       "headers": {
-        "authorization": "TOKEN"
+        "authorization": "XXX..."
       }
     }
-    "#);
+    "###);
 
     let (_value, code) = server.delete_webhook(uuid).await;
     snapshot!(code, @"204 No Content");
@@ -430,31 +430,31 @@ async fn create_and_patch() {
     let (value, code) =
         server.patch_webhook(&uuid, json!({ "headers": { "authorization": "TOKEN" } })).await;
     snapshot!(code, @"200 OK");
-    snapshot!(json_string!(value, { ".uuid" => "[uuid]" }), @r#"
+    snapshot!(json_string!(value, { ".uuid" => "[uuid]" }), @r###"
     {
       "uuid": "[uuid]",
       "isEditable": true,
       "url": "https://example.com/hook",
       "headers": {
-        "authorization": "TOKEN"
+        "authorization": "XXX..."
       }
     }
-    "#);
+    "###);
 
     let (value, code) =
         server.patch_webhook(&uuid, json!({ "headers": { "authorization2": "TOKEN" } })).await;
     snapshot!(code, @"200 OK");
-    snapshot!(json_string!(value, { ".uuid" => "[uuid]" }), @r#"
+    snapshot!(json_string!(value, { ".uuid" => "[uuid]" }), @r###"
     {
       "uuid": "[uuid]",
       "isEditable": true,
       "url": "https://example.com/hook",
       "headers": {
-        "authorization": "TOKEN",
+        "authorization": "XXX...",
         "authorization2": "TOKEN"
       }
     }
-    "#);
+    "###);
 
     let (value, code) =
         server.patch_webhook(&uuid, json!({ "headers": { "authorization": null } })).await;

crates/meilisearch/tests/upgrade/mod.rs

@@ -43,7 +43,7 @@ async fn version_too_old() {
     std::fs::write(db_path.join("VERSION"), "1.11.9999").unwrap();
     let options = Opt { experimental_dumpless_upgrade: true, ..default_settings };
     let err = Server::new_with_options(options).await.map(|_| ()).unwrap_err();
-    snapshot!(err, @"Database version 1.11.9999 is too old for the experimental dumpless upgrade feature. Please generate a dump using the v1.11.9999 and import it in the v1.23.0");
+    snapshot!(err, @"Database version 1.11.9999 is too old for the experimental dumpless upgrade feature. Please generate a dump using the v1.11.9999 and import it in the v1.24.0");
 }
 
 #[actix_rt::test]
@@ -58,7 +58,7 @@ async fn version_requires_downgrade() {
     std::fs::write(db_path.join("VERSION"), format!("{major}.{minor}.{patch}")).unwrap();
     let options = Opt { experimental_dumpless_upgrade: true, ..default_settings };
     let err = Server::new_with_options(options).await.map(|_| ()).unwrap_err();
-    snapshot!(err, @"Database version 1.23.1 is higher than the Meilisearch version 1.23.0. Downgrade is not supported");
+    snapshot!(err, @"Database version 1.24.1 is higher than the Meilisearch version 1.24.0. Downgrade is not supported");
 }
 
 #[actix_rt::test]

crates/meilisearch/tests/upgrade/v1_12/snapshots/v1_12_0.rs/check_the_index_scheduler/batches_filter_afterEnqueuedAt_equal_2025-01-16T16_47_41.snap

@@ -8,7 +8,7 @@ source: crates/meilisearch/tests/upgrade/v1_12/v1_12_0.rs
       "progress": null,
       "details": {
         "upgradeFrom": "v1.12.0",
-        "upgradeTo": "v1.23.0"
+        "upgradeTo": "v1.24.0"
       },
       "stats": {
         "totalNbTasks": 1,

crates/meilisearch/tests/upgrade/v1_12/snapshots/v1_12_0.rs/check_the_index_scheduler/batches_filter_afterFinishedAt_equal_2025-01-16T16_47_41.snap

@@ -8,7 +8,7 @@ source: crates/meilisearch/tests/upgrade/v1_12/v1_12_0.rs
       "progress": null,
       "details": {
         "upgradeFrom": "v1.12.0",
-        "upgradeTo": "v1.23.0"
+        "upgradeTo": "v1.24.0"
       },
       "stats": {
         "totalNbTasks": 1,

crates/meilisearch/tests/upgrade/v1_12/snapshots/v1_12_0.rs/check_the_index_scheduler/batches_filter_afterStartedAt_equal_2025-01-16T16_47_41.snap

@@ -8,7 +8,7 @@ source: crates/meilisearch/tests/upgrade/v1_12/v1_12_0.rs
       "progress": null,
       "details": {
         "upgradeFrom": "v1.12.0",
-        "upgradeTo": "v1.23.0"
+        "upgradeTo": "v1.24.0"
       },
       "stats": {
         "totalNbTasks": 1,

crates/meilisearch/tests/upgrade/v1_12/snapshots/v1_12_0.rs/check_the_index_scheduler/tasks_filter_afterEnqueuedAt_equal_2025-01-16T16_47_41.snap

@@ -12,7 +12,7 @@ source: crates/meilisearch/tests/upgrade/v1_12/v1_12_0.rs
       "canceledBy": null,
       "details": {
         "upgradeFrom": "v1.12.0",
-        "upgradeTo": "v1.23.0"
+        "upgradeTo": "v1.24.0"
       },
       "error": null,
       "duration": "[duration]",

crates/meilisearch/tests/upgrade/v1_12/snapshots/v1_12_0.rs/check_the_index_scheduler/tasks_filter_afterFinishedAt_equal_2025-01-16T16_47_41.snap

@@ -12,7 +12,7 @@ source: crates/meilisearch/tests/upgrade/v1_12/v1_12_0.rs
       "canceledBy": null,
       "details": {
         "upgradeFrom": "v1.12.0",
-        "upgradeTo": "v1.23.0"
+        "upgradeTo": "v1.24.0"
       },
       "error": null,
       "duration": "[duration]",

crates/meilisearch/tests/upgrade/v1_12/snapshots/v1_12_0.rs/check_the_index_scheduler/tasks_filter_afterStartedAt_equal_2025-01-16T16_47_41.snap

@@ -12,7 +12,7 @@ source: crates/meilisearch/tests/upgrade/v1_12/v1_12_0.rs
       "canceledBy": null,
       "details": {
         "upgradeFrom": "v1.12.0",
-        "upgradeTo": "v1.23.0"
+        "upgradeTo": "v1.24.0"
       },
       "error": null,
       "duration": "[duration]",

crates/meilisearch/tests/upgrade/v1_12/snapshots/v1_12_0.rs/check_the_index_scheduler/the_whole_batch_queue_once_everything_has_been_processed.snap

@@ -8,7 +8,7 @@ source: crates/meilisearch/tests/upgrade/v1_12/v1_12_0.rs
       "progress": null,
       "details": {
         "upgradeFrom": "v1.12.0",
-        "upgradeTo": "v1.23.0"
+        "upgradeTo": "v1.24.0"
       },
       "stats": {
         "totalNbTasks": 1,

crates/meilisearch/tests/upgrade/v1_12/snapshots/v1_12_0.rs/check_the_index_scheduler/the_whole_task_queue_once_everything_has_been_processed.snap

@@ -12,7 +12,7 @@ source: crates/meilisearch/tests/upgrade/v1_12/v1_12_0.rs
       "canceledBy": null,
       "details": {
         "upgradeFrom": "v1.12.0",
-        "upgradeTo": "v1.23.0"
+        "upgradeTo": "v1.24.0"
       },
       "error": null,
       "duration": "[duration]",

crates/meilitool/src/upgrade/v1_12.rs

@@ -68,7 +68,7 @@ fn convert_update_files(db_path: &Path) -> anyhow::Result<()> {
 
     for uuid in file_store.all_uuids().context("while retrieving uuids from file store")? {
         let uuid = uuid.context("while retrieving uuid from file store")?;
-        let update_file_path = file_store.get_update_path(uuid);
+        let update_file_path = file_store.update_path(uuid);
         let update_file = file_store
             .get_update(uuid)
             .with_context(|| format!("while getting update file for uuid {uuid:?}"))?;

crates/milli/Cargo.toml

@@ -18,7 +18,7 @@ bincode = "1.3.3"
 bstr = "1.12.0"
 bytemuck = { version = "1.23.1", features = ["extern_crate_alloc"] }
 byteorder = "1.5.0"
-charabia = { version = "0.9.7", default-features = false }
+charabia = { version = "0.9.8", default-features = false }
 cellulite = "0.3.1-nested-rtxns-2"
 concat-arrays = "0.1.2"
 convert_case = "0.8.0"
@@ -34,7 +34,7 @@ grenad = { version = "0.5.0", default-features = false, features = [
     "rayon",
     "tempfile",
 ] }
-heed = { version = "0.22.1-nested-rtxns", default-features = false, features = [
+heed = { version = "0.22.1-nested-rtxns-6", default-features = false, features = [
     "serde-json",
     "serde-bincode",
 ] }
@@ -90,7 +90,7 @@ rhai = { version = "1.22.2", features = [
     "sync",
 ] }
 arroy = "0.6.4-nested-rtxns"
-hannoy = { version = "0.0.9-nested-rtxns", features = ["arroy"] }
+hannoy = { version = "0.0.9-nested-rtxns-2", features = ["arroy"] }
 rand = "0.8.5"
 tracing = "0.1.41"
 ureq = { version = "2.12.1", features = ["json"] }
@@ -120,14 +120,16 @@ twox-hash = { version = "2.1.1", default-features = false, features = [
 ] }
 geo-types = "0.7.16"
 zerometry = "0.3.0"
+bitpacking = "0.9.2"
 
 [dev-dependencies]
-mimalloc = { version = "0.1.47", default-features = false }
 # fixed version due to format breakages in v1.40
 insta = "=1.39.0"
+mimalloc = { version = "0.1.47", default-features = false }
 maplit = "1.0.2"
 md5 = "0.7.0"
 meili-snap = { path = "../meili-snap" }
+quickcheck = "1.0.3"
 rand = { version = "0.8.5", features = ["small_rng"] }
 
 [features]

crates/milli/src/heed_codec/de_bitmap.rs

@@ -0,0 +1,299 @@
+use std::io::{self, ErrorKind};
+use std::mem::{size_of, size_of_val};
+
+use bitpacking::{BitPacker, BitPacker1x, BitPacker4x, BitPacker8x};
+use roaring::RoaringBitmap;
+
+/// The magic header for our custom encoding format
+const MAGIC_HEADER: u16 = 36869;
+
+pub struct DeBitmapCodec;
+
+// TODO reintroduce:
+//  - serialized_size?
+//  - serialize_into_vec
+//  - intersection_with_serialized
+//  - merge_into
+//  - merge_deladd_into
+impl DeBitmapCodec {
+    /// Returns the delta-encoded compressed version of the given roaring bitmap.
+    pub fn serialize_into<W: io::Write>(bitmap: &RoaringBitmap, writer: W) -> io::Result<()> {
+        let mut tmp_buffer = Vec::new();
+        Self::serialize_into_with_tmp_buffer(bitmap, writer, &mut tmp_buffer)
+    }
+
+    /// Same as [Self::serialize_into] but accepts a buffer to avoid allocating one.
+    pub fn serialize_into_with_tmp_buffer<W: io::Write>(
+        bitmap: &RoaringBitmap,
+        mut writer: W,
+        tmp_buffer: &mut Vec<u32>,
+    ) -> io::Result<()> {
+        // Insert the magic header
+        writer.write_all(&MAGIC_HEADER.to_ne_bytes())?;
+
+        let bitpacker8x = BitPacker8x::new();
+        let bitpacker4x = BitPacker4x::new();
+        let bitpacker1x = BitPacker1x::new();
+
+        // This temporary buffer is used to store each chunk of decompressed and
+        // compressed and delta-encoded u32s. We need room for the decompressed
+        // u32s coming from the roaring bitmap, the compressed output that can
+        // be as large as the decompressed u32s, and the chunk header.
+        tmp_buffer.resize((BitPacker8x::BLOCK_LEN * 2) + 1, 0u32);
+        let (decompressed, compressed) = tmp_buffer.split_at_mut(BitPacker8x::BLOCK_LEN);
+        let compressed = bytemuck::cast_slice_mut(compressed);
+
+        let mut buffer_index = 0;
+        let mut initial = None;
+        // We initially collect all the integers into a flat buffer of the size
+        // of the largest bitpacker. We encode them with it until we don't have
+        // enough of them...
+        for n in bitmap {
+            decompressed[buffer_index] = n;
+            buffer_index += 1;
+            if buffer_index == BitPacker8x::BLOCK_LEN {
+                let output = encode_with_packer(&bitpacker8x, decompressed, initial, compressed);
+                writer.write_all(output)?;
+                initial = Some(n);
+                buffer_index = 0;
+            }
+        }
+
+        // ...We then switch to a smaller bitpacker to encode the remaining chunks...
+        let decompressed = &decompressed[..buffer_index];
+        let mut chunks = decompressed.chunks_exact(BitPacker4x::BLOCK_LEN);
+        for decompressed in chunks.by_ref() {
+            let output = encode_with_packer(&bitpacker4x, decompressed, initial, compressed);
+            writer.write_all(output)?;
+            initial = decompressed.iter().last().copied();
+        }
+
+        // ...And so on...
+        let decompressed = chunks.remainder();
+        let mut chunks = decompressed.chunks_exact(BitPacker1x::BLOCK_LEN);
+        for decompressed in chunks.by_ref() {
+            let output = encode_with_packer(&bitpacker1x, decompressed, initial, compressed);
+            writer.write_all(output)?;
+            initial = decompressed.iter().last().copied();
+        }
+
+        // ...Until we don't have any small enough bitpacker. We put them raw
+        // at the end of out buffer with a header indicating the matter.
+        let decompressed = chunks.remainder();
+        if !decompressed.is_empty() {
+            let header = encode_chunk_header(BitPackerLevel::None, u32::BITS as u8);
+            // Note: Not convinced about the performance of writing a single
+            //       byte followed by a larger write. However, we will use this
+            //       codec with a BufWriter or directly with a Vec of bytes.
+            writer.write_all(&[header])?;
+            writer.write_all(bytemuck::cast_slice(decompressed))?;
+        }
+
+        Ok(())
+    }
+
+    /// Returns the delta-decoded roaring bitmap from the compressed bytes.
+    pub fn deserialize_from(compressed: &[u8]) -> io::Result<RoaringBitmap> {
+        let mut tmp_buffer = Vec::new();
+        Self::deserialize_from_with_tmp_buffer(compressed, &mut tmp_buffer)
+    }
+
+    /// Same as [Self::deserialize_from] but accepts a buffer to avoid allocating one.
+    pub fn deserialize_from_with_tmp_buffer(
+        input: &[u8],
+        tmp_buffer: &mut Vec<u32>,
+    ) -> io::Result<RoaringBitmap> {
+        let Some((header, mut compressed)) = input.split_at_checked(size_of_val(&MAGIC_HEADER))
+        else {
+            return Err(io::Error::new(ErrorKind::UnexpectedEof, "expecting a two-bytes header"));
+        };
+
+        // Safety: This unwrap cannot happen as the header buffer is the right size
+        let header = u16::from_ne_bytes(header.try_into().unwrap());
+
+        if header != MAGIC_HEADER {
+            return Err(io::Error::other("invalid header value"));
+        }
+
+        let bitpacker8x = BitPacker8x::new();
+        let bitpacker4x = BitPacker4x::new();
+        let bitpacker1x = BitPacker1x::new();
+
+        let mut bitmap = RoaringBitmap::new();
+        tmp_buffer.resize(BitPacker8x::BLOCK_LEN, 0u32);
+        let decompressed = &mut tmp_buffer[..];
+        let mut initial = None;
+
+        while let Some((&chunk_header, encoded)) = compressed.split_first() {
+            let (level, num_bits) = decode_chunk_header(chunk_header);
+            let (bytes_read, decompressed) = match level {
+                BitPackerLevel::None => {
+                    if num_bits != u32::BITS as u8 {
+                        return Err(io::Error::new(
+                            ErrorKind::InvalidData,
+                            "invalid number of bits to encode non-compressed u32s",
+                        ));
+                    }
+
+                    let chunks = encoded.chunks_exact(size_of::<u32>());
+                    if !chunks.remainder().is_empty() {
+                        return Err(io::Error::new(
+                            io::ErrorKind::InvalidData,
+                            "expecting last chunk to be a multiple of the size of an u32",
+                        ));
+                    }
+
+                    let integers = chunks
+                        // safety: This unwrap cannot happen as
+                        //         the size of u32 is set correctly.
+                        .map(|b| b.try_into().unwrap())
+                        .map(u32::from_ne_bytes);
+
+                    bitmap
+                        .append(integers)
+                        .map_err(|e| io::Error::new(ErrorKind::InvalidData, e))?;
+
+                    // This is basically always the last chunk that exists in
+                    // this delta-encoded format as the raw u32s are appended
+                    // when there is not enough of them to fit in a bitpacker.
+                    break;
+                }
+                BitPackerLevel::BitPacker1x => {
+                    decode_with_packer(&bitpacker1x, decompressed, initial, encoded, num_bits)
+                }
+                BitPackerLevel::BitPacker4x => {
+                    decode_with_packer(&bitpacker4x, decompressed, initial, encoded, num_bits)
+                }
+                BitPackerLevel::BitPacker8x => {
+                    decode_with_packer(&bitpacker8x, decompressed, initial, encoded, num_bits)
+                }
+            };
+
+            initial = decompressed.iter().last().copied();
+            // TODO investigate perf
+            // Safety: Bitpackers cannot output unsorter integers when
+            //         used with the compress_strictly_sorted function.
+            bitmap.append(decompressed.iter().copied()).unwrap();
+            // What the delta-decoding read plus the chunk header size
+            compressed = &compressed[bytes_read + 1..];
+        }
+
+        Ok(bitmap)
+    }
+}
+
+/// Takes a strickly sorted list of u32s and outputs delta-encoded
+/// bytes with a chunk header. We expect the output buffer to be
+/// at least BLOCK_LEN + 1.
+fn encode_with_packer<'c, B: BitPackerExt>(
+    bitpacker: &B,
+    decompressed: &[u32],
+    initial: Option<u32>,
+    output: &'c mut [u8],
+) -> &'c [u8] {
+    let num_bits = bitpacker.num_bits_strictly_sorted(initial, decompressed);
+    let compressed_len = B::compressed_block_size(num_bits);
+    let chunk_header = encode_chunk_header(B::level(), num_bits);
+    let buffer = &mut output[..compressed_len + 1];
+    // Safety: The buffer is at least one byte
+    let (header_in_buffer, encoded) = buffer.split_first_mut().unwrap();
+    *header_in_buffer = chunk_header;
+    bitpacker.compress_strictly_sorted(initial, decompressed, encoded, num_bits);
+    buffer
+}
+
+/// Returns the number of bytes read and the decoded unsigned integers.
+fn decode_with_packer<'d, B: BitPacker>(
+    bitpacker: &B,
+    decompressed: &'d mut [u32],
+    initial: Option<u32>,
+    compressed: &[u8],
+    num_bits: u8,
+) -> (usize, &'d [u32]) {
+    let decompressed = &mut decompressed[..B::BLOCK_LEN];
+    let read = bitpacker.decompress_strictly_sorted(initial, compressed, decompressed, num_bits);
+    (read, decompressed)
+}
+
+/// An identifier for the bitpacker to be able
+/// to correctly decode the compressed integers.
+#[derive(Debug, PartialEq, Eq)]
+#[repr(u8)]
+enum BitPackerLevel {
+    /// The remaining bytes are raw little endian encoded u32s.
+    None,
+    /// The remaining bits are encoded using a `BitPacker1x`.
+    BitPacker1x,
+    /// The remaining bits are encoded using a `BitPacker4x`.
+    BitPacker4x,
+    /// The remaining bits are encoded using a `BitPacker8x`.
+    BitPacker8x,
+}
+
+/// Returns the chunk header based on the bitpacker level
+/// and the number of bits to encode the list of integers.
+fn encode_chunk_header(level: BitPackerLevel, num_bits: u8) -> u8 {
+    debug_assert!(num_bits as u32 <= 2_u32.pow(6));
+    let level = level as u8;
+    debug_assert!(level <= 3);
+    num_bits | (level << 6)
+}
+
+/// Decodes the chunk header and output the bitpacker level
+/// and the number of bits to decode the following bytes.
+fn decode_chunk_header(data: u8) -> (BitPackerLevel, u8) {
+    let num_bits = data & 0b00111111;
+    let level = match data >> 6 {
+        0 => BitPackerLevel::None,
+        1 => BitPackerLevel::BitPacker1x,
+        2 => BitPackerLevel::BitPacker4x,
+        3 => BitPackerLevel::BitPacker8x,
+        invalid => panic!("Invalid bitpacker level: {invalid}"),
+    };
+    debug_assert!(num_bits as u32 <= 2_u32.pow(6));
+    (level, num_bits)
+}
+
+/// A simple helper trait to get the BitPackerLevel
+/// and correctly generate the chunk header.
+trait BitPackerExt: BitPacker {
+    /// Returns the level of the bitpacker: an identifier to be
+    /// able to decode the numbers with the right bitpacker.
+    fn level() -> BitPackerLevel;
+}
+
+impl BitPackerExt for BitPacker8x {
+    fn level() -> BitPackerLevel {
+        BitPackerLevel::BitPacker8x
+    }
+}
+
+impl BitPackerExt for BitPacker4x {
+    fn level() -> BitPackerLevel {
+        BitPackerLevel::BitPacker4x
+    }
+}
+
+impl BitPackerExt for BitPacker1x {
+    fn level() -> BitPackerLevel {
+        BitPackerLevel::BitPacker1x
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use quickcheck::quickcheck;
+    use roaring::RoaringBitmap;
+
+    use crate::heed_codec::de_bitmap::DeBitmapCodec;
+
+    quickcheck! {
+        fn qc_random(xs: Vec<u32>) -> bool {
+            let bitmap = RoaringBitmap::from_iter(xs);
+            let mut compressed = Vec::new();
+            DeBitmapCodec::serialize_into(&bitmap, &mut compressed).unwrap();
+            let decompressed = DeBitmapCodec::deserialize_from(&compressed[..]).unwrap();
+            decompressed == bitmap
+        }
+    }
+}

crates/milli/src/heed_codec/mod.rs

@@ -1,6 +1,7 @@
 mod beu16_str_codec;
 mod beu32_str_codec;
 mod byte_slice_ref;
+mod de_bitmap;
 pub mod facet;
 mod field_id_word_count_codec;
 mod fst_set_codec;
@@ -19,6 +20,7 @@ use thiserror::Error;
 
 pub use self::beu16_str_codec::BEU16StrCodec;
 pub use self::beu32_str_codec::BEU32StrCodec;
+pub use self::de_bitmap::DeBitmapCodec;
 pub use self::field_id_word_count_codec::FieldIdWordCountCodec;
 pub use self::fst_set_codec::FstSetCodec;
 pub use self::obkv_codec::ObkvCodec;

crates/milli/src/index.rs

@@ -425,6 +425,10 @@ impl Index {
         self.env.info().map_size
     }
 
+    pub fn try_clone_inner_file(&self) -> heed::Result<File> {
+        self.env.try_clone_inner_file()
+    }
+
     pub fn copy_to_file(&self, file: &mut File, option: CompactionOption) -> Result<()> {
         self.env.copy_to_file(file, option).map_err(Into::into)
     }

crates/milli/src/search/new/bucket_sort.rs

@@ -97,7 +97,7 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
 
     logger.start_iteration_ranking_rule(0, ranking_rules[0].as_ref(), query, universe);
 
-    ranking_rules[0].start_iteration(ctx, logger, universe, query)?;
+    ranking_rules[0].start_iteration(ctx, logger, universe, query, &time_budget)?;
 
     let mut ranking_rule_scores: Vec<ScoreDetails> = vec![];
 
@@ -168,42 +168,6 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
         };
 
     while valid_docids.len() < max_len_to_evaluate {
-        if time_budget.exceeded() {
-            loop {
-                let bucket = std::mem::take(&mut ranking_rule_universes[cur_ranking_rule_index]);
-                ranking_rule_scores.push(ScoreDetails::Skipped);
-
-                // remove candidates from the universe without adding them to result if their score is below the threshold
-                let is_below_threshold =
-                    ranking_score_threshold.is_some_and(|ranking_score_threshold| {
-                        let current_score = ScoreDetails::global_score(ranking_rule_scores.iter());
-                        current_score < ranking_score_threshold
-                    });
-
-                if is_below_threshold {
-                    all_candidates -= &bucket;
-                    all_candidates -= &ranking_rule_universes[cur_ranking_rule_index];
-                } else {
-                    maybe_add_to_results!(bucket);
-                }
-
-                ranking_rule_scores.pop();
-
-                if cur_ranking_rule_index == 0 {
-                    break;
-                }
-
-                back!();
-            }
-
-            return Ok(BucketSortOutput {
-                scores: valid_scores,
-                docids: valid_docids,
-                all_candidates,
-                degraded: true,
-            });
-        }
-
         // The universe for this bucket is zero, so we don't need to sort
         // anything, just go back to the parent ranking rule.
         if ranking_rule_universes[cur_ranking_rule_index].is_empty()
@@ -216,14 +180,63 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
             continue;
         }
 
-        let Some(next_bucket) = ranking_rules[cur_ranking_rule_index].next_bucket(
-            ctx,
-            logger,
-            &ranking_rule_universes[cur_ranking_rule_index],
-        )?
-        else {
-            back!();
-            continue;
+        let next_bucket = if time_budget.exceeded() {
+            match ranking_rules[cur_ranking_rule_index].non_blocking_next_bucket(
+                ctx,
+                logger,
+                &ranking_rule_universes[cur_ranking_rule_index],
+            )? {
+                std::task::Poll::Ready(bucket) => bucket,
+                std::task::Poll::Pending => {
+                    loop {
+                        let bucket =
+                            std::mem::take(&mut ranking_rule_universes[cur_ranking_rule_index]);
+                        ranking_rule_scores.push(ScoreDetails::Skipped);
+
+                        // remove candidates from the universe without adding them to result if their score is below the threshold
+                        let is_below_threshold =
+                            ranking_score_threshold.is_some_and(|ranking_score_threshold| {
+                                let current_score =
+                                    ScoreDetails::global_score(ranking_rule_scores.iter());
+                                current_score < ranking_score_threshold
+                            });
+
+                        if is_below_threshold {
+                            all_candidates -= &bucket;
+                            all_candidates -= &ranking_rule_universes[cur_ranking_rule_index];
+                        } else {
+                            maybe_add_to_results!(bucket);
+                        }
+
+                        ranking_rule_scores.pop();
+
+                        if cur_ranking_rule_index == 0 {
+                            break;
+                        }
+
+                        back!();
+                    }
+
+                    return Ok(BucketSortOutput {
+                        scores: valid_scores,
+                        docids: valid_docids,
+                        all_candidates,
+                        degraded: true,
+                    });
+                }
+            }
+        } else {
+            let Some(next_bucket) = ranking_rules[cur_ranking_rule_index].next_bucket(
+                ctx,
+                logger,
+                &ranking_rule_universes[cur_ranking_rule_index],
+                &time_budget,
+            )?
+            else {
+                back!();
+                continue;
+            };
+            next_bucket
         };
 
         ranking_rule_scores.push(next_bucket.score);
@@ -275,6 +288,7 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
             logger,
             &next_bucket.candidates,
             &next_bucket.query,
+            &time_budget,
         )?;
     }
 

crates/milli/src/search/new/exact_attribute.rs

@@ -6,7 +6,7 @@ use super::ranking_rules::{RankingRule, RankingRuleOutput};
 use crate::score_details::{self, ScoreDetails};
 use crate::search::new::query_graph::QueryNodeData;
 use crate::search::new::query_term::ExactTerm;
-use crate::{CboRoaringBitmapCodec, Result, SearchContext, SearchLogger};
+use crate::{CboRoaringBitmapCodec, Result, SearchContext, SearchLogger, TimeBudget};
 
 /// A ranking rule that produces 3 disjoint buckets:
 ///
@@ -35,6 +35,7 @@ impl<'ctx> RankingRule<'ctx, QueryGraph> for ExactAttribute {
         _logger: &mut dyn SearchLogger<QueryGraph>,
         universe: &roaring::RoaringBitmap,
         query: &QueryGraph,
+        _time_budget: &TimeBudget,
     ) -> Result<()> {
         self.state = State::start_iteration(ctx, universe, query)?;
         Ok(())
@@ -46,6 +47,7 @@ impl<'ctx> RankingRule<'ctx, QueryGraph> for ExactAttribute {
         _ctx: &mut SearchContext<'ctx>,
         _logger: &mut dyn SearchLogger<QueryGraph>,
         universe: &roaring::RoaringBitmap,
+        _time_budget: &TimeBudget,
     ) -> Result<Option<RankingRuleOutput<QueryGraph>>> {
         let state = std::mem::take(&mut self.state);
         let (state, output) = State::next(state, universe);

crates/milli/src/search/new/geo_sort.rs

@@ -7,7 +7,7 @@ use super::ranking_rules::{RankingRule, RankingRuleOutput, RankingRuleQueryTrait
 use crate::documents::geo_sort::{fill_cache, next_bucket};
 use crate::documents::{GeoSortParameter, GeoSortStrategy};
 use crate::score_details::{self, ScoreDetails};
-use crate::{GeoPoint, Result, SearchContext, SearchLogger};
+use crate::{GeoPoint, Result, SearchContext, SearchLogger, TimeBudget};
 
 pub struct GeoSort<Q: RankingRuleQueryTrait> {
     query: Option<Q>,
@@ -84,6 +84,7 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for GeoSort<Q> {
         _logger: &mut dyn SearchLogger<Q>,
         universe: &RoaringBitmap,
         query: &Q,
+        _time_budget: &TimeBudget,
     ) -> Result<()> {
         assert!(self.query.is_none());
 
@@ -110,6 +111,7 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for GeoSort<Q> {
         ctx: &mut SearchContext<'ctx>,
         _logger: &mut dyn SearchLogger<Q>,
         universe: &RoaringBitmap,
+        _time_budget: &TimeBudget,
     ) -> Result<Option<RankingRuleOutput<Q>>> {
         let query = self.query.as_ref().unwrap().clone();
 

crates/milli/src/search/new/graph_based_ranking_rule.rs

@@ -53,7 +53,7 @@ use super::{QueryGraph, RankingRule, RankingRuleOutput, SearchContext};
 use crate::score_details::Rank;
 use crate::search::new::query_term::LocatedQueryTermSubset;
 use crate::search::new::ranking_rule_graph::PathVisitor;
-use crate::{Result, TermsMatchingStrategy};
+use crate::{Result, TermsMatchingStrategy, TimeBudget};
 
 pub type Words = GraphBasedRankingRule<WordsGraph>;
 impl GraphBasedRankingRule<WordsGraph> {
@@ -135,6 +135,7 @@ impl<'ctx, G: RankingRuleGraphTrait> RankingRule<'ctx, QueryGraph> for GraphBase
         _logger: &mut dyn SearchLogger<QueryGraph>,
         _universe: &RoaringBitmap,
         query_graph: &QueryGraph,
+        _time_budget: &TimeBudget,
     ) -> Result<()> {
         // the `next_max_cost` is the successor integer to the maximum cost of the paths in the graph.
         //
@@ -217,6 +218,7 @@ impl<'ctx, G: RankingRuleGraphTrait> RankingRule<'ctx, QueryGraph> for GraphBase
         ctx: &mut SearchContext<'ctx>,
         logger: &mut dyn SearchLogger<QueryGraph>,
         universe: &RoaringBitmap,
+        _time_budget: &TimeBudget,
     ) -> Result<Option<RankingRuleOutput<QueryGraph>>> {
         // Will crash if `next_bucket` is called before `start_iteration` or after `end_iteration`,
         // should never happen

crates/milli/src/search/new/ranking_rules.rs

@@ -1,9 +1,11 @@
+use std::task::Poll;
+
 use roaring::RoaringBitmap;
 
 use super::logger::SearchLogger;
 use super::{QueryGraph, SearchContext};
 use crate::score_details::ScoreDetails;
-use crate::Result;
+use crate::{Result, TimeBudget};
 
 /// An internal trait implemented by only [`PlaceholderQuery`] and [`QueryGraph`]
 pub trait RankingRuleQueryTrait: Sized + Clone + 'static {}
@@ -28,12 +30,15 @@ pub trait RankingRule<'ctx, Query: RankingRuleQueryTrait> {
     /// buckets using [`next_bucket`](RankingRule::next_bucket).
     ///
     /// The given universe is the universe that will be given to [`next_bucket`](RankingRule::next_bucket).
+    ///
+    /// If this function may take a long time, it should check the `time_budget` and return early if exceeded.
     fn start_iteration(
         &mut self,
         ctx: &mut SearchContext<'ctx>,
         logger: &mut dyn SearchLogger<Query>,
         universe: &RoaringBitmap,
         query: &Query,
+        time_budget: &TimeBudget,
     ) -> Result<()>;
 
     /// Return the next bucket of this ranking rule.
@@ -43,13 +48,31 @@ pub trait RankingRule<'ctx, Query: RankingRuleQueryTrait> {
     /// The universe given as argument is either:
     /// - a subset of the universe given to the previous call to [`next_bucket`](RankingRule::next_bucket); OR
     /// - the universe given to [`start_iteration`](RankingRule::start_iteration)
+    ///
+    /// If this function may take a long time, it should check the `time_budget` and return early if exceeded.
     fn next_bucket(
         &mut self,
         ctx: &mut SearchContext<'ctx>,
         logger: &mut dyn SearchLogger<Query>,
         universe: &RoaringBitmap,
+        time_budget: &TimeBudget,
     ) -> Result<Option<RankingRuleOutput<Query>>>;
 
+    /// Return the next bucket of this ranking rule, if doing so can be done without blocking
+    ///
+    /// Even if the time budget is exceeded, when getting the next bucket is a fast operation, this should return `true`
+    /// to allow Meilisearch to collect the results.
+    ///
+    /// Default implementation conservatively returns that it would block.
+    fn non_blocking_next_bucket(
+        &mut self,
+        _ctx: &mut SearchContext<'ctx>,
+        _logger: &mut dyn SearchLogger<Query>,
+        _universe: &RoaringBitmap,
+    ) -> Result<Poll<RankingRuleOutput<Query>>> {
+        Ok(Poll::Pending)
+    }
+
     /// Finish iterating over the buckets, which yields control to the parent ranking rule
     /// The next call to this ranking rule, if any, will be [`start_iteration`](RankingRule::start_iteration).
     fn end_iteration(

crates/milli/src/search/new/sort.rs

@@ -7,7 +7,7 @@ use crate::heed_codec::facet::{FacetGroupKeyCodec, OrderedF64Codec};
 use crate::heed_codec::{BytesRefCodec, StrRefCodec};
 use crate::score_details::{self, ScoreDetails};
 use crate::search::facet::{ascending_facet_sort, descending_facet_sort};
-use crate::{FieldId, Index, Result};
+use crate::{FieldId, Index, Result, TimeBudget};
 
 pub trait RankingRuleOutputIter<'ctx, Query> {
     fn next_bucket(&mut self) -> Result<Option<RankingRuleOutput<Query>>>;
@@ -96,6 +96,7 @@ impl<'ctx, Query: RankingRuleQueryTrait> RankingRule<'ctx, Query> for Sort<'ctx,
         _logger: &mut dyn SearchLogger<Query>,
         parent_candidates: &RoaringBitmap,
         parent_query: &Query,
+        _time_budget: &TimeBudget,
     ) -> Result<()> {
         let iter: RankingRuleOutputIterWrapper<'ctx, Query> = match self.field_id {
             Some(field_id) => {
@@ -194,6 +195,7 @@ impl<'ctx, Query: RankingRuleQueryTrait> RankingRule<'ctx, Query> for Sort<'ctx,
         _ctx: &mut SearchContext<'ctx>,
         _logger: &mut dyn SearchLogger<Query>,
         universe: &RoaringBitmap,
+        _time_budget: &TimeBudget,
     ) -> Result<Option<RankingRuleOutput<Query>>> {
         let iter = self.iter.as_mut().unwrap();
         if let Some(mut bucket) = iter.next_bucket()? {

crates/milli/src/search/new/tests/cutoff.rs

@@ -3,12 +3,17 @@
 //! 2. A test that ensure the filters are affectively applied even with a cutoff of 0
 //! 3. A test that ensure the cutoff works well with the ranking scores
 
+use std::collections::BTreeMap;
+use std::sync::Arc;
 use std::time::Duration;
 
 use meili_snap::snapshot;
 
 use crate::index::tests::TempIndex;
 use crate::score_details::{ScoreDetails, ScoringStrategy};
+use crate::update::Setting;
+use crate::vector::settings::EmbeddingSettings;
+use crate::vector::{Embedder, EmbedderOptions};
 use crate::{Criterion, Filter, FilterableAttributesRule, Search, TimeBudget};
 
 fn create_index() -> TempIndex {
@@ -361,9 +366,8 @@ fn degraded_search_and_score_details() {
     ]
     "###);
 
-    // After SIX loop iteration. The words ranking rule gave us a new bucket.
-    // Since we reached the limit we were able to early exit without checking the typo ranking rule.
-    search.time_budget(TimeBudget::max().with_stop_after(6));
+    // After FIVE loop iterations. The words ranking rule gave us a new bucket.
+    search.time_budget(TimeBudget::max().with_stop_after(5));
 
     let result = search.execute().unwrap();
     snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
@@ -424,4 +428,399 @@ fn degraded_search_and_score_details() {
         ],
     ]
     "###);
+
+    // After SIX loop iterations.
+    // we finished
+    search.time_budget(TimeBudget::max().with_stop_after(6));
+
+    let result = search.execute().unwrap();
+    snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
+    IDs: [4, 1, 0, 3]
+    Scores: 1.0000 0.9167 0.8333 0.6667 
+    Score Details:
+    [
+        [
+            Words(
+                Words {
+                    matching_words: 3,
+                    max_matching_words: 3,
+                },
+            ),
+            Typo(
+                Typo {
+                    typo_count: 0,
+                    max_typo_count: 3,
+                },
+            ),
+        ],
+        [
+            Words(
+                Words {
+                    matching_words: 3,
+                    max_matching_words: 3,
+                },
+            ),
+            Typo(
+                Typo {
+                    typo_count: 1,
+                    max_typo_count: 3,
+                },
+            ),
+        ],
+        [
+            Words(
+                Words {
+                    matching_words: 3,
+                    max_matching_words: 3,
+                },
+            ),
+            Typo(
+                Typo {
+                    typo_count: 2,
+                    max_typo_count: 3,
+                },
+            ),
+        ],
+        [
+            Words(
+                Words {
+                    matching_words: 2,
+                    max_matching_words: 3,
+                },
+            ),
+            Typo(
+                Typo {
+                    typo_count: 0,
+                    max_typo_count: 2,
+                },
+            ),
+        ],
+    ]
+    "###);
+}
+
+#[test]
+fn degraded_search_and_score_details_vector() {
+    let index = create_index();
+
+    index
+        .add_documents(documents!([
+            {
+                "id": 4,
+                "text": "hella puppo kefir",
+                "_vectors": {
+                    "default": [0.1, 0.1]
+                }
+            },
+            {
+                "id": 3,
+                "text": "hella puppy kefir",
+                "_vectors": {
+                    "default": [-0.1, 0.1]
+                }
+            },
+            {
+                "id": 2,
+                "text": "hello",
+                "_vectors": {
+                    "default": [0.1, -0.1]
+                }
+            },
+            {
+                "id": 1,
+                "text": "hello puppy",
+                "_vectors": {
+                    "default": [-0.1, -0.1]
+                }
+            },
+            {
+                "id": 0,
+                "text": "hello puppy kefir",
+                "_vectors": {
+                    "default": null
+                }
+            },
+        ]))
+        .unwrap();
+
+    index
+        .update_settings(|settings| {
+            let mut embedders = BTreeMap::new();
+            embedders.insert(
+                "default".into(),
+                Setting::Set(EmbeddingSettings {
+                    source: Setting::Set(crate::vector::settings::EmbedderSource::UserProvided),
+                    dimensions: Setting::Set(2),
+                    ..Default::default()
+                }),
+            );
+            settings.set_embedder_settings(embedders);
+            settings.set_vector_store(crate::vector::VectorStoreBackend::Hannoy);
+        })
+        .unwrap();
+
+    let rtxn = index.read_txn().unwrap();
+    let mut search = Search::new(&rtxn, &index);
+
+    let embedder = Arc::new(
+        Embedder::new(
+            EmbedderOptions::UserProvided(crate::vector::embedder::manual::EmbedderOptions {
+                dimensions: 2,
+                distribution: None,
+            }),
+            0,
+        )
+        .unwrap(),
+    );
+
+    search.semantic("default".into(), embedder, false, Some(vec![1., -1.]), None);
+
+    search.limit(4);
+    search.scoring_strategy(ScoringStrategy::Detailed);
+    search.time_budget(TimeBudget::max());
+
+    let result = search.execute().unwrap();
+    snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
+    IDs: [2, 0, 3, 1]
+    Scores: 1.0000 0.5000 0.5000 0.0000 
+    Score Details:
+    [
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        1.0,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.0,
+                    ),
+                },
+            ),
+        ],
+    ]
+    "###);
+
+    // Do ONE loop iteration. Not much can be deduced, almost everyone matched the words first bucket.
+    search.time_budget(TimeBudget::max().with_stop_after(1));
+
+    let result = search.execute().unwrap();
+    snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
+    IDs: [0, 1, 2, 3]
+    Scores: 0.5000 0.0000 0.0000 0.0000 
+    Score Details:
+    [
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Skipped,
+        ],
+        [
+            Skipped,
+        ],
+        [
+            Skipped,
+        ],
+    ]
+    "###);
+
+    search.time_budget(TimeBudget::max().with_stop_after(2));
+
+    let result = search.execute().unwrap();
+    snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
+    IDs: [0, 1, 2, 3]
+    Scores: 0.5000 0.0000 0.0000 0.0000 
+    Score Details:
+    [
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.0,
+                    ),
+                },
+            ),
+        ],
+        [
+            Skipped,
+        ],
+        [
+            Skipped,
+        ],
+    ]
+    "###);
+
+    search.time_budget(TimeBudget::max().with_stop_after(3));
+
+    let result = search.execute().unwrap();
+    snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
+    IDs: [2, 0, 1, 3]
+    Scores: 1.0000 0.5000 0.0000 0.0000 
+    Score Details:
+    [
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        1.0,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.0,
+                    ),
+                },
+            ),
+        ],
+        [
+            Skipped,
+        ],
+    ]
+    "###);
+
+    search.time_budget(TimeBudget::max().with_stop_after(4));
+
+    let result = search.execute().unwrap();
+    snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
+    IDs: [2, 0, 3, 1]
+    Scores: 1.0000 0.5000 0.5000 0.0000 
+    Score Details:
+    [
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        1.0,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.0,
+                    ),
+                },
+            ),
+        ],
+    ]
+    "###);
+
+    search.time_budget(TimeBudget::max().with_stop_after(5));
+
+    let result = search.execute().unwrap();
+    snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
+    IDs: [2, 0, 3, 1]
+    Scores: 1.0000 0.5000 0.5000 0.0000 
+    Score Details:
+    [
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        1.0,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.5,
+                    ),
+                },
+            ),
+        ],
+        [
+            Vector(
+                Vector {
+                    similarity: Some(
+                        0.0,
+                    ),
+                },
+            ),
+        ],
+    ]
+    "###);
 }

crates/milli/src/search/new/vector_sort.rs

@@ -1,4 +1,5 @@
 use std::iter::FromIterator;
+use std::task::Poll;
 use std::time::Instant;
 
 use roaring::RoaringBitmap;
@@ -7,7 +8,7 @@ use super::ranking_rules::{RankingRule, RankingRuleOutput, RankingRuleQueryTrait
 use super::VectorStoreStats;
 use crate::score_details::{self, ScoreDetails};
 use crate::vector::{DistributionShift, Embedder, VectorStore};
-use crate::{DocumentId, Result, SearchContext, SearchLogger};
+use crate::{DocumentId, Result, SearchContext, SearchLogger, TimeBudget};
 
 pub struct VectorSort<Q: RankingRuleQueryTrait> {
     query: Option<Q>,
@@ -52,6 +53,7 @@ impl<Q: RankingRuleQueryTrait> VectorSort<Q> {
         &mut self,
         ctx: &mut SearchContext<'_>,
         vector_candidates: &RoaringBitmap,
+        time_budget: &TimeBudget,
     ) -> Result<()> {
         let target = &self.target;
         let backend = ctx.index.get_vector_store(ctx.txn)?.unwrap_or_default();
@@ -59,7 +61,13 @@ impl<Q: RankingRuleQueryTrait> VectorSort<Q> {
         let before = Instant::now();
         let reader =
             VectorStore::new(backend, ctx.index.vector_store, self.embedder_index, self.quantized);
-        let results = reader.nns_by_vector(ctx.txn, target, self.limit, Some(vector_candidates))?;
+        let results = reader.nns_by_vector(
+            ctx.txn,
+            target,
+            self.limit,
+            Some(vector_candidates),
+            time_budget,
+        )?;
         self.cached_sorted_docids = results.into_iter();
         *ctx.vector_store_stats.get_or_insert_default() += VectorStoreStats {
             total_time: before.elapsed(),
@@ -69,6 +77,20 @@ impl<Q: RankingRuleQueryTrait> VectorSort<Q> {
 
         Ok(())
     }
+
+    fn next_result(&mut self, vector_candidates: &RoaringBitmap) -> Option<(DocumentId, f32)> {
+        for (docid, distance) in self.cached_sorted_docids.by_ref() {
+            if vector_candidates.contains(docid) {
+                let score = 1.0 - distance;
+                let score = self
+                    .distribution_shift
+                    .map(|distribution| distribution.shift(score))
+                    .unwrap_or(score);
+                return Some((docid, score));
+            }
+        }
+        None
+    }
 }
 
 impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
@@ -83,12 +105,13 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
         _logger: &mut dyn SearchLogger<Q>,
         universe: &RoaringBitmap,
         query: &Q,
+        time_budget: &TimeBudget,
     ) -> Result<()> {
         assert!(self.query.is_none());
 
         self.query = Some(query.clone());
         let vector_candidates = &self.vector_candidates & universe;
-        self.fill_buffer(ctx, &vector_candidates)?;
+        self.fill_buffer(ctx, &vector_candidates, time_budget)?;
         Ok(())
     }
 
@@ -99,6 +122,7 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
         ctx: &mut SearchContext<'ctx>,
         _logger: &mut dyn SearchLogger<Q>,
         universe: &RoaringBitmap,
+        time_budget: &TimeBudget,
     ) -> Result<Option<RankingRuleOutput<Q>>> {
         let query = self.query.as_ref().unwrap().clone();
         let vector_candidates = &self.vector_candidates & universe;
@@ -111,24 +135,17 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
             }));
         }
 
-        for (docid, distance) in self.cached_sorted_docids.by_ref() {
-            if vector_candidates.contains(docid) {
-                let score = 1.0 - distance;
-                let score = self
-                    .distribution_shift
-                    .map(|distribution| distribution.shift(score))
-                    .unwrap_or(score);
-                return Ok(Some(RankingRuleOutput {
-                    query,
-                    candidates: RoaringBitmap::from_iter([docid]),
-                    score: ScoreDetails::Vector(score_details::Vector { similarity: Some(score) }),
-                }));
-            }
+        if let Some((docid, score)) = self.next_result(&vector_candidates) {
+            return Ok(Some(RankingRuleOutput {
+                query,
+                candidates: RoaringBitmap::from_iter([docid]),
+                score: ScoreDetails::Vector(score_details::Vector { similarity: Some(score) }),
+            }));
         }
 
         // if we got out of this loop it means we've exhausted our cache.
         // we need to refill it and run the function again.
-        self.fill_buffer(ctx, &vector_candidates)?;
+        self.fill_buffer(ctx, &vector_candidates, time_budget)?;
 
         // we tried filling the buffer, but it remained empty 😢
         // it means we don't actually have any document remaining in the universe with a vector.
@@ -141,11 +158,39 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
             }));
         }
 
-        self.next_bucket(ctx, _logger, universe)
+        self.next_bucket(ctx, _logger, universe, time_budget)
     }
 
     #[tracing::instrument(level = "trace", skip_all, target = "search::vector_sort")]
     fn end_iteration(&mut self, _ctx: &mut SearchContext<'ctx>, _logger: &mut dyn SearchLogger<Q>) {
         self.query = None;
     }
+
+    fn non_blocking_next_bucket(
+        &mut self,
+        _ctx: &mut SearchContext<'ctx>,
+        _logger: &mut dyn SearchLogger<Q>,
+        universe: &RoaringBitmap,
+    ) -> Result<Poll<RankingRuleOutput<Q>>> {
+        let query = self.query.as_ref().unwrap().clone();
+        let vector_candidates = &self.vector_candidates & universe;
+
+        if vector_candidates.is_empty() {
+            return Ok(Poll::Ready(RankingRuleOutput {
+                query,
+                candidates: universe.clone(),
+                score: ScoreDetails::Vector(score_details::Vector { similarity: None }),
+            }));
+        }
+
+        if let Some((docid, score)) = self.next_result(&vector_candidates) {
+            Ok(Poll::Ready(RankingRuleOutput {
+                query,
+                candidates: RoaringBitmap::from_iter([docid]),
+                score: ScoreDetails::Vector(score_details::Vector { similarity: Some(score) }),
+            }))
+        } else {
+            Ok(Poll::Pending)
+        }
+    }
 }

crates/milli/src/update/indexer_config.rs

@@ -1,3 +1,6 @@
+use std::num::NonZeroUsize;
+use std::time::Duration;
+
 use grenad::CompressionType;
 
 use super::GrenadParameters;
@@ -20,6 +23,7 @@ pub struct IndexerConfig {
     pub experimental_no_edition_2024_for_dumps: bool,
     pub experimental_no_edition_2024_for_prefix_post_processing: bool,
     pub experimental_no_edition_2024_for_facet_post_processing: bool,
+    pub s3_snapshot_options: Option<S3SnapshotOptions>,
 }
 
 impl IndexerConfig {
@@ -37,6 +41,20 @@ impl IndexerConfig {
     }
 }
 
+#[derive(Debug, Clone)]
+pub struct S3SnapshotOptions {
+    pub s3_bucket_url: String,
+    pub s3_bucket_region: String,
+    pub s3_bucket_name: String,
+    pub s3_snapshot_prefix: String,
+    pub s3_access_key: String,
+    pub s3_secret_key: String,
+    pub s3_max_in_flight_parts: NonZeroUsize,
+    pub s3_compression_level: u32,
+    pub s3_signature_duration: Duration,
+    pub s3_multipart_part_size: u64,
+}
+
 /// By default use only 1 thread for indexing in tests
 #[cfg(test)]
 pub fn default_thread_pool_and_threads() -> (ThreadPoolNoAbort, Option<usize>) {
@@ -76,6 +94,7 @@ impl Default for IndexerConfig {
             experimental_no_edition_2024_for_dumps: false,
             experimental_no_edition_2024_for_prefix_post_processing: false,
             experimental_no_edition_2024_for_facet_post_processing: false,
+            s3_snapshot_options: None,
         }
     }
 }

crates/milli/src/update/mod.rs

@@ -5,7 +5,7 @@ pub use self::concurrent_available_ids::ConcurrentAvailableIds;
 pub use self::facet::bulk::FacetsUpdateBulk;
 pub use self::facet::incremental::FacetsUpdateIncrementalInner;
 pub use self::index_documents::{request_threads, *};
-pub use self::indexer_config::{default_thread_pool_and_threads, IndexerConfig};
+pub use self::indexer_config::{default_thread_pool_and_threads, IndexerConfig, S3SnapshotOptions};
 pub use self::new::ChannelCongestion;
 pub use self::settings::{validate_embedding_settings, Setting, Settings};
 pub use self::update_step::UpdateIndexingStep;

crates/milli/src/update/new/words_prefix_docids.rs

@@ -4,7 +4,7 @@ use std::io::{BufReader, BufWriter, Read, Seek, Write};
 use std::iter;
 
 use hashbrown::HashMap;
-use heed::types::{Bytes, DecodeIgnore};
+use heed::types::{Bytes, DecodeIgnore, Str};
 use heed::{BytesDecode, Database, Error, RoTxn, RwTxn};
 use rayon::iter::{IndexedParallelIterator as _, IntoParallelIterator, ParallelIterator as _};
 use roaring::MultiOps;
@@ -16,22 +16,29 @@ use crate::heed_codec::StrBEU16Codec;
 use crate::update::GrenadParameters;
 use crate::{CboRoaringBitmapCodec, Index, Prefix, Result};
 
-struct WordPrefixDocids {
+struct WordPrefixDocids<'i> {
+    index: &'i Index,
     database: Database<Bytes, CboRoaringBitmapCodec>,
     prefix_database: Database<Bytes, CboRoaringBitmapCodec>,
     max_memory_by_thread: Option<usize>,
+    /// Do not use an experimental LMDB feature to read uncommitted data in parallel.
+    no_experimental_post_processing: bool,
 }
 
-impl WordPrefixDocids {
+impl<'i> WordPrefixDocids<'i> {
     fn new(
+        index: &'i Index,
         database: Database<Bytes, CboRoaringBitmapCodec>,
         prefix_database: Database<Bytes, CboRoaringBitmapCodec>,
         grenad_parameters: &GrenadParameters,
-    ) -> WordPrefixDocids {
+    ) -> WordPrefixDocids<'i> {
         WordPrefixDocids {
+            index,
             database,
             prefix_database,
             max_memory_by_thread: grenad_parameters.max_memory_by_thread(),
+            no_experimental_post_processing: grenad_parameters
+                .experimental_no_edition_2024_for_prefix_post_processing,
         }
     }
 
@@ -42,7 +49,77 @@ impl WordPrefixDocids {
         prefix_to_delete: &BTreeSet<Prefix>,
     ) -> Result<()> {
         delete_prefixes(wtxn, &self.prefix_database, prefix_to_delete)?;
-        self.recompute_modified_prefixes(wtxn, prefix_to_compute)
+        if self.no_experimental_post_processing {
+            self.recompute_modified_prefixes(wtxn, prefix_to_compute)
+        } else {
+            self.recompute_modified_prefixes_no_frozen(wtxn, prefix_to_compute)
+        }
+    }
+
+    #[tracing::instrument(level = "trace", skip_all, target = "indexing::prefix")]
+    fn recompute_modified_prefixes_no_frozen(
+        &self,
+        wtxn: &mut RwTxn,
+        prefix_to_compute: &BTreeSet<Prefix>,
+    ) -> Result<()> {
+        let thread_count = rayon::current_num_threads();
+        let rtxns = iter::repeat_with(|| self.index.env.nested_read_txn(wtxn))
+            .take(thread_count)
+            .collect::<heed::Result<Vec<_>>>()?;
+
+        let outputs = rtxns
+            .into_par_iter()
+            .enumerate()
+            .map(|(thread_id, rtxn)| {
+                // `indexes` represent offsets at which prefixes computations were stored in the `file`.
+                let mut indexes = Vec::new();
+                let mut file = BufWriter::new(spooled_tempfile(
+                    self.max_memory_by_thread.unwrap_or(usize::MAX),
+                ));
+
+                let mut buffer = Vec::new();
+                for (prefix_index, prefix) in prefix_to_compute.iter().enumerate() {
+                    // Is prefix for another thread?
+                    if prefix_index % thread_count != thread_id {
+                        continue;
+                    }
+
+                    let output = self
+                        .database
+                        .prefix_iter(&rtxn, prefix.as_bytes())?
+                        .remap_types::<Str, CboRoaringBitmapCodec>()
+                        .map(|result| result.map(|(_word, bitmap)| bitmap))
+                        .union()?;
+
+                    buffer.clear();
+                    CboRoaringBitmapCodec::serialize_into_vec(&output, &mut buffer);
+                    indexes.push(PrefixEntry { prefix, serialized_length: buffer.len() });
+                    file.write_all(&buffer)?;
+                }
+
+                Ok((indexes, file))
+            })
+            .collect::<Result<Vec<_>>>()?;
+
+        // We iterate over all the collected and serialized bitmaps through
+        // the files and entries to eventually put them in the final database.
+        let mut buffer = Vec::new();
+        for (index, file) in outputs {
+            let mut file = file.into_inner().map_err(|e| e.into_error())?;
+            file.rewind()?;
+            let mut file = BufReader::new(file);
+            for PrefixEntry { prefix, serialized_length } in index {
+                buffer.resize(serialized_length, 0);
+                file.read_exact(&mut buffer)?;
+                self.prefix_database.remap_data_type::<Bytes>().put(
+                    wtxn,
+                    prefix.as_bytes(),
+                    &buffer,
+                )?;
+            }
+        }
+
+        Ok(())
     }
 
     #[tracing::instrument(level = "trace", skip_all, target = "indexing::prefix")]
@@ -463,6 +540,7 @@ pub fn compute_word_prefix_docids(
     grenad_parameters: &GrenadParameters,
 ) -> Result<()> {
     WordPrefixDocids::new(
+        index,
         index.word_docids.remap_key_type(),
         index.word_prefix_docids.remap_key_type(),
         grenad_parameters,
@@ -479,6 +557,7 @@ pub fn compute_exact_word_prefix_docids(
     grenad_parameters: &GrenadParameters,
 ) -> Result<()> {
     WordPrefixDocids::new(
+        index,
         index.exact_word_docids.remap_key_type(),
         index.exact_word_prefix_docids.remap_key_type(),
         grenad_parameters,

crates/milli/src/update/upgrade/mod.rs

@@ -41,6 +41,7 @@ const UPGRADE_FUNCTIONS: &[&dyn UpgradeIndex] = &[
     &ToTargetNoOp { target: (1, 21, 0) },
     &ToTargetNoOp { target: (1, 22, 0) },
     &ToTargetNoOp { target: (1, 23, 0) },
+    &ToTargetNoOp { target: (1, 24, 0) },
     // This is the last upgrade function, it will be called when the index is up to date.
     // any other upgrade function should be added before this one.
     &ToCurrentNoOp {},
@@ -75,6 +76,7 @@ const fn start(from: (u32, u32, u32)) -> Option<usize> {
         (1, 21, _) => function_index!(11),
         (1, 22, _) => function_index!(12),
         (1, 23, _) => function_index!(13),
+        (1, 24, _) => function_index!(14),
         // We deliberately don't add a placeholder with (VERSION_MAJOR, VERSION_MINOR, VERSION_PATCH) here to force manually
         // considering dumpless upgrade.
         (_major, _minor, _patch) => return None,

crates/milli/src/vector/store.rs

@@ -8,6 +8,7 @@ use serde::{Deserialize, Serialize};
 
 use crate::progress::Progress;
 use crate::vector::Embeddings;
+use crate::TimeBudget;
 
 const HANNOY_EF_CONSTRUCTION: usize = 125;
 const HANNOY_M: usize = 16;
@@ -591,6 +592,7 @@ impl VectorStore {
         vector: &[f32],
         limit: usize,
         filter: Option<&RoaringBitmap>,
+        time_budget: &TimeBudget,
     ) -> crate::Result<Vec<(ItemId, f32)>> {
         if self.backend == VectorStoreBackend::Arroy {
             if self.quantized {
@@ -601,11 +603,25 @@ impl VectorStore {
                     .map_err(Into::into)
             }
         } else if self.quantized {
-            self._hannoy_nns_by_vector(rtxn, self._hannoy_quantized_db(), vector, limit, filter)
-                .map_err(Into::into)
+            self._hannoy_nns_by_vector(
+                rtxn,
+                self._hannoy_quantized_db(),
+                vector,
+                limit,
+                filter,
+                time_budget,
+            )
+            .map_err(Into::into)
         } else {
-            self._hannoy_nns_by_vector(rtxn, self._hannoy_angular_db(), vector, limit, filter)
-                .map_err(Into::into)
+            self._hannoy_nns_by_vector(
+                rtxn,
+                self._hannoy_angular_db(),
+                vector,
+                limit,
+                filter,
+                time_budget,
+            )
+            .map_err(Into::into)
         }
     }
     pub fn item_vectors(&self, rtxn: &RoTxn, item_id: u32) -> crate::Result<Vec<Vec<f32>>> {
@@ -1000,6 +1016,7 @@ impl VectorStore {
         vector: &[f32],
         limit: usize,
         filter: Option<&RoaringBitmap>,
+        time_budget: &TimeBudget,
     ) -> Result<Vec<(ItemId, f32)>, hannoy::Error> {
         let mut results = Vec::new();
 
@@ -1011,7 +1028,10 @@ impl VectorStore {
                 searcher.candidates(filter);
             }
 
-            results.append(&mut searcher.by_vector(rtxn, vector)?);
+            let (res, _degraded) =
+                &mut searcher
+                    .by_vector_with_cancellation(rtxn, vector, || time_budget.exceeded())?;
+            results.append(res);
         }
 
         results.sort_unstable_by_key(|(_, distance)| OrderedFloat(*distance));

documentation/prototypes.md

@@ -20,14 +20,16 @@ These make us iterate fast before stabilizing it for the current release.
 
 ### Release steps
 
-The prototype name must follow this convention: `prototype-X-Y` where
-- `X` is the feature name formatted in `kebab-case`. It should not end with a single number.
+The prototype name must follow this convention: `prototype-v<version>.<name>-<number>` where
+- `version` is the version of Meilisearch on which the prototype is based.
+- `name` is the feature name formatted in `kebab-case`. It should not end with a single number.
 - `Y` is the version of the prototype, starting from `0`.
 
-✅ Example: `prototype-auto-resize-0`. </br>
-❌ Bad example: `auto-resize-0`: lacks the `prototype` prefix. </br>
-❌ Bad example: `prototype-auto-resize`: lacks the version suffix. </br>
-❌ Bad example: `prototype-auto-resize-0-0`: feature name ends with a single number.
+✅ Example: `prototype-v1.23.0.search-personalization-0`. </br>
+❌ Bad example: `prototype-search-personalization-0`: version is missing.</br>
+❌ Bad example: `v1.23.0.auto-resize-0`: lacks the `prototype` prefix. </br>
+❌ Bad example: `prototype-v1.23.0.auto-resize`: lacks the version suffix. </br>
+❌ Bad example: `prototype-v1.23.0.auto-resize-0-0`: feature name ends with a single number.
 
 Steps to create a prototype: