Improve code readability

Fix chat settings dumpless upgrade

.gitignore

@@ -11,12 +11,18 @@
 /bench
 /_xtask_benchmark.ms
 /benchmarks
+.DS_Store
 
 # Snapshots
 ## ... large
 *.full.snap
-##  ... unreviewed
+## ... unreviewed
 *.snap.new
+## ... pending
+*.pending-snap
+
+# Tmp files
+.tmp*
 
 # Database snapshot
 crates/meilisearch/db.snapshot

crates/benchmarks/Cargo.toml

@@ -51,3 +51,8 @@ harness = false
 [[bench]]
 name = "indexing"
 harness = false
+
+[[bench]]
+name = "sort"
+harness = false
+

crates/benchmarks/benches/sort.rs

@@ -0,0 +1,114 @@
+//! This benchmark module is used to compare the performance of sorting documents in /search VS /documents
+//!
+//! The tests/benchmarks were designed in the context of a query returning only 20 documents.
+
+mod datasets_paths;
+mod utils;
+
+use criterion::{criterion_group, criterion_main};
+use milli::update::Settings;
+use utils::Conf;
+
+#[cfg(not(windows))]
+#[global_allocator]
+static ALLOC: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
+fn base_conf(builder: &mut Settings) {
+    let displayed_fields =
+        ["geonameid", "name", "asciiname", "alternatenames", "_geo", "population"]
+            .iter()
+            .map(|s| s.to_string())
+            .collect();
+    builder.set_displayed_fields(displayed_fields);
+
+    let sortable_fields =
+        ["_geo", "name", "population", "elevation", "timezone", "modification-date"]
+            .iter()
+            .map(|s| s.to_string())
+            .collect();
+    builder.set_sortable_fields(sortable_fields);
+}
+
+#[rustfmt::skip]
+const BASE_CONF: Conf = Conf {
+    dataset: datasets_paths::SMOL_ALL_COUNTRIES,
+    dataset_format: "jsonl",
+    configure: base_conf,
+    primary_key: Some("geonameid"),
+    queries: &[""],
+    offsets: &[
+        Some((0, 20)), // The most common query in the real world
+        Some((0, 500)), // A query that ranges over many documents
+        Some((980, 20)), // The worst query that could happen in the real world
+        Some((800_000, 20)) // The worst query
+    ],
+    get_documents: true,
+    ..Conf::BASE
+};
+
+fn bench_sort(c: &mut criterion::Criterion) {
+    #[rustfmt::skip]
+    let confs = &[
+        utils::Conf {
+            group_name: "without sort",
+            sort: None,
+            ..BASE_CONF
+        },
+
+        utils::Conf {
+            group_name: "sort on many different values",
+            sort: Some(vec!["name:asc"]),
+            ..BASE_CONF
+        },
+
+        utils::Conf {
+            group_name: "sort on many similar values",
+            sort: Some(vec!["timezone:desc"]),
+            ..BASE_CONF
+        },
+
+        utils::Conf {
+            group_name: "sort on many similar then different values",
+            sort: Some(vec!["timezone:desc", "name:asc"]),
+            ..BASE_CONF
+        },
+
+        utils::Conf {
+            group_name: "sort on many different then similar values",
+            sort: Some(vec!["timezone:desc", "name:asc"]),
+            ..BASE_CONF
+        },
+
+        utils::Conf {
+            group_name: "geo sort",
+            sample_size: Some(10),
+            sort: Some(vec!["_geoPoint(45.4777599, 9.1967508):asc"]),
+            ..BASE_CONF
+        },
+
+        utils::Conf {
+            group_name: "sort on many similar values then geo sort",
+            sample_size: Some(50),
+            sort: Some(vec!["timezone:desc", "_geoPoint(45.4777599, 9.1967508):asc"]),
+            ..BASE_CONF
+        },
+
+        utils::Conf {
+            group_name: "sort on many different values then geo sort",
+            sample_size: Some(50),
+            sort: Some(vec!["name:desc", "_geoPoint(45.4777599, 9.1967508):asc"]),
+            ..BASE_CONF
+        },
+
+        utils::Conf {
+            group_name: "sort on many fields",
+            sort: Some(vec!["population:asc", "name:asc", "elevation:asc", "timezone:asc"]),
+            ..BASE_CONF
+        },
+    ];
+
+    utils::run_benches(c, confs);
+}
+
+criterion_group!(benches, bench_sort);
+criterion_main!(benches);

crates/benchmarks/benches/utils.rs

@@ -9,6 +9,7 @@ use anyhow::Context;
 use bumpalo::Bump;
 use criterion::BenchmarkId;
 use memmap2::Mmap;
+use milli::documents::sort::recursive_sort;
 use milli::heed::EnvOpenOptions;
 use milli::progress::Progress;
 use milli::update::new::indexer;
@@ -35,6 +36,12 @@ pub struct Conf<'a> {
     pub configure: fn(&mut Settings),
     pub filter: Option<&'a str>,
     pub sort: Option<Vec<&'a str>>,
+    /// set to skip documents (offset, limit)
+    pub offsets: &'a [Option<(usize, usize)>],
+    /// enable if you want to bench getting documents without querying
+    pub get_documents: bool,
+    /// configure the benchmark sample size
+    pub sample_size: Option<usize>,
     /// enable or disable the optional words on the query
     pub optional_words: bool,
     /// primary key, if there is None we'll auto-generate docids for every documents
@@ -52,6 +59,9 @@ impl Conf<'_> {
         configure: |_| (),
         filter: None,
         sort: None,
+        offsets: &[None],
+        get_documents: false,
+        sample_size: None,
         optional_words: true,
         primary_key: None,
     };
@@ -145,25 +155,79 @@ pub fn run_benches(c: &mut criterion::Criterion, confs: &[Conf]) {
         let file_name = Path::new(conf.dataset).file_name().and_then(|f| f.to_str()).unwrap();
         let name = format!("{}: {}", file_name, conf.group_name);
         let mut group = c.benchmark_group(&name);
+        if let Some(sample_size) = conf.sample_size {
+            group.sample_size(sample_size);
+        }
 
         for &query in conf.queries {
-            group.bench_with_input(BenchmarkId::from_parameter(query), &query, |b, &query| {
-                b.iter(|| {
-                    let rtxn = index.read_txn().unwrap();
-                    let mut search = index.search(&rtxn);
-                    search.query(query).terms_matching_strategy(TermsMatchingStrategy::default());
-                    if let Some(filter) = conf.filter {
-                        let filter = Filter::from_str(filter).unwrap().unwrap();
-                        search.filter(filter);
-                    }
-                    if let Some(sort) = &conf.sort {
-                        let sort = sort.iter().map(|sort| sort.parse().unwrap()).collect();
-                        search.sort_criteria(sort);
-                    }
-                    let _ids = search.execute().unwrap();
-                });
-            });
+            for offset in conf.offsets {
+                let parameter = match offset {
+                    None => query.to_string(),
+                    Some((offset, limit)) => format!("{query}[{offset}:{limit}]"),
+                };
+                group.bench_with_input(
+                    BenchmarkId::from_parameter(parameter),
+                    &query,
+                    |b, &query| {
+                        b.iter(|| {
+                            let rtxn = index.read_txn().unwrap();
+                            let mut search = index.search(&rtxn);
+                            search
+                                .query(query)
+                                .terms_matching_strategy(TermsMatchingStrategy::default());
+                            if let Some(filter) = conf.filter {
+                                let filter = Filter::from_str(filter).unwrap().unwrap();
+                                search.filter(filter);
+                            }
+                            if let Some(sort) = &conf.sort {
+                                let sort = sort.iter().map(|sort| sort.parse().unwrap()).collect();
+                                search.sort_criteria(sort);
+                            }
+                            if let Some((offset, limit)) = offset {
+                                search.offset(*offset).limit(*limit);
+                            }
+
+                            let _ids = search.execute().unwrap();
+                        });
+                    },
+                );
+            }
         }
+
+        if conf.get_documents {
+            for offset in conf.offsets {
+                let parameter = match offset {
+                    None => String::from("get_documents"),
+                    Some((offset, limit)) => format!("get_documents[{offset}:{limit}]"),
+                };
+                group.bench_with_input(BenchmarkId::from_parameter(parameter), &(), |b, &()| {
+                    b.iter(|| {
+                        let rtxn = index.read_txn().unwrap();
+                        if let Some(sort) = &conf.sort {
+                            let sort = sort.iter().map(|sort| sort.parse().unwrap()).collect();
+                            let all_docs = index.documents_ids(&rtxn).unwrap();
+                            let facet_sort =
+                                recursive_sort(&index, &rtxn, sort, &all_docs).unwrap();
+                            let iter = facet_sort.iter().unwrap();
+                            if let Some((offset, limit)) = offset {
+                                let _results = iter.skip(*offset).take(*limit).collect::<Vec<_>>();
+                            } else {
+                                let _results = iter.collect::<Vec<_>>();
+                            }
+                        } else {
+                            let all_docs = index.documents_ids(&rtxn).unwrap();
+                            if let Some((offset, limit)) = offset {
+                                let _results =
+                                    all_docs.iter().skip(*offset).take(*limit).collect::<Vec<_>>();
+                            } else {
+                                let _results = all_docs.iter().collect::<Vec<_>>();
+                            }
+                        }
+                    });
+                });
+            }
+        }
+
         group.finish();
 
         index.prepare_for_closing().wait();

crates/filter-parser/src/lib.rs

@@ -60,7 +60,7 @@ use nom::combinator::{cut, eof, map, opt};
 use nom::multi::{many0, separated_list1};
 use nom::number::complete::recognize_float;
 use nom::sequence::{delimited, preceded, terminated, tuple};
-use nom::Finish;
+use nom::{Finish, Slice};
 use nom_locate::LocatedSpan;
 pub(crate) use value::parse_value;
 use value::word_exact;
@@ -121,6 +121,16 @@ impl<'a> Token<'a> {
             Err(Error::new_from_kind(self.span, ErrorKind::NonFiniteFloat))
         }
     }
+
+    /// Split the token by a delimiter and return an iterator of tokens.
+    /// Each token in the iterator will have its own span that corresponds to a slice of the original token's span.
+    pub fn split(&self, delimiter: &'a str) -> impl Iterator<Item = Token<'a>> + '_ {
+        let original_addr = self.value().as_ptr() as usize;
+        self.value().split(delimiter).map(move |part| {
+            let offset = part.as_ptr() as usize - original_addr;
+            Token::new(self.span.slice(offset..offset + part.len()), Some(part.to_string()))
+        })
+    }
 }
 
 impl<'a> From<Span<'a>> for Token<'a> {
@@ -179,6 +189,25 @@ impl<'a> FilterCondition<'a> {
         }
     }
 
+    pub fn use_vector_filter(&self) -> Option<&Token> {
+        match self {
+            FilterCondition::Condition { fid, op: _ } => {
+                if fid.value().starts_with("_vectors.") || fid.value() == "_vectors" {
+                    Some(fid)
+                } else {
+                    None
+                }
+            }
+            FilterCondition::Not(this) => this.use_vector_filter(),
+            FilterCondition::Or(seq) | FilterCondition::And(seq) => {
+                seq.iter().find_map(|filter| filter.use_vector_filter())
+            }
+            FilterCondition::GeoLowerThan { .. }
+            | FilterCondition::GeoBoundingBox { .. }
+            | FilterCondition::In { .. } => None,
+        }
+    }
+
     pub fn fids(&self, depth: usize) -> Box<dyn Iterator<Item = &Token> + '_> {
         if depth == 0 {
             return Box::new(std::iter::empty());
@@ -1043,4 +1072,103 @@ pub mod tests {
         let token: Token = s.into();
         assert_eq!(token.value(), s);
     }
+
+    #[test]
+    fn split() {
+        let s = "test string that should not be parsed\n newline";
+        let token: Token = s.into();
+        let parts: Vec<_> = token.split(" ").collect();
+        insta::assert_snapshot!(format!("{parts:#?}"), @r#"
+        [
+            Token {
+                span: LocatedSpan {
+                    offset: 0,
+                    line: 1,
+                    fragment: "test",
+                    extra: "test string that should not be parsed\n newline",
+                },
+                value: Some(
+                    "test",
+                ),
+            },
+            Token {
+                span: LocatedSpan {
+                    offset: 5,
+                    line: 1,
+                    fragment: "string",
+                    extra: "test string that should not be parsed\n newline",
+                },
+                value: Some(
+                    "string",
+                ),
+            },
+            Token {
+                span: LocatedSpan {
+                    offset: 12,
+                    line: 1,
+                    fragment: "that",
+                    extra: "test string that should not be parsed\n newline",
+                },
+                value: Some(
+                    "that",
+                ),
+            },
+            Token {
+                span: LocatedSpan {
+                    offset: 17,
+                    line: 1,
+                    fragment: "should",
+                    extra: "test string that should not be parsed\n newline",
+                },
+                value: Some(
+                    "should",
+                ),
+            },
+            Token {
+                span: LocatedSpan {
+                    offset: 24,
+                    line: 1,
+                    fragment: "not",
+                    extra: "test string that should not be parsed\n newline",
+                },
+                value: Some(
+                    "not",
+                ),
+            },
+            Token {
+                span: LocatedSpan {
+                    offset: 28,
+                    line: 1,
+                    fragment: "be",
+                    extra: "test string that should not be parsed\n newline",
+                },
+                value: Some(
+                    "be",
+                ),
+            },
+            Token {
+                span: LocatedSpan {
+                    offset: 31,
+                    line: 1,
+                    fragment: "parsed\n",
+                    extra: "test string that should not be parsed\n newline",
+                },
+                value: Some(
+                    "parsed\n",
+                ),
+            },
+            Token {
+                span: LocatedSpan {
+                    offset: 39,
+                    line: 2,
+                    fragment: "newline",
+                    extra: "test string that should not be parsed\n newline",
+                },
+                value: Some(
+                    "newline",
+                ),
+            },
+        ]
+        "#);
+    }
 }

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

@@ -62,13 +62,14 @@ impl IndexScheduler {
             let ExportIndexSettings { filter, override_settings } = export_settings;
             let index = self.index(uid)?;
             let index_rtxn = index.read_txn()?;
+            let bearer = api_key.map(|api_key| format!("Bearer {api_key}"));
 
             // First, check if the index already exists
             let url = format!("{base_url}/indexes/{uid}");
             let response = retry(&must_stop_processing, || {
                 let mut request = agent.get(&url);
-                if let Some(api_key) = api_key {
-                    request = request.set("Authorization", &format!("Bearer {api_key}"));
+                if let Some(bearer) = &bearer {
+                    request = request.set("Authorization", bearer);
                 }
 
                 request.send_bytes(Default::default()).map_err(into_backoff_error)
@@ -90,8 +91,8 @@ impl IndexScheduler {
                 let url = format!("{base_url}/indexes");
                 retry(&must_stop_processing, || {
                     let mut request = agent.post(&url);
-                    if let Some(api_key) = api_key {
-                        request = request.set("Authorization", &format!("Bearer {api_key}"));
+                    if let Some(bearer) = &bearer {
+                        request = request.set("Authorization", bearer);
                     }
                     let index_param = json!({ "uid": uid, "primaryKey": primary_key });
                     request.send_json(&index_param).map_err(into_backoff_error)
@@ -103,8 +104,8 @@ impl IndexScheduler {
                 let url = format!("{base_url}/indexes/{uid}");
                 retry(&must_stop_processing, || {
                     let mut request = agent.patch(&url);
-                    if let Some(api_key) = api_key {
-                        request = request.set("Authorization", &format!("Bearer {api_key}"));
+                    if let Some(bearer) = &bearer {
+                        request = request.set("Authorization", bearer);
                     }
                     let index_param = json!({ "primaryKey": primary_key });
                     request.send_json(&index_param).map_err(into_backoff_error)
@@ -122,7 +123,6 @@ impl IndexScheduler {
                 }
                 // Retry logic for sending settings
                 let url = format!("{base_url}/indexes/{uid}/settings");
-                let bearer = api_key.map(|api_key| format!("Bearer {api_key}"));
                 retry(&must_stop_processing, || {
                     let mut request = agent.patch(&url);
                     if let Some(bearer) = bearer.as_ref() {
@@ -167,10 +167,10 @@ impl IndexScheduler {
                 },
             );
 
-            let limit = payload_size.map(|ps| ps.as_u64() as usize).unwrap_or(50 * 1024 * 1024); // defaults to 50 MiB
+            let limit = payload_size.map(|ps| ps.as_u64() as usize).unwrap_or(20 * 1024 * 1024); // defaults to 20 MiB
             let documents_url = format!("{base_url}/indexes/{uid}/documents");
 
-            request_threads()
+            let results = request_threads()
                 .broadcast(|ctx| {
                     let index_rtxn = index
                         .read_txn()
@@ -265,9 +265,8 @@ impl IndexScheduler {
                                 let mut request = agent.post(&documents_url);
                                 request = request.set("Content-Type", "application/x-ndjson");
                                 request = request.set("Content-Encoding", "gzip");
-                                if let Some(api_key) = api_key {
-                                    request = request
-                                        .set("Authorization", &(format!("Bearer {api_key}")));
+                                if let Some(bearer) = &bearer {
+                                    request = request.set("Authorization", bearer);
                                 }
                                 request.send_bytes(&compressed_buffer).map_err(into_backoff_error)
                             })?;
@@ -276,7 +275,7 @@ impl IndexScheduler {
                         }
                         buffer.extend_from_slice(&tmp_buffer);
 
-                        if i % 100 == 0 {
+                        if i > 0 && i % 100 == 0 {
                             step.fetch_add(100, atomic::Ordering::Relaxed);
                         }
                     }
@@ -284,8 +283,8 @@ impl IndexScheduler {
                     retry(&must_stop_processing, || {
                         let mut request = agent.post(&documents_url);
                         request = request.set("Content-Type", "application/x-ndjson");
-                        if let Some(api_key) = api_key {
-                            request = request.set("Authorization", &(format!("Bearer {api_key}")));
+                        if let Some(bearer) = &bearer {
+                            request = request.set("Authorization", bearer);
                         }
                         request.send_bytes(&buffer).map_err(into_backoff_error)
                     })?;
@@ -298,6 +297,9 @@ impl IndexScheduler {
                         Some(uid.to_string()),
                     )
                 })?;
+            for result in results {
+                result?;
+            }
 
             step.store(total_documents, atomic::Ordering::Relaxed);
         }

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

@@ -736,7 +736,7 @@ fn test_document_addition_mixed_rights_with_index() {
 #[test]
 fn test_document_addition_mixed_right_without_index_starts_with_cant_create() {
     // We're going to autobatch multiple document addition.
-    // - The index does not exists
+    // - The index does not exist
     // - The first document addition don't have the right to create an index
     // - The second do. They should not batch together.
     // - The second should batch with everything else as it's going to create an index.

crates/meilisearch-types/src/error.rs

@@ -237,6 +237,7 @@ InvalidDocumentRetrieveVectors                 , InvalidRequest       , BAD_REQU
 MissingDocumentFilter                          , InvalidRequest       , BAD_REQUEST ;
 MissingDocumentEditionFunction                 , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentFilter                          , InvalidRequest       , BAD_REQUEST ;
+InvalidDocumentSort                            , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentGeoField                        , InvalidRequest       , BAD_REQUEST ;
 InvalidVectorDimensions                        , InvalidRequest       , BAD_REQUEST ;
 InvalidVectorsType                             , InvalidRequest       , BAD_REQUEST ;
@@ -415,6 +416,7 @@ InvalidChatCompletionPrompts                   , InvalidRequest       , BAD_REQU
 InvalidChatCompletionSystemPrompt              , InvalidRequest       , BAD_REQUEST ;
 InvalidChatCompletionSearchDescriptionPrompt   , InvalidRequest       , BAD_REQUEST ;
 InvalidChatCompletionSearchQueryParamPrompt    , InvalidRequest       , BAD_REQUEST ;
+InvalidChatCompletionSearchFilterParamPrompt   , InvalidRequest       , BAD_REQUEST ;
 InvalidChatCompletionSearchIndexUidParamPrompt , InvalidRequest       , BAD_REQUEST ;
 InvalidChatCompletionPreQueryPrompt            , InvalidRequest       , BAD_REQUEST
 }
@@ -476,7 +478,8 @@ impl ErrorCode for milli::Error {
                     UserError::InvalidDistinctAttribute { .. } => Code::InvalidSearchDistinct,
                     UserError::SortRankingRuleMissing => Code::InvalidSearchSort,
                     UserError::InvalidFacetsDistribution { .. } => Code::InvalidSearchFacets,
-                    UserError::InvalidSortableAttribute { .. } => Code::InvalidSearchSort,
+                    UserError::InvalidSearchSortableAttribute { .. } => Code::InvalidSearchSort,
+                    UserError::InvalidDocumentSortableAttribute { .. } => Code::InvalidDocumentSort,
                     UserError::InvalidSearchableAttribute { .. } => {
                         Code::InvalidSearchAttributesToSearchOn
                     }
@@ -492,7 +495,8 @@ impl ErrorCode for milli::Error {
                     UserError::InvalidVectorsMapType { .. }
                     | UserError::InvalidVectorsEmbedderConf { .. } => Code::InvalidVectorsType,
                     UserError::TooManyVectors(_, _) => Code::TooManyVectors,
-                    UserError::SortError(_) => Code::InvalidSearchSort,
+                    UserError::SortError { search: true, .. } => Code::InvalidSearchSort,
+                    UserError::SortError { search: false, .. } => Code::InvalidDocumentSort,
                     UserError::InvalidMinTypoWordLenSetting(_, _) => {
                         Code::InvalidSettingsTypoTolerance
                     }

crates/meilisearch-types/src/features.rs

@@ -4,10 +4,11 @@ use serde::{Deserialize, Serialize};
 
 use crate::error::{Code, ResponseError};
 
-pub const DEFAULT_CHAT_SYSTEM_PROMPT: &str = "You are a highly capable research assistant with access to powerful search tools. IMPORTANT INSTRUCTIONS:1. When answering questions, you MUST make multiple tool calls (at least 2-3) to gather comprehensive information.2. Use different search queries for each tool call - vary keywords, rephrase questions, and explore different semantic angles to ensure broad coverage.3. Always explicitly announce BEFORE making each tool call by saying: \"I'll search for [specific information] now.\"4. Combine information from ALL tool calls to provide complete, nuanced answers rather than relying on a single source.5. For complex topics, break down your research into multiple targeted queries rather than using a single generic search.";
+pub const DEFAULT_CHAT_SYSTEM_PROMPT: &str = "You are a highly capable research assistant with access to powerful search tools. IMPORTANT INSTRUCTIONS:1. When answering questions, you MUST make multiple tool calls (at least 2-3) to gather comprehensive information.2. Use different search queries for each tool call - vary keywords, rephrase questions, and explore different semantic angles to ensure broad coverage.3. Always explicitly announce BEFORE making each tool call by saying: \"I'll search for [specific information] now.\"4. Combine information from ALL tool calls to provide complete, nuanced answers rather than relying on a single source.5. For complex topics, break down your research into multiple targeted queries rather than using a single generic search. Meilisearch doesn't use the colon (:) syntax to filter but rather the equal (=) one. Separate filters from query and keep the q parameter empty if needed. Same for the filter parameter: keep it empty if need be. If you need to find documents that CONTAINS keywords simply put the keywords in the q parameter do no use a filter for this purpose. Whenever you get an error, read the error message and fix your error. ";
 pub const DEFAULT_CHAT_SEARCH_DESCRIPTION_PROMPT: &str =
-    "Search the database for relevant JSON documents using an optional query.";
+    "Query: 'best story about Rust before 2018' with year: 2018, 2020, 2021\nlabel: analysis, golang, javascript\ntype: story, link\nvote: 300, 298, 278\n: {\"q\": \"\", \"filter\": \"category = Rust AND type = story AND year < 2018 AND vote > 100\"}\nQuery: 'A black or green car that can go fast with red brakes' with maxspeed_kmh: 200, 150, 130\ncolor: black, grey, red, green\nbrand: Toyota, Renault, Jeep, Ferrari\n: {\"q\": \"red brakes\", \"filter\": \"maxspeed_kmh > 150 AND color IN ['black', green]\"}\nQuery: 'Superman movie released in 2018 or after' with year: 2018, 2020, 2021\ngenres: Drama, Comedy, Adventure, Fiction\n: {\"q\":\"Superman\",\"filter\":\"genres IN [Adventure, Fiction] AND year >= 2018\"}";
 pub const DEFAULT_CHAT_SEARCH_Q_PARAM_PROMPT: &str = "The search query string used to find relevant documents in the index. This should contain keywords or phrases that best represent what the user is looking for. More specific queries will yield more precise results.";
+pub const DEFAULT_CHAT_SEARCH_FILTER_PARAM_PROMPT: &str = "The search filter string used to find relevant documents in the index. It supports parentheses, `=`, `!=`, `>=`, `>`, `<=`, `<`, `IN`, `NOT IN`, `TO`, `EXISTS`, `NOT EXISTS`, `IS NULL`, `IS NOT NULL`, `IS EMPTY`, `IS NOT EMPTY`, `_geoRadius`, or `_geoBoundingBox`. Here is an example: \"price > 100 AND category = 'electronics'\". The following is a list of fields that can be filtered on: ";
 pub const DEFAULT_CHAT_SEARCH_INDEX_UID_PARAM_PROMPT: &str = "The name of the index to search within. An index is a collection of documents organized for search. Selecting the right index ensures the most relevant results for the user query.";
 
 #[derive(Serialize, Deserialize, Debug, Clone, Copy, Default, PartialEq, Eq)]
@@ -161,18 +162,31 @@ impl ChatCompletionSource {
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Eq)]
 #[serde(rename_all = "camelCase")]
 pub struct ChatCompletionPrompts {
+    #[serde(default)]
     pub system: String,
+    #[serde(default)]
     pub search_description: String,
+    #[serde(default)]
     pub search_q_param: String,
+    #[serde(default = "default_search_filter_param")]
+    pub search_filter_param: String,
+    #[serde(default)]
     pub search_index_uid_param: String,
 }
 
+/// This function is used for when the search_filter_param is
+/// not provided and this can happen when the database is in v1.15.
+fn default_search_filter_param() -> String {
+    DEFAULT_CHAT_SEARCH_FILTER_PARAM_PROMPT.to_string()
+}
+
 impl Default for ChatCompletionPrompts {
     fn default() -> Self {
         Self {
             system: DEFAULT_CHAT_SYSTEM_PROMPT.to_string(),
             search_description: DEFAULT_CHAT_SEARCH_DESCRIPTION_PROMPT.to_string(),
             search_q_param: DEFAULT_CHAT_SEARCH_Q_PARAM_PROMPT.to_string(),
+            search_filter_param: DEFAULT_CHAT_SEARCH_FILTER_PARAM_PROMPT.to_string(),
             search_index_uid_param: DEFAULT_CHAT_SEARCH_INDEX_UID_PARAM_PROMPT.to_string(),
         }
     }

crates/meilisearch/src/analytics/mock_analytics.rs

@@ -104,6 +104,4 @@ impl Analytics for MockAnalytics {
         _request: &HttpRequest,
     ) {
     }
-    fn get_fetch_documents(&self, _documents_query: &DocumentFetchKind, _request: &HttpRequest) {}
-    fn post_fetch_documents(&self, _documents_query: &DocumentFetchKind, _request: &HttpRequest) {}
 }

crates/meilisearch/src/analytics/mod.rs

@@ -73,12 +73,6 @@ pub enum DocumentDeletionKind {
     PerFilter,
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq)]
-pub enum DocumentFetchKind {
-    PerDocumentId { retrieve_vectors: bool },
-    Normal { with_filter: bool, limit: usize, offset: usize, retrieve_vectors: bool },
-}
-
 /// To send an event to segment, your event must be able to aggregate itself with another event of the same type.
 pub trait Aggregate: 'static + mopa::Any + Send {
     /// The name of the event that will be sent to segment.

crates/meilisearch/src/error.rs

@@ -49,7 +49,7 @@ pub enum MeilisearchHttpError {
     TooManySearchRequests(usize),
     #[error("Internal error: Search limiter is down.")]
     SearchLimiterIsDown,
-    #[error("The provided payload reached the size limit. The maximum accepted payload size is {}.",  Byte::from_u64(*.0 as u64).get_appropriate_unit(UnitType::Binary))]
+    #[error("The provided payload reached the size limit. The maximum accepted payload size is {}.", Byte::from_u64(*.0 as u64).get_appropriate_unit(if *.0 % 1024 == 0 { UnitType::Binary } else { UnitType::Decimal }))]
     PayloadTooLarge(usize),
     #[error("Two indexes must be given for each swap. The list `[{}]` contains {} indexes.",
         .0.iter().map(|uid| format!("\"{uid}\"")).collect::<Vec<_>>().join(", "), .0.len()

crates/meilisearch/src/routes/chats/chat_completions.rs

@@ -27,9 +27,10 @@ use meilisearch_types::features::{
     ChatCompletionPrompts as DbChatCompletionPrompts,
     ChatCompletionSource as DbChatCompletionSource, SystemRole,
 };
+use meilisearch_types::heed::RoTxn;
 use meilisearch_types::keys::actions;
 use meilisearch_types::milli::index::ChatConfig;
-use meilisearch_types::milli::{all_obkv_to_json, obkv_to_json, TimeBudget};
+use meilisearch_types::milli::{all_obkv_to_json, obkv_to_json, OrderBy, PatternMatch, TimeBudget};
 use meilisearch_types::{Document, Index};
 use serde::Deserialize;
 use serde_json::json;
@@ -169,6 +170,7 @@ fn setup_search_tool(
 
     let mut index_uids = Vec::new();
     let mut function_description = prompts.search_description.clone();
+    let mut filter_description = prompts.search_filter_param.clone();
     index_scheduler.try_for_each_index::<_, ()>(|name, index| {
         // Make sure to skip unauthorized indexes
         if !filters.is_index_authorized(name) {
@@ -180,16 +182,22 @@ fn setup_search_tool(
         let index_description = chat_config.description;
         let _ = writeln!(&mut function_description, "\n\n - {name}: {index_description}\n");
         index_uids.push(name.to_string());
+        let facet_distributions = format_facet_distributions(index, &rtxn, 10).unwrap(); // TODO do not unwrap
+        let _ = writeln!(&mut filter_description, "\n## Facet distributions of the {name} index");
+        let _ = writeln!(&mut filter_description, "{facet_distributions}");
 
         Ok(())
     })?;
 
+    tracing::debug!("LLM function description: {function_description}");
+    tracing::debug!("LLM filter description: {filter_description}");
+
     let tool = ChatCompletionToolArgs::default()
         .r#type(ChatCompletionToolType::Function)
         .function(
             FunctionObjectArgs::default()
                 .name(MEILI_SEARCH_IN_INDEX_FUNCTION_NAME)
-                .description(&function_description)
+                .description(function_description)
                 .parameters(json!({
                     "type": "object",
                     "properties": {
@@ -203,9 +211,13 @@ fn setup_search_tool(
                             // "type": ["string", "null"],
                             "type": "string",
                             "description": prompts.search_q_param,
+                        },
+                        "filter": {
+                            "type": "string",
+                            "description": filter_description,
                         }
                     },
-                    "required": ["index_uid", "q"],
+                    "required": ["index_uid", "q", "filter"],
                     "additionalProperties": false,
                 }))
                 .strict(true)
@@ -247,11 +259,19 @@ async fn process_search_request(
     auth_token: &str,
     index_uid: String,
     q: Option<String>,
+    filter: Option<String>,
 ) -> Result<(Index, Vec<Document>, String), ResponseError> {
     let index = index_scheduler.index(&index_uid)?;
     let rtxn = index.static_read_txn()?;
     let ChatConfig { description: _, prompt: _, search_parameters } = index.chat_config(&rtxn)?;
-    let mut query = SearchQuery { q, ..SearchQuery::from(search_parameters) };
+    let mut query = SearchQuery {
+        q,
+        filter: filter.map(serde_json::Value::from),
+        ..SearchQuery::from(search_parameters)
+    };
+
+    tracing::debug!("LLM query: {:?}", query);
+
     let auth_filter = ActionPolicy::<{ actions::SEARCH }>::authenticate(
         auth_ctrl,
         auth_token,
@@ -280,14 +300,23 @@ async fn process_search_request(
         let (search, _is_finite_pagination, _max_total_hits, _offset) =
             prepare_search(&index_cloned, &rtxn, &query, &search_kind, time_budget, features)?;
 
-        search_from_kind(index_uid, search_kind, search)
-            .map(|(search_results, _)| (rtxn, search_results))
-            .map_err(ResponseError::from)
+        match search_from_kind(index_uid, search_kind, search) {
+            Ok((search_results, _)) => Ok((rtxn, Ok(search_results))),
+            Err(MeilisearchHttpError::Milli {
+                error: meilisearch_types::milli::Error::UserError(user_error),
+                index_name: _,
+            }) => Ok((rtxn, Err(user_error))),
+            Err(err) => Err(ResponseError::from(err)),
+        }
     })
     .await;
     permit.drop().await;
 
-    let output = output?;
+    let output = match output? {
+        Ok((rtxn, Ok(search_results))) => Ok((rtxn, search_results)),
+        Ok((_rtxn, Err(error))) => return Ok((index, Vec::new(), error.to_string())),
+        Err(err) => Err(err),
+    };
     let mut documents = Vec::new();
     if let Ok((ref rtxn, ref search_result)) = output {
         MEILISEARCH_CHAT_SEARCH_REQUESTS.with_label_values(&["internal"]).inc();
@@ -395,16 +424,19 @@ async fn non_streamed_chat(
 
                 for call in meili_calls {
                     let result = match serde_json::from_str(&call.function.arguments) {
-                        Ok(SearchInIndexParameters { index_uid, q }) => process_search_request(
-                            &index_scheduler,
-                            auth_ctrl.clone(),
-                            &search_queue,
-                            auth_token,
-                            index_uid,
-                            q,
-                        )
-                        .await
-                        .map_err(|e| e.to_string()),
+                        Ok(SearchInIndexParameters { index_uid, q, filter }) => {
+                            process_search_request(
+                                &index_scheduler,
+                                auth_ctrl.clone(),
+                                &search_queue,
+                                auth_token,
+                                index_uid,
+                                q,
+                                filter,
+                            )
+                            .await
+                            .map_err(|e| e.to_string())
+                        }
                         Err(err) => Err(err.to_string()),
                     };
 
@@ -719,13 +751,14 @@ async fn handle_meili_tools(
         let mut error = None;
 
         let result = match serde_json::from_str(&call.function.arguments) {
-            Ok(SearchInIndexParameters { index_uid, q }) => match process_search_request(
+            Ok(SearchInIndexParameters { index_uid, q, filter }) => match process_search_request(
                 index_scheduler,
                 auth_ctrl.clone(),
                 search_queue,
                 auth_token,
                 index_uid,
                 q,
+                filter,
             )
             .await
             {
@@ -801,4 +834,42 @@ struct SearchInIndexParameters {
     index_uid: String,
     /// The query parameter to use.
     q: Option<String>,
+    /// The filter parameter to use.
+    filter: Option<String>,
+}
+
+fn format_facet_distributions(
+    index: &Index,
+    rtxn: &RoTxn,
+    max_values_per_facet: usize,
+) -> meilisearch_types::milli::Result<String> {
+    let universe = index.documents_ids(rtxn)?;
+    let rules = index.filterable_attributes_rules(rtxn)?;
+    let fields_ids_map = index.fields_ids_map(rtxn)?;
+    let filterable_attributes = fields_ids_map
+        .names()
+        .filter(|name| rules.iter().any(|rule| matches!(rule.match_str(name), PatternMatch::Match)))
+        .map(|name| (name, OrderBy::Count));
+    let facets_distribution = index
+        .facets_distribution(rtxn)
+        .max_values_per_facet(max_values_per_facet)
+        .candidates(universe)
+        .facets(filterable_attributes)
+        .execute()?;
+
+    let mut output = String::new();
+    for (facet_name, entries) in facets_distribution {
+        let _ = write!(&mut output, "{}: ", facet_name);
+        let total_entries = entries.len();
+        for (i, (value, _count)) in entries.into_iter().enumerate() {
+            let _ = if total_entries.saturating_sub(1) == i {
+                write!(&mut output, "{value}.")
+            } else {
+                write!(&mut output, "{value}, ")
+            };
+        }
+        let _ = writeln!(&mut output);
+    }
+
+    Ok(output)
 }

crates/meilisearch/src/routes/chats/settings.rs

@@ -8,8 +8,8 @@ use meilisearch_types::error::{Code, ResponseError};
 use meilisearch_types::features::{
     ChatCompletionPrompts as DbChatCompletionPrompts, ChatCompletionSettings,
     ChatCompletionSource as DbChatCompletionSource, DEFAULT_CHAT_SEARCH_DESCRIPTION_PROMPT,
-    DEFAULT_CHAT_SEARCH_INDEX_UID_PARAM_PROMPT, DEFAULT_CHAT_SEARCH_Q_PARAM_PROMPT,
-    DEFAULT_CHAT_SYSTEM_PROMPT,
+    DEFAULT_CHAT_SEARCH_FILTER_PARAM_PROMPT, DEFAULT_CHAT_SEARCH_INDEX_UID_PARAM_PROMPT,
+    DEFAULT_CHAT_SEARCH_Q_PARAM_PROMPT, DEFAULT_CHAT_SYSTEM_PROMPT,
 };
 use meilisearch_types::keys::actions;
 use meilisearch_types::milli::update::Setting;
@@ -84,6 +84,11 @@ async fn patch_settings(
                 Setting::Reset => DEFAULT_CHAT_SEARCH_Q_PARAM_PROMPT.to_string(),
                 Setting::NotSet => old_settings.prompts.search_q_param,
             },
+            search_filter_param: match new_prompts.search_filter_param {
+                Setting::Set(new_description) => new_description,
+                Setting::Reset => DEFAULT_CHAT_SEARCH_FILTER_PARAM_PROMPT.to_string(),
+                Setting::NotSet => old_settings.prompts.search_filter_param,
+            },
             search_index_uid_param: match new_prompts.search_index_uid_param {
                 Setting::Set(new_description) => new_description,
                 Setting::Reset => DEFAULT_CHAT_SEARCH_INDEX_UID_PARAM_PROMPT.to_string(),
@@ -252,6 +257,10 @@ pub struct ChatPrompts {
     #[schema(value_type = Option<String>, example = json!("This is query parameter..."))]
     pub search_q_param: Setting<String>,
     #[serde(default)]
+    #[deserr(default, error = DeserrJsonError<InvalidChatCompletionSearchFilterParamPrompt>)]
+    #[schema(value_type = Option<String>, example = json!("This is filter parameter..."))]
+    pub search_filter_param: Setting<String>,
+    #[serde(default)]
     #[deserr(default, error = DeserrJsonError<InvalidChatCompletionSearchIndexUidParamPrompt>)]
     #[schema(value_type = Option<String>, example = json!("This is index you want to search in..."))]
     pub search_index_uid_param: Setting<String>,

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

@@ -1,6 +1,7 @@
 use std::collections::HashSet;
 use std::io::{ErrorKind, Seek as _};
 use std::marker::PhantomData;
+use std::str::FromStr;
 
 use actix_web::http::header::CONTENT_TYPE;
 use actix_web::web::Data;
@@ -17,9 +18,10 @@ use meilisearch_types::error::deserr_codes::*;
 use meilisearch_types::error::{Code, ResponseError};
 use meilisearch_types::heed::RoTxn;
 use meilisearch_types::index_uid::IndexUid;
+use meilisearch_types::milli::documents::sort::recursive_sort;
 use meilisearch_types::milli::update::IndexDocumentsMethod;
 use meilisearch_types::milli::vector::parsed_vectors::ExplicitVectors;
-use meilisearch_types::milli::DocumentId;
+use meilisearch_types::milli::{AscDesc, DocumentId};
 use meilisearch_types::serde_cs::vec::CS;
 use meilisearch_types::star_or::OptionStarOrList;
 use meilisearch_types::tasks::KindWithContent;
@@ -42,6 +44,7 @@ use crate::extractors::authentication::policies::*;
 use crate::extractors::authentication::GuardedData;
 use crate::extractors::payload::Payload;
 use crate::extractors::sequential_extractor::SeqHandler;
+use crate::routes::indexes::search::fix_sort_query_parameters;
 use crate::routes::{
     get_task_id, is_dry_run, PaginationView, SummarizedTaskView, PAGINATION_DEFAULT_LIMIT,
 };
@@ -135,6 +138,10 @@ pub struct DocumentsFetchAggregator<Method: AggregateMethod> {
     per_document_id: bool,
     // if a filter was used
     per_filter: bool,
+    with_vector_filter: bool,
+
+    // if documents were sorted
+    sort: bool,
 
     #[serde(rename = "vector.retrieve_vectors")]
     retrieve_vectors: bool,
@@ -151,39 +158,6 @@ pub struct DocumentsFetchAggregator<Method: AggregateMethod> {
     marker: std::marker::PhantomData<Method>,
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq)]
-pub enum DocumentFetchKind {
-    PerDocumentId { retrieve_vectors: bool },
-    Normal { with_filter: bool, limit: usize, offset: usize, retrieve_vectors: bool, ids: usize },
-}
-
-impl<Method: AggregateMethod> DocumentsFetchAggregator<Method> {
-    pub fn from_query(query: &DocumentFetchKind) -> Self {
-        let (limit, offset, retrieve_vectors) = match query {
-            DocumentFetchKind::PerDocumentId { retrieve_vectors } => (1, 0, *retrieve_vectors),
-            DocumentFetchKind::Normal { limit, offset, retrieve_vectors, .. } => {
-                (*limit, *offset, *retrieve_vectors)
-            }
-        };
-
-        let ids = match query {
-            DocumentFetchKind::Normal { ids, .. } => *ids,
-            DocumentFetchKind::PerDocumentId { .. } => 0,
-        };
-
-        Self {
-            per_document_id: matches!(query, DocumentFetchKind::PerDocumentId { .. }),
-            per_filter: matches!(query, DocumentFetchKind::Normal { with_filter, .. } if *with_filter),
-            max_limit: limit,
-            max_offset: offset,
-            retrieve_vectors,
-            max_document_ids: ids,
-
-            marker: PhantomData,
-        }
-    }
-}
-
 impl<Method: AggregateMethod> Aggregate for DocumentsFetchAggregator<Method> {
     fn event_name(&self) -> &'static str {
         Method::event_name()
@@ -193,6 +167,8 @@ impl<Method: AggregateMethod> Aggregate for DocumentsFetchAggregator<Method> {
         Box::new(Self {
             per_document_id: self.per_document_id | new.per_document_id,
             per_filter: self.per_filter | new.per_filter,
+            with_vector_filter: self.with_vector_filter | new.with_vector_filter,
+            sort: self.sort | new.sort,
             retrieve_vectors: self.retrieve_vectors | new.retrieve_vectors,
             max_limit: self.max_limit.max(new.max_limit),
             max_offset: self.max_offset.max(new.max_offset),
@@ -276,6 +252,8 @@ pub async fn get_document(
             retrieve_vectors: param_retrieve_vectors.0,
             per_document_id: true,
             per_filter: false,
+            with_vector_filter: false,
+            sort: false,
             max_limit: 0,
             max_offset: 0,
             max_document_ids: 0,
@@ -406,6 +384,8 @@ pub struct BrowseQueryGet {
     #[param(default, value_type = Option<String>, example = "popularity > 1000")]
     #[deserr(default, error = DeserrQueryParamError<InvalidDocumentFilter>)]
     filter: Option<String>,
+    #[deserr(default, error = DeserrQueryParamError<InvalidDocumentSort>)]
+    sort: Option<String>,
 }
 
 #[derive(Debug, Deserr, ToSchema)]
@@ -430,6 +410,9 @@ pub struct BrowseQuery {
     #[schema(default, value_type = Option<Value>, example = "popularity > 1000")]
     #[deserr(default, error = DeserrJsonError<InvalidDocumentFilter>)]
     filter: Option<Value>,
+    #[schema(default, value_type = Option<Vec<String>>, example = json!(["title:asc", "rating:desc"]))]
+    #[deserr(default, error = DeserrJsonError<InvalidDocumentSort>)]
+    sort: Option<Vec<String>>,
 }
 
 /// Get documents with POST
@@ -495,6 +478,11 @@ pub async fn documents_by_query_post(
     analytics.publish(
         DocumentsFetchAggregator::<DocumentsPOST> {
             per_filter: body.filter.is_some(),
+            with_vector_filter: body
+                .filter
+                .as_ref()
+                .is_some_and(|f| f.to_string().contains("_vectors")),
+            sort: body.sort.is_some(),
             retrieve_vectors: body.retrieve_vectors,
             max_limit: body.limit,
             max_offset: body.offset,
@@ -571,7 +559,7 @@ pub async fn get_documents(
 ) -> Result<HttpResponse, ResponseError> {
     debug!(parameters = ?params, "Get documents GET");
 
-    let BrowseQueryGet { limit, offset, fields, retrieve_vectors, filter, ids } =
+    let BrowseQueryGet { limit, offset, fields, retrieve_vectors, filter, ids, sort } =
         params.into_inner();
 
     let filter = match filter {
@@ -582,20 +570,24 @@ pub async fn get_documents(
         None => None,
     };
 
-    let ids = ids.map(|ids| ids.into_iter().map(Into::into).collect());
-
     let query = BrowseQuery {
         offset: offset.0,
         limit: limit.0,
         fields: fields.merge_star_and_none(),
         retrieve_vectors: retrieve_vectors.0,
         filter,
-        ids,
+        ids: ids.map(|ids| ids.into_iter().map(Into::into).collect()),
+        sort: sort.map(|attr| fix_sort_query_parameters(&attr)),
     };
 
     analytics.publish(
         DocumentsFetchAggregator::<DocumentsGET> {
             per_filter: query.filter.is_some(),
+            with_vector_filter: query
+                .filter
+                .as_ref()
+                .is_some_and(|f| f.to_string().contains("_vectors")),
+            sort: query.sort.is_some(),
             retrieve_vectors: query.retrieve_vectors,
             max_limit: query.limit,
             max_offset: query.offset,
@@ -615,7 +607,7 @@ fn documents_by_query(
     query: BrowseQuery,
 ) -> Result<HttpResponse, ResponseError> {
     let index_uid = IndexUid::try_from(index_uid.into_inner())?;
-    let BrowseQuery { offset, limit, fields, retrieve_vectors, filter, ids } = query;
+    let BrowseQuery { offset, limit, fields, retrieve_vectors, filter, ids, sort } = query;
 
     let retrieve_vectors = RetrieveVectors::new(retrieve_vectors);
 
@@ -633,6 +625,18 @@ fn documents_by_query(
         None
     };
 
+    let sort_criteria = if let Some(sort) = &sort {
+        let sorts: Vec<_> = match sort.iter().map(|s| milli::AscDesc::from_str(s)).collect() {
+            Ok(sorts) => sorts,
+            Err(asc_desc_error) => {
+                return Err(milli::SortError::from(asc_desc_error).into_document_error().into())
+            }
+        };
+        Some(sorts)
+    } else {
+        None
+    };
+
     let index = index_scheduler.index(&index_uid)?;
     let (total, documents) = retrieve_documents(
         &index,
@@ -643,6 +647,7 @@ fn documents_by_query(
         fields,
         retrieve_vectors,
         index_scheduler.features(),
+        sort_criteria,
     )?;
 
     let ret = PaginationView::new(offset, limit, total as usize, documents);
@@ -1461,8 +1466,6 @@ fn some_documents<'a, 't: 'a>(
                     document.remove("_vectors");
                 }
                 RetrieveVectors::Retrieve => {
-                    // Clippy is simply wrong
-                    #[allow(clippy::manual_unwrap_or_default)]
                     let mut vectors = match document.remove("_vectors") {
                         Some(Value::Object(map)) => map,
                         _ => Default::default(),
@@ -1494,6 +1497,7 @@ fn retrieve_documents<S: AsRef<str>>(
     attributes_to_retrieve: Option<Vec<S>>,
     retrieve_vectors: RetrieveVectors,
     features: RoFeatures,
+    sort_criteria: Option<Vec<AscDesc>>,
 ) -> Result<(u64, Vec<Document>), ResponseError> {
     let rtxn = index.read_txn()?;
     let filter = &filter;
@@ -1526,15 +1530,32 @@ fn retrieve_documents<S: AsRef<str>>(
         })?
     }
 
-    let (it, number_of_documents) = {
+    let (it, number_of_documents) = if let Some(sort) = sort_criteria {
+        let number_of_documents = candidates.len();
+        let facet_sort = recursive_sort(index, &rtxn, sort, &candidates)?;
+        let iter = facet_sort.iter()?;
+        let mut documents = Vec::with_capacity(limit);
+        for result in iter.skip(offset).take(limit) {
+            documents.push(result?);
+        }
+        (
+            itertools::Either::Left(some_documents(
+                index,
+                &rtxn,
+                documents.into_iter(),
+                retrieve_vectors,
+            )?),
+            number_of_documents,
+        )
+    } else {
         let number_of_documents = candidates.len();
         (
-            some_documents(
+            itertools::Either::Right(some_documents(
                 index,
                 &rtxn,
                 candidates.into_iter().skip(offset).take(limit),
                 retrieve_vectors,
-            )?,
+            )?),
             number_of_documents,
         )
     };

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

@@ -40,6 +40,7 @@ pub struct SearchAggregator<Method: AggregateMethod> {
     // filter
     filter_with_geo_radius: bool,
     filter_with_geo_bounding_box: bool,
+    filter_on_vectors: bool,
     // every time a request has a filter, this field must be incremented by the number of terms it contains
     filter_sum_of_criteria_terms: usize,
     // every time a request has a filter, this field must be incremented by one
@@ -163,6 +164,7 @@ impl<Method: AggregateMethod> SearchAggregator<Method> {
             let stringified_filters = filter.to_string();
             ret.filter_with_geo_radius = stringified_filters.contains("_geoRadius(");
             ret.filter_with_geo_bounding_box = stringified_filters.contains("_geoBoundingBox(");
+            ret.filter_on_vectors = stringified_filters.contains("_vectors");
             ret.filter_sum_of_criteria_terms = RE.split(&stringified_filters).count();
         }
 
@@ -260,6 +262,7 @@ impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
             distinct,
             filter_with_geo_radius,
             filter_with_geo_bounding_box,
+            filter_on_vectors,
             filter_sum_of_criteria_terms,
             filter_total_number_of_criteria,
             used_syntax,
@@ -314,6 +317,7 @@ impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
         // filter
         self.filter_with_geo_radius |= filter_with_geo_radius;
         self.filter_with_geo_bounding_box |= filter_with_geo_bounding_box;
+        self.filter_on_vectors |= filter_on_vectors;
         self.filter_sum_of_criteria_terms =
             self.filter_sum_of_criteria_terms.saturating_add(filter_sum_of_criteria_terms);
         self.filter_total_number_of_criteria =
@@ -388,6 +392,7 @@ impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
             distinct,
             filter_with_geo_radius,
             filter_with_geo_bounding_box,
+            filter_on_vectors,
             filter_sum_of_criteria_terms,
             filter_total_number_of_criteria,
             used_syntax,
@@ -445,6 +450,7 @@ impl<Method: AggregateMethod> Aggregate for SearchAggregator<Method> {
             "filter": {
                "with_geoRadius": filter_with_geo_radius,
                "with_geoBoundingBox": filter_with_geo_bounding_box,
+               "on_vectors": filter_on_vectors,
                "avg_criteria_number": format!("{:.2}", filter_sum_of_criteria_terms as f64 / filter_total_number_of_criteria as f64),
                "most_used_syntax": used_syntax.iter().max_by_key(|(_, v)| *v).map(|(k, _)| json!(k)).unwrap_or_else(|| json!(null)),
             },

crates/meilisearch/src/routes/tasks.rs

@@ -336,7 +336,7 @@ impl<Method: AggregateMethod + 'static> Aggregate for TaskFilterAnalytics<Method
                 "link": "https://docs.meilisearch.com/errors#missing_authorization_header"
             }
         )),
-        (status = 404, description = "The task uid does not exists", body = ResponseError, content_type = "application/json", example = json!(
+        (status = 404, description = "The task uid does not exist", body = ResponseError, content_type = "application/json", example = json!(
             {
                 "message": "Task :taskUid not found.",
                 "code": "task_not_found",
@@ -430,7 +430,7 @@ async fn cancel_tasks(
                 "link": "https://docs.meilisearch.com/errors#missing_authorization_header"
             }
         )),
-        (status = 404, description = "The task uid does not exists", body = ResponseError, content_type = "application/json", example = json!(
+        (status = 404, description = "The task uid does not exist", body = ResponseError, content_type = "application/json", example = json!(
             {
                 "message": "Task :taskUid not found.",
                 "code": "task_not_found",
@@ -611,7 +611,7 @@ async fn get_tasks(
                 "link": "https://docs.meilisearch.com/errors#missing_authorization_header"
             }
         )),
-        (status = 404, description = "The task uid does not exists", body = ResponseError, content_type = "application/json", example = json!(
+        (status = 404, description = "The task uid does not exist", body = ResponseError, content_type = "application/json", example = json!(
             {
                 "message": "Task :taskUid not found.",
                 "code": "task_not_found",
@@ -665,7 +665,7 @@ async fn get_task(
                 "link": "https://docs.meilisearch.com/errors#missing_authorization_header"
             }
         )),
-        (status = 404, description = "The task uid does not exists", body = ResponseError, content_type = "application/json", example = json!(
+        (status = 404, description = "The task uid does not exist", body = ResponseError, content_type = "application/json", example = json!(
             {
                 "message": "Task :taskUid not found.",
                 "code": "task_not_found",

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

@@ -745,10 +745,9 @@ impl SearchByIndex {
                         match sort.iter().map(|s| milli::AscDesc::from_str(s)).collect() {
                             Ok(sorts) => sorts,
                             Err(asc_desc_error) => {
-                                return Err(milli::Error::from(milli::SortError::from(
-                                    asc_desc_error,
-                                ))
-                                .into())
+                                return Err(milli::SortError::from(asc_desc_error)
+                                    .into_search_error()
+                                    .into())
                             }
                         };
                     Some(sorts)

crates/meilisearch/src/search/mod.rs

@@ -1051,6 +1051,7 @@ pub fn prepare_search<'t>(
         .unwrap_or(DEFAULT_PAGINATION_MAX_TOTAL_HITS);
 
     search.exhaustive_number_hits(is_finite_pagination);
+    search.max_total_hits(Some(max_total_hits));
     search.scoring_strategy(
         if query.show_ranking_score
             || query.show_ranking_score_details
@@ -1091,7 +1092,7 @@ pub fn prepare_search<'t>(
         let sort = match sort.iter().map(|s| AscDesc::from_str(s)).collect() {
             Ok(sorts) => sorts,
             Err(asc_desc_error) => {
-                return Err(milli::Error::from(SortError::from(asc_desc_error)).into())
+                return Err(SortError::from(asc_desc_error).into_search_error().into())
             }
         };
 
@@ -2077,7 +2078,7 @@ pub(crate) fn parse_filter(
     })?;
 
     if let Some(ref filter) = filter {
-        // If the contains operator is used while the contains filter features is not enabled, errors out
+        // If the contains operator is used while the contains filter feature is not enabled, errors out
         if let Some((token, error)) =
             filter.use_contains_operator().zip(features.check_contains_filter().err())
         {
@@ -2088,6 +2089,18 @@ pub(crate) fn parse_filter(
         }
     }
 
+    if let Some(ref filter) = filter {
+        // If a vector filter is used while the multi modal feature is not enabled, errors out
+        if let Some((token, error)) =
+            filter.use_vector_filter().zip(features.check_multimodal("using a vector filter").err())
+        {
+            return Err(ResponseError::from_msg(
+                token.as_external_error(error).to_string(),
+                Code::FeatureNotEnabled,
+            ));
+        }
+    }
+
     Ok(filter)
 }
 

crates/meilisearch/tests/common/index.rs

@@ -562,5 +562,7 @@ pub struct GetAllDocumentsOptions {
     pub offset: Option<usize>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub fields: Option<Vec<&'static str>>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub sort: Option<Vec<&'static str>>,
     pub retrieve_vectors: bool,
 }

crates/meilisearch/tests/common/mod.rs

@@ -3,8 +3,12 @@ pub mod index;
 pub mod server;
 pub mod service;
 
-use std::fmt::{self, Display};
+use std::{
+    collections::BTreeMap,
+    fmt::{self, Display},
+};
 
+use actix_http::StatusCode;
 #[allow(unused)]
 pub use index::GetAllDocumentsOptions;
 use meili_snap::json_string;
@@ -13,6 +17,10 @@ use serde::{Deserialize, Serialize};
 #[allow(unused)]
 pub use server::{default_settings, Server};
 use tokio::sync::OnceCell;
+use wiremock::{
+    matchers::{method, path},
+    Mock, MockServer, Request, ResponseTemplate,
+};
 
 use crate::common::index::Index;
 
@@ -508,3 +516,166 @@ pub async fn shared_index_with_geo_documents() -> &'static Index<'static, Shared
         })
         .await
 }
+
+pub async fn shared_index_for_fragments() -> Index<'static, Shared> {
+    static INDEX: OnceCell<(Server<Shared>, String)> = OnceCell::const_new();
+    let (server, uid) = INDEX
+        .get_or_init(|| async {
+            let (server, uid, _) = init_fragments_index().await;
+            (server.into_shared(), uid)
+        })
+        .await;
+    server._index(uid).to_shared()
+}
+
+async fn fragment_mock_server() -> String {
+    let text_to_embedding: BTreeMap<_, _> = vec![
+        ("kefir", [0.5, -0.5, 0.0]),
+        ("intel", [1.0, 1.0, 0.0]),
+        ("dustin", [-0.5, 0.5, 0.0]),
+        ("bulldog", [0.0, 0.0, 1.0]),
+        ("labrador", [0.0, 0.0, -1.0]),
+        ("{{ doc.", [-9999.0, -9999.0, -9999.0]), // If a template didn't render
+    ]
+    .into_iter()
+    .collect();
+
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
+
+    Mock::given(method("POST"))
+        .and(path("/"))
+        .respond_with(move |req: &Request| {
+            let text = String::from_utf8_lossy(&req.body).to_string();
+
+            let mut data = [0.0, 0.0, 0.0];
+            for (inner_text, inner_data) in &text_to_embedding {
+                if text.contains(inner_text) {
+                    for (i, &value) in inner_data.iter().enumerate() {
+                        data[i] += value;
+                    }
+                }
+            }
+            ResponseTemplate::new(200).set_body_json(json!({ "data": data }))
+        })
+        .mount(mock_server)
+        .await;
+
+    mock_server.uri()
+}
+
+pub async fn init_fragments_index() -> (Server<Owned>, String, crate::common::Value) {
+    let url = fragment_mock_server().await;
+    let server = Server::new().await;
+    let index = server.unique_index();
+
+    let (_response, code) = server.set_features(json!({"multimodal": true})).await;
+    assert_eq!(code, StatusCode::OK);
+
+    // Configure the index to use our mock embedder
+    let settings = json!({
+        "embedders": {
+            "rest": {
+                "source": "rest",
+                "url": url,
+                "dimensions": 3,
+                "request": "{{fragment}}",
+                "response": {
+                    "data": "{{embedding}}"
+                },
+                "indexingFragments": {
+                    "withBreed": {"value": "{{ doc.name }} is a {{ doc.breed }}"},
+                    "basic": {"value": "{{ doc.name }} is a dog"},
+                },
+                "searchFragments": {
+                    "justBreed": {"value": "It's a {{ media.breed }}"},
+                    "justName": {"value": "{{ media.name }} is a dog"},
+                    "query": {"value": "Some pre-prompt for query {{ q }}"},
+                }
+            },
+        },
+    });
+    let (response, code) = index.update_settings(settings.clone()).await;
+    assert_eq!(code, StatusCode::ACCEPTED);
+
+    server.wait_task(response.uid()).await.succeeded();
+
+    // Send documents
+    let documents = json!([
+        {"id": 0, "name": "kefir"},
+        {"id": 1, "name": "echo", "_vectors": { "rest": [1, 1, 1] }},
+        {"id": 2, "name": "intel", "breed": "labrador"},
+        {"id": 3, "name": "dustin", "breed": "bulldog"},
+    ]);
+    let (value, code) = index.add_documents(documents, None).await;
+    assert_eq!(code, StatusCode::ACCEPTED);
+
+    let _task = index.wait_task(value.uid()).await.succeeded();
+
+    let uid = index.uid.clone();
+    (server, uid, settings)
+}
+
+pub async fn init_fragments_index_composite() -> (Server<Owned>, String, crate::common::Value) {
+    let url = fragment_mock_server().await;
+    let server = Server::new().await;
+    let index = server.unique_index();
+
+    let (_response, code) = server.set_features(json!({"multimodal": true})).await;
+    assert_eq!(code, StatusCode::OK);
+
+    let (_response, code) = server.set_features(json!({"compositeEmbedders": true})).await;
+    assert_eq!(code, StatusCode::OK);
+
+    // Configure the index to use our mock embedder
+    let settings = json!({
+        "embedders": {
+            "rest": {
+                "source": "composite",
+                "searchEmbedder": {
+                    "source": "rest",
+                    "url": url,
+                    "dimensions": 3,
+                    "request": "{{fragment}}",
+                    "response": {
+                        "data": "{{embedding}}"
+                    },
+                    "searchFragments": {
+                        "query": {"value": "Some pre-prompt for query {{ q }}"},
+                    }
+                },
+                "indexingEmbedder": {
+                    "source": "rest",
+                    "url": url,
+                    "dimensions": 3,
+                    "request": "{{fragment}}",
+                    "response": {
+                        "data": "{{embedding}}"
+                    },
+                    "indexingFragments": {
+                        "withBreed": {"value": "{{ doc.name }} is a {{ doc.breed }}"},
+                        "basic": {"value": "{{ doc.name }} is a dog"},
+                    }
+                },
+            },
+        },
+    });
+    let (response, code) = index.update_settings(settings.clone()).await;
+    assert_eq!(code, StatusCode::ACCEPTED);
+
+    server.wait_task(response.uid()).await.succeeded();
+
+    // Send documents
+    let documents = json!([
+        {"id": 0, "name": "kefir"},
+        {"id": 1, "name": "echo", "_vectors": { "rest": [1, 1, 1] }},
+        {"id": 2, "name": "intel", "breed": "labrador"},
+        {"id": 3, "name": "dustin", "breed": "bulldog"},
+    ]);
+    let (value, code) = index.add_documents(documents, None).await;
+    assert_eq!(code, StatusCode::ACCEPTED);
+
+    index.wait_task(value.uid()).await.succeeded();
+
+    let uid = index.uid.clone();
+    (server, uid, settings)
+}

crates/meilisearch/tests/common/server.rs

@@ -35,7 +35,7 @@ pub struct Server<State = Owned> {
 pub static TEST_TEMP_DIR: Lazy<TempDir> = Lazy::new(|| TempDir::new().unwrap());
 
 impl Server<Owned> {
-    fn into_shared(self) -> Server<Shared> {
+    pub(super) fn into_shared(self) -> Server<Shared> {
         Server { service: self.service, _dir: self._dir, _marker: PhantomData }
     }
 

crates/meilisearch/tests/documents/errors.rs

@@ -557,7 +557,7 @@ async fn delete_document_by_filter() {
     "###);
 
     let index = shared_does_not_exists_index().await;
-    // index does not exists
+    // index does not exist
     let (response, _code) =
         index.delete_document_by_filter_fail(json!({ "filter": "doggo = bernese"})).await;
     snapshot!(response, @r###"

crates/meilisearch/tests/documents/get_documents.rs

@@ -5,8 +5,8 @@ use urlencoding::encode as urlencode;
 
 use crate::common::encoder::Encoder;
 use crate::common::{
-    shared_does_not_exists_index, shared_empty_index, shared_index_with_test_set,
-    GetAllDocumentsOptions, Server, Value,
+    shared_does_not_exists_index, shared_empty_index, shared_index_with_geo_documents,
+    shared_index_with_test_set, GetAllDocumentsOptions, Server, Value,
 };
 use crate::json;
 
@@ -83,6 +83,311 @@ async fn get_document() {
     );
 }
 
+#[actix_rt::test]
+async fn get_document_sorted() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+    index.load_test_set().await;
+
+    let (task, _status_code) =
+        index.update_settings_sortable_attributes(json!(["age", "email", "gender", "name"])).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    let (response, _code) = index
+        .get_all_documents(GetAllDocumentsOptions {
+            fields: Some(vec!["id", "age", "email"]),
+            sort: Some(vec!["age:asc", "email:desc"]),
+            ..Default::default()
+        })
+        .await;
+    let results = response["results"].as_array().unwrap();
+    snapshot!(json_string!(results), @r#"
+    [
+      {
+        "id": 5,
+        "age": 20,
+        "email": "warrenwatson@chorizon.com"
+      },
+      {
+        "id": 6,
+        "age": 20,
+        "email": "sheliaberry@chorizon.com"
+      },
+      {
+        "id": 57,
+        "age": 20,
+        "email": "kaitlinconner@chorizon.com"
+      },
+      {
+        "id": 45,
+        "age": 20,
+        "email": "irenebennett@chorizon.com"
+      },
+      {
+        "id": 40,
+        "age": 21,
+        "email": "staffordemerson@chorizon.com"
+      },
+      {
+        "id": 41,
+        "age": 21,
+        "email": "salinasgamble@chorizon.com"
+      },
+      {
+        "id": 63,
+        "age": 21,
+        "email": "knowleshebert@chorizon.com"
+      },
+      {
+        "id": 50,
+        "age": 21,
+        "email": "guerramcintyre@chorizon.com"
+      },
+      {
+        "id": 44,
+        "age": 22,
+        "email": "jonispears@chorizon.com"
+      },
+      {
+        "id": 56,
+        "age": 23,
+        "email": "tuckerbarry@chorizon.com"
+      },
+      {
+        "id": 51,
+        "age": 23,
+        "email": "keycervantes@chorizon.com"
+      },
+      {
+        "id": 60,
+        "age": 23,
+        "email": "jodyherrera@chorizon.com"
+      },
+      {
+        "id": 70,
+        "age": 23,
+        "email": "glassperkins@chorizon.com"
+      },
+      {
+        "id": 75,
+        "age": 24,
+        "email": "emmajacobs@chorizon.com"
+      },
+      {
+        "id": 68,
+        "age": 24,
+        "email": "angelinadyer@chorizon.com"
+      },
+      {
+        "id": 17,
+        "age": 25,
+        "email": "ortegabrennan@chorizon.com"
+      },
+      {
+        "id": 76,
+        "age": 25,
+        "email": "claricegardner@chorizon.com"
+      },
+      {
+        "id": 43,
+        "age": 25,
+        "email": "arnoldbender@chorizon.com"
+      },
+      {
+        "id": 12,
+        "age": 25,
+        "email": "aidakirby@chorizon.com"
+      },
+      {
+        "id": 9,
+        "age": 26,
+        "email": "kellimendez@chorizon.com"
+      }
+    ]
+    "#);
+
+    let (response, _code) = index
+        .get_all_documents(GetAllDocumentsOptions {
+            fields: Some(vec!["id", "gender", "name"]),
+            sort: Some(vec!["gender:asc", "name:asc"]),
+            ..Default::default()
+        })
+        .await;
+    let results = response["results"].as_array().unwrap();
+    snapshot!(json_string!(results), @r#"
+    [
+      {
+        "id": 3,
+        "name": "Adeline Flynn",
+        "gender": "female"
+      },
+      {
+        "id": 12,
+        "name": "Aida Kirby",
+        "gender": "female"
+      },
+      {
+        "id": 68,
+        "name": "Angelina Dyer",
+        "gender": "female"
+      },
+      {
+        "id": 15,
+        "name": "Aurelia Contreras",
+        "gender": "female"
+      },
+      {
+        "id": 36,
+        "name": "Barbra Valenzuela",
+        "gender": "female"
+      },
+      {
+        "id": 23,
+        "name": "Blanca Mcclain",
+        "gender": "female"
+      },
+      {
+        "id": 53,
+        "name": "Caitlin Burnett",
+        "gender": "female"
+      },
+      {
+        "id": 71,
+        "name": "Candace Sawyer",
+        "gender": "female"
+      },
+      {
+        "id": 65,
+        "name": "Carole Rowland",
+        "gender": "female"
+      },
+      {
+        "id": 33,
+        "name": "Cecilia Greer",
+        "gender": "female"
+      },
+      {
+        "id": 1,
+        "name": "Cherry Orr",
+        "gender": "female"
+      },
+      {
+        "id": 38,
+        "name": "Christina Short",
+        "gender": "female"
+      },
+      {
+        "id": 7,
+        "name": "Chrystal Boyd",
+        "gender": "female"
+      },
+      {
+        "id": 76,
+        "name": "Clarice Gardner",
+        "gender": "female"
+      },
+      {
+        "id": 73,
+        "name": "Eleanor Shepherd",
+        "gender": "female"
+      },
+      {
+        "id": 75,
+        "name": "Emma Jacobs",
+        "gender": "female"
+      },
+      {
+        "id": 16,
+        "name": "Estella Bass",
+        "gender": "female"
+      },
+      {
+        "id": 62,
+        "name": "Estelle Ramirez",
+        "gender": "female"
+      },
+      {
+        "id": 20,
+        "name": "Florence Long",
+        "gender": "female"
+      },
+      {
+        "id": 42,
+        "name": "Graciela Russell",
+        "gender": "female"
+      }
+    ]
+    "#);
+}
+
+#[actix_rt::test]
+async fn get_document_geosorted() {
+    let index = shared_index_with_geo_documents().await;
+
+    let (response, _code) = index
+        .get_all_documents(GetAllDocumentsOptions {
+            sort: Some(vec!["_geoPoint(45.4777599, 9.1967508):asc"]),
+            ..Default::default()
+        })
+        .await;
+    let results = response["results"].as_array().unwrap();
+    snapshot!(json_string!(results), @r#"
+    [
+      {
+        "id": 2,
+        "name": "La Bella Italia",
+        "address": "456 Elm Street, Townsville",
+        "type": "Italian",
+        "rating": 9,
+        "_geo": {
+          "lat": "45.4777599",
+          "lng": "9.1967508"
+        }
+      },
+      {
+        "id": 1,
+        "name": "Taco Truck",
+        "address": "444 Salsa Street, Burritoville",
+        "type": "Mexican",
+        "rating": 9,
+        "_geo": {
+          "lat": 34.0522,
+          "lng": -118.2437
+        }
+      },
+      {
+        "id": 3,
+        "name": "Crêpe Truck",
+        "address": "2 Billig Avenue, Rouenville",
+        "type": "French",
+        "rating": 10
+      }
+    ]
+    "#);
+}
+
+#[actix_rt::test]
+async fn get_document_sort_the_unsortable() {
+    let index = shared_index_with_test_set().await;
+
+    let (response, _code) = index
+        .get_all_documents(GetAllDocumentsOptions {
+            fields: Some(vec!["id", "name"]),
+            sort: Some(vec!["name:asc"]),
+            ..Default::default()
+        })
+        .await;
+
+    snapshot!(json_string!(response), @r#"
+    {
+      "message": "Attribute `name` is not sortable. This index does not have configured sortable attributes.",
+      "code": "invalid_document_sort",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_document_sort"
+    }
+    "#);
+}
+
 #[actix_rt::test]
 async fn error_get_unexisting_index_all_documents() {
     let index = shared_does_not_exists_index().await;

crates/meilisearch/tests/search/errors.rs

@@ -304,7 +304,7 @@ async fn search_bad_filter() {
     let server = Server::new_shared();
     let index = server.unique_index();
     // Also, to trigger the error message we need to effectively create the index or else it'll throw an
-    // index does not exists error.
+    // index does not exist error.
     let (response, _code) = index.create(None).await;
     server.wait_task(response.uid()).await.succeeded();
 
@@ -1263,7 +1263,7 @@ async fn search_with_contains_without_enabling_the_feature() {
     let server = Server::new_shared();
     let index = server.unique_index();
     // Also, to trigger the error message we need to effectively create the index or else it'll throw an
-    // index does not exists error.
+    // index does not exist error.
     let (task, _code) = index.create(None).await;
     server.wait_task(task.uid()).await.succeeded();
 

crates/meilisearch/tests/search/filters.rs

@@ -4,8 +4,8 @@ use tempfile::TempDir;
 
 use super::test_settings_documents_indexing_swapping_and_search;
 use crate::common::{
-    default_settings, shared_index_with_documents, shared_index_with_nested_documents, Server,
-    DOCUMENTS, NESTED_DOCUMENTS,
+    default_settings, shared_index_for_fragments, shared_index_with_documents,
+    shared_index_with_nested_documents, Server, DOCUMENTS, NESTED_DOCUMENTS,
 };
 use crate::json;
 
@@ -731,3 +731,457 @@ async fn test_filterable_attributes_priority() {
     )
     .await;
 }
+
+#[actix_rt::test]
+async fn vector_filter_all_embedders() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [
+        {
+          "name": "kefir"
+        },
+        {
+          "name": "echo"
+        },
+        {
+          "name": "intel"
+        },
+        {
+          "name": "dustin"
+        }
+      ],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 4
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_missing_fragment() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.fragments EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "message": "Index `[uuid]`: Vector filter is inconsistent: either specify a fragment name or remove the `fragments` part.\n15:24 _vectors.rest.fragments EXISTS",
+      "code": "invalid_search_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_filter"
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_non_existant_embedder() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.other EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "message": "Index `[uuid]`: The embedder `other` does not exist. Available embedders are: `rest`.\n10:15 _vectors.other EXISTS",
+      "code": "invalid_search_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_filter"
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_all_embedders_user_provided() {
+    let index = shared_index_for_fragments().await;
+
+    // This one is counterintuitive, but it is the same as the previous one.
+    // It's because userProvided is interpreted as an embedder name
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.userProvided EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "message": "Index `[uuid]`: The embedder `userProvided` does not exist. Available embedders are: `rest`.\n10:22 _vectors.userProvided EXISTS",
+      "code": "invalid_search_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_filter"
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_specific_embedder() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [
+        {
+          "name": "kefir"
+        },
+        {
+          "name": "echo"
+        },
+        {
+          "name": "intel"
+        },
+        {
+          "name": "dustin"
+        }
+      ],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 4
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_user_provided() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.userProvided EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [
+        {
+          "name": "echo"
+        }
+      ],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 1
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_specific_fragment() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.fragments.withBreed EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [
+        {
+          "name": "intel"
+        },
+        {
+          "name": "dustin"
+        }
+      ],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 2
+    }
+    "#);
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.fragments.basic EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [
+        {
+          "name": "kefir"
+        },
+        {
+          "name": "echo"
+        },
+        {
+          "name": "intel"
+        },
+        {
+          "name": "dustin"
+        }
+      ],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 4
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_non_existant_fragment() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.fragments.other EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "message": "Index `[uuid]`: The fragment `other` does not exist on embedder `rest`. Available fragments on this embedder are: `basic`, `withBreed`.\n25:30 _vectors.rest.fragments.other EXISTS",
+      "code": "invalid_search_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_filter"
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_specific_fragment_user_provided() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.fragments.other.userProvided EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "message": "Index `[uuid]`: Vector filter cannot have both `other` and `userProvided`.\n31:43 _vectors.rest.fragments.other.userProvided EXISTS",
+      "code": "invalid_search_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_filter"
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_document_template_but_fragments_used() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.documentTemplate EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 0
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_document_template() {
+    let (_mock, setting) = crate::vector::create_mock().await;
+    let server = crate::vector::get_server_vector().await;
+    let index = server.index("doggo");
+
+    let (_response, code) = server.set_features(json!({"multimodal": true})).await;
+    snapshot!(code, @"200 OK");
+
+    let (response, code) = index
+        .update_settings(json!({
+            "embedders": {
+                "rest": setting,
+            },
+        }))
+        .await;
+    snapshot!(code, @"202 Accepted");
+    server.wait_task(response.uid()).await.succeeded();
+
+    let documents = json!([
+        {"id": 0, "name": "kefir"},
+        {"id": 1, "name": "echo", "_vectors": { "rest": [1, 1, 1] }},
+        {"id": 2, "name": "intel"},
+        {"id": 3, "name": "iko" }
+    ]);
+    let (value, code) = index.add_documents(documents, None).await;
+    snapshot!(code, @"202 Accepted");
+    index.wait_task(value.uid()).await.succeeded();
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.documentTemplate EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [
+        {
+          "name": "kefir"
+        },
+        {
+          "name": "intel"
+        },
+        {
+          "name": "iko"
+        }
+      ],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 3
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_feature_gate() {
+    let index = shared_index_with_documents().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "message": "using a vector filter requires enabling the `multimodal` experimental feature. See https://github.com/orgs/meilisearch/discussions/846\n1:9 _vectors EXISTS",
+      "code": "feature_not_enabled",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_negation() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.userProvided NOT EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [
+        {
+          "name": "kefir"
+        },
+        {
+          "name": "intel"
+        },
+        {
+          "name": "dustin"
+        }
+      ],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 3
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_or_combination() {
+    let index = shared_index_for_fragments().await;
+
+    let (value, _code) = index
+        .search_post(json!({
+            "filter": "_vectors.rest.fragments.withBreed EXISTS OR _vectors.rest.userProvided EXISTS",
+            "attributesToRetrieve": ["name"]
+        }))
+        .await;
+    snapshot!(value, @r#"
+    {
+      "hits": [
+        {
+          "name": "echo"
+        },
+        {
+          "name": "intel"
+        },
+        {
+          "name": "dustin"
+        }
+      ],
+      "query": "",
+      "processingTimeMs": "[duration]",
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 3
+    }
+    "#);
+}
+
+#[actix_rt::test]
+async fn vector_filter_regenerate() {
+    let index = shared_index_for_fragments().await;
+
+    for selector in ["_vectors.rest.regenerate", "_vectors.rest.fragments.basic.regenerate"] {
+        let (value, _code) = index
+            .search_post(json!({
+                "filter": format!("{selector} EXISTS"),
+                "attributesToRetrieve": ["name"]
+            }))
+            .await;
+        snapshot!(value, @r#"
+            {
+              "hits": [
+                {
+                  "name": "kefir"
+                },
+                {
+                  "name": "intel"
+                },
+                {
+                  "name": "dustin"
+                }
+              ],
+              "query": "",
+              "processingTimeMs": "[duration]",
+              "limit": 20,
+              "offset": 0,
+              "estimatedTotalHits": 3
+            }
+            "#);
+    }
+}

crates/meilisearch/tests/search/multi/proxy.rs

@@ -2499,7 +2499,7 @@ pub struct LocalMeiliParams {
 
 /// A server that exploits [`MockServer`] to provide an URL for testing network and the network.
 pub struct LocalMeili {
-    mock_server: MockServer,
+    mock_server: &'static MockServer,
 }
 
 impl LocalMeili {
@@ -2508,7 +2508,7 @@ impl LocalMeili {
     }
 
     pub async fn with_params(server: Arc<Server>, params: LocalMeiliParams) -> Self {
-        let mock_server = MockServer::start().await;
+        let mock_server = Box::leak(Box::new(MockServer::start().await));
 
         // tokio won't let us execute asynchronous code from a sync function inside of an async test,
         // so instead we spawn another thread that will call the service on a brand new tokio runtime
@@ -2572,7 +2572,7 @@ impl LocalMeili {
                     response.set_body_json(value)
                 }
             })
-            .mount(&mock_server)
+            .mount(mock_server)
             .await;
         Self { mock_server }
     }

crates/meilisearch/tests/search/pagination.rs

@@ -1,6 +1,7 @@
 use super::shared_index_with_documents;
 use crate::common::Server;
 use crate::json;
+use meili_snap::{json_string, snapshot};
 
 #[actix_rt::test]
 async fn default_search_should_return_estimated_total_hit() {
@@ -133,3 +134,61 @@ async fn ensure_placeholder_search_hit_count_valid() {
             .await;
     }
 }
+
+#[actix_rt::test]
+async fn test_issue_5274() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    let documents = json!([
+        {
+            "id": 1,
+            "title": "Document 1",
+            "content": "This is the first."
+        },
+        {
+            "id": 2,
+            "title": "Document 2",
+            "content": "This is the second doc."
+        }
+    ]);
+    let (task, _code) = index.add_documents(documents, None).await;
+    server.wait_task(task.uid()).await.succeeded();
+
+    // Find out the lowest ranking score among the documents
+    let (rep, _status) = index
+        .search_post(json!({"q": "doc", "page": 1, "hitsPerPage": 2, "showRankingScore": true}))
+        .await;
+    let hits = rep["hits"].as_array().expect("Missing hits array");
+    let second_hit = hits.get(1).expect("Missing second hit");
+    let ranking_score = second_hit
+        .get("_rankingScore")
+        .expect("Missing _rankingScore field")
+        .as_f64()
+        .expect("Expected _rankingScore to be a f64");
+
+    // Search with a ranking score threshold just above and expect to be a single hit
+    let (rep, _status) = index
+        .search_post(json!({"q": "doc", "page": 1, "hitsPerPage": 1, "rankingScoreThreshold": ranking_score + 0.0001}))
+        .await;
+
+    snapshot!(json_string!(rep, {
+        ".processingTimeMs" => "[ignored]",
+    }), @r#"
+    {
+      "hits": [
+        {
+          "id": 2,
+          "title": "Document 2",
+          "content": "This is the second doc."
+        }
+      ],
+      "query": "doc",
+      "processingTimeMs": "[ignored]",
+      "hitsPerPage": 1,
+      "page": 1,
+      "totalPages": 1,
+      "totalHits": 1
+    }
+    "#);
+}

crates/meilisearch/tests/settings/get_settings.rs

@@ -692,3 +692,68 @@ async fn granular_filterable_attributes() {
     ]
     "###);
 }
+
+#[actix_rt::test]
+async fn test_searchable_attributes_order() {
+    let server = Server::new_shared();
+    let index = server.unique_index();
+
+    // 1) Create an index with settings "searchableAttributes": ["title", "overview"]
+    let (response, code) = index.create(None).await;
+    assert_eq!(code, 202, "{response}");
+    server.wait_task(response.uid()).await.succeeded();
+
+    let (task, code) = index
+        .update_settings(json!({
+            "searchableAttributes": ["title", "overview"]
+        }))
+        .await;
+    assert_eq!(code, 202, "{task}");
+    server.wait_task(task.uid()).await.succeeded();
+
+    // 2) Add documents in the index
+    let documents = json!([
+        {
+            "id": 1,
+            "title": "The Matrix",
+            "overview": "A computer hacker learns from mysterious rebels about the true nature of his reality."
+        },
+        {
+            "id": 2,
+            "title": "Inception",
+            "overview": "A thief who steals corporate secrets through dream-sharing technology."
+        }
+    ]);
+
+    let (response, code) = index.add_documents(documents, None).await;
+    assert_eq!(code, 202, "{response}");
+    server.wait_task(response.uid()).await.succeeded();
+
+    // 3) Modify the settings "searchableAttributes": ["overview", "title"] (overview is put first)
+    let (task, code) = index
+        .update_settings(json!({
+            "searchableAttributes": ["overview", "title"]
+        }))
+        .await;
+    assert_eq!(code, 202, "{task}");
+    server.wait_task(task.uid()).await.succeeded();
+
+    // 4) Check if it has been applied
+    let (response, code) = index.settings().await;
+    assert_eq!(code, 200, "{response}");
+    assert_eq!(response["searchableAttributes"], json!(["overview", "title"]));
+
+    // 5) Re-modify the settings "searchableAttributes": ["title", "overview"] (title is put first)
+    let (task, code) = index
+        .update_settings(json!({
+            "searchableAttributes": ["title", "overview"]
+        }))
+        .await;
+    assert_eq!(code, 202, "{task}");
+    server.wait_task(task.uid()).await.succeeded();
+
+    // 6) Check if it has been applied
+    let (response, code) = index.settings().await;
+    assert_eq!(code, 200, "{response}");
+    assert_eq!(response["searchableAttributes"], json!(["title", "overview"]));
+}

crates/meilisearch/tests/vector/mod.rs

@@ -1,4 +1,5 @@
 mod binary_quantized;
+mod fragments;
 #[cfg(feature = "test-ollama")]
 mod ollama;
 mod openai;
@@ -13,8 +14,9 @@ use meilisearch::option::MaxThreads;
 use crate::common::index::Index;
 use crate::common::{default_settings, GetAllDocumentsOptions, Server};
 use crate::json;
+pub use rest::create_mock;
 
-async fn get_server_vector() -> Server {
+pub async fn get_server_vector() -> Server {
     Server::new().await
 }
 

crates/meilisearch/tests/vector/openai.rs

@@ -136,7 +136,7 @@ fn long_text() -> &'static str {
     })
 }
 
-async fn create_mock_tokenized() -> (MockServer, Value) {
+async fn create_mock_tokenized() -> (&'static MockServer, Value) {
     create_mock_with_template("{{doc.text}}", ModelDimensions::Large, false, false).await
 }
 
@@ -145,8 +145,8 @@ async fn create_mock_with_template(
     model_dimensions: ModelDimensions,
     fallible: bool,
     slow: bool,
-) -> (MockServer, Value) {
-    let mock_server = MockServer::start().await;
+) -> (&'static MockServer, Value) {
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
     const API_KEY: &str = "my-api-key";
     const API_KEY_BEARER: &str = "Bearer my-api-key";
 
@@ -299,7 +299,7 @@ async fn create_mock_with_template(
                 }
             }))
         })
-        .mount(&mock_server)
+        .mount(mock_server)
         .await;
     let url = mock_server.uri();
 
@@ -321,27 +321,27 @@ const DOGGO_TEMPLATE: &str = r#"{%- if doc.gender == "F" -%}Une chienne nommée
         Un chien nommé {{doc.name}}, né en {{doc.birthyear}}
         {%- endif %}, de race {{doc.breed}}."#;
 
-async fn create_mock() -> (MockServer, Value) {
+async fn create_mock() -> (&'static MockServer, Value) {
     create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large, false, false).await
 }
 
-async fn create_mock_dimensions() -> (MockServer, Value) {
+async fn create_mock_dimensions() -> (&'static MockServer, Value) {
     create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large512, false, false).await
 }
 
-async fn create_mock_small_embedding_model() -> (MockServer, Value) {
+async fn create_mock_small_embedding_model() -> (&'static MockServer, Value) {
     create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Small, false, false).await
 }
 
-async fn create_mock_legacy_embedding_model() -> (MockServer, Value) {
+async fn create_mock_legacy_embedding_model() -> (&'static MockServer, Value) {
     create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Ada, false, false).await
 }
 
-async fn create_fallible_mock() -> (MockServer, Value) {
+async fn create_fallible_mock() -> (&'static MockServer, Value) {
     create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large, true, false).await
 }
 
-async fn create_slow_mock() -> (MockServer, Value) {
+async fn create_slow_mock() -> (&'static MockServer, Value) {
     create_mock_with_template(DOGGO_TEMPLATE, ModelDimensions::Large, true, true).await
 }
 

crates/meilisearch/tests/vector/rest.rs

@@ -12,8 +12,8 @@ use crate::common::Value;
 use crate::json;
 use crate::vector::{get_server_vector, GetAllDocumentsOptions};
 
-async fn create_mock() -> (MockServer, Value) {
-    let mock_server = MockServer::start().await;
+pub async fn create_mock() -> (&'static MockServer, Value) {
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
 
     let text_to_embedding: BTreeMap<_, _> = vec![
         // text -> embedding
@@ -32,7 +32,7 @@ async fn create_mock() -> (MockServer, Value) {
                 json!({ "data": text_to_embedding.get(text.as_str()).unwrap_or(&[99., 99., 99.]) }),
             )
         })
-        .mount(&mock_server)
+        .mount(mock_server)
         .await;
     let url = mock_server.uri();
 
@@ -50,8 +50,8 @@ async fn create_mock() -> (MockServer, Value) {
     (mock_server, embedder_settings)
 }
 
-async fn create_mock_default_template() -> (MockServer, Value) {
-    let mock_server = MockServer::start().await;
+async fn create_mock_default_template() -> (&'static MockServer, Value) {
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
 
     let text_to_embedding: BTreeMap<_, _> = vec![
         // text -> embedding
@@ -73,7 +73,7 @@ async fn create_mock_default_template() -> (MockServer, Value) {
                     .set_body_json(json!({"error": "text not found", "text": text})),
             }
         })
-        .mount(&mock_server)
+        .mount(mock_server)
         .await;
     let url = mock_server.uri();
 
@@ -106,8 +106,8 @@ struct SingleResponse {
     embedding: Vec<f32>,
 }
 
-async fn create_mock_multiple() -> (MockServer, Value) {
-    let mock_server = MockServer::start().await;
+async fn create_mock_multiple() -> (&'static MockServer, Value) {
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
 
     let text_to_embedding: BTreeMap<_, _> = vec![
         // text -> embedding
@@ -146,7 +146,7 @@ async fn create_mock_multiple() -> (MockServer, Value) {
 
             ResponseTemplate::new(200).set_body_json(response)
         })
-        .mount(&mock_server)
+        .mount(mock_server)
         .await;
     let url = mock_server.uri();
 
@@ -176,8 +176,8 @@ struct SingleRequest {
     input: String,
 }
 
-async fn create_mock_single_response_in_array() -> (MockServer, Value) {
-    let mock_server = MockServer::start().await;
+async fn create_mock_single_response_in_array() -> (&'static MockServer, Value) {
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
 
     let text_to_embedding: BTreeMap<_, _> = vec![
         // text -> embedding
@@ -212,7 +212,7 @@ async fn create_mock_single_response_in_array() -> (MockServer, Value) {
 
             ResponseTemplate::new(200).set_body_json(response)
         })
-        .mount(&mock_server)
+        .mount(mock_server)
         .await;
     let url = mock_server.uri();
 
@@ -236,8 +236,8 @@ async fn create_mock_single_response_in_array() -> (MockServer, Value) {
     (mock_server, embedder_settings)
 }
 
-async fn create_mock_raw_with_custom_header() -> (MockServer, Value) {
-    let mock_server = MockServer::start().await;
+async fn create_mock_raw_with_custom_header() -> (&'static MockServer, Value) {
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
 
     let text_to_embedding: BTreeMap<_, _> = vec![
         // text -> embedding
@@ -277,7 +277,7 @@ async fn create_mock_raw_with_custom_header() -> (MockServer, Value) {
 
             ResponseTemplate::new(200).set_body_json(output)
         })
-        .mount(&mock_server)
+        .mount(mock_server)
         .await;
     let url = mock_server.uri();
 
@@ -293,8 +293,8 @@ async fn create_mock_raw_with_custom_header() -> (MockServer, Value) {
     (mock_server, embedder_settings)
 }
 
-async fn create_mock_raw() -> (MockServer, Value) {
-    let mock_server = MockServer::start().await;
+async fn create_mock_raw() -> (&'static MockServer, Value) {
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
 
     let text_to_embedding: BTreeMap<_, _> = vec![
         // text -> embedding
@@ -321,7 +321,7 @@ async fn create_mock_raw() -> (MockServer, Value) {
 
             ResponseTemplate::new(200).set_body_json(output)
         })
-        .mount(&mock_server)
+        .mount(mock_server)
         .await;
     let url = mock_server.uri();
 
@@ -337,8 +337,8 @@ async fn create_mock_raw() -> (MockServer, Value) {
     (mock_server, embedder_settings)
 }
 
-async fn create_faulty_mock_raw(sender: mpsc::Sender<()>) -> (MockServer, Value) {
-    let mock_server = MockServer::start().await;
+async fn create_faulty_mock_raw(sender: mpsc::Sender<()>) -> (&'static MockServer, Value) {
+    let mock_server = Box::leak(Box::new(MockServer::start().await));
     let count = AtomicUsize::new(0);
 
     Mock::given(method("POST"))
@@ -355,7 +355,7 @@ async fn create_faulty_mock_raw(sender: mpsc::Sender<()>) -> (MockServer, Value)
                 ResponseTemplate::new(500).set_body_string("Service Unavailable")
             }
         })
-        .mount(&mock_server)
+        .mount(mock_server)
         .await;
 
     let url = mock_server.uri();

crates/meilisearch/tests/vector/settings.rs

@@ -101,14 +101,7 @@ async fn reset_embedder_documents() {
     server.wait_task(response.uid()).await;
 
     // Make sure the documents are still present
-    let (documents, _code) = index
-        .get_all_documents(GetAllDocumentsOptions {
-            limit: None,
-            offset: None,
-            retrieve_vectors: false,
-            fields: None,
-        })
-        .await;
+    let (documents, _code) = index.get_all_documents(GetAllDocumentsOptions::default()).await;
     snapshot!(json_string!(documents), @r###"
     {
       "results": [

crates/milli/src/asc_desc.rs

@@ -168,6 +168,16 @@ pub enum SortError {
     ReservedNameForFilter { name: String },
 }
 
+impl SortError {
+    pub fn into_search_error(self) -> Error {
+        Error::UserError(UserError::SortError { error: self, search: true })
+    }
+
+    pub fn into_document_error(self) -> Error {
+        Error::UserError(UserError::SortError { error: self, search: false })
+    }
+}
+
 impl From<AscDescError> for SortError {
     fn from(error: AscDescError) -> Self {
         match error {
@@ -190,12 +200,6 @@ impl From<AscDescError> for SortError {
     }
 }
 
-impl From<SortError> for Error {
-    fn from(error: SortError) -> Self {
-        Self::UserError(UserError::SortError(error))
-    }
-}
-
 #[cfg(test)]
 mod tests {
     use big_s::S;

crates/milli/src/documents/geo_sort.rs

@@ -0,0 +1,294 @@
+use crate::{
+    distance_between_two_points,
+    heed_codec::facet::{FieldDocIdFacetCodec, OrderedF64Codec},
+    lat_lng_to_xyz,
+    search::new::{facet_string_values, facet_values_prefix_key},
+    GeoPoint, Index,
+};
+use heed::{
+    types::{Bytes, Unit},
+    RoPrefix, RoTxn,
+};
+use roaring::RoaringBitmap;
+use rstar::RTree;
+use std::collections::VecDeque;
+
+#[derive(Debug, Clone, Copy)]
+pub struct GeoSortParameter {
+    // Define the strategy used by the geo sort
+    pub strategy: GeoSortStrategy,
+    // Limit the number of docs in a single bucket to avoid unexpectedly large overhead
+    pub max_bucket_size: u64,
+    // Considering the errors of GPS and geographical calculations, distances less than distance_error_margin will be treated as equal
+    pub distance_error_margin: f64,
+}
+
+impl Default for GeoSortParameter {
+    fn default() -> Self {
+        Self {
+            strategy: GeoSortStrategy::default(),
+            max_bucket_size: 1000,
+            distance_error_margin: 1.0,
+        }
+    }
+}
+/// Define the strategy used by the geo sort.
+/// The parameter represents the cache size, and, in the case of the Dynamic strategy,
+/// the point where we move from using the iterative strategy to the rtree.
+#[derive(Debug, Clone, Copy)]
+pub enum GeoSortStrategy {
+    AlwaysIterative(usize),
+    AlwaysRtree(usize),
+    Dynamic(usize),
+}
+
+impl Default for GeoSortStrategy {
+    fn default() -> Self {
+        GeoSortStrategy::Dynamic(1000)
+    }
+}
+
+impl GeoSortStrategy {
+    pub fn use_rtree(&self, candidates: usize) -> bool {
+        match self {
+            GeoSortStrategy::AlwaysIterative(_) => false,
+            GeoSortStrategy::AlwaysRtree(_) => true,
+            GeoSortStrategy::Dynamic(i) => candidates >= *i,
+        }
+    }
+
+    pub fn cache_size(&self) -> usize {
+        match self {
+            GeoSortStrategy::AlwaysIterative(i)
+            | GeoSortStrategy::AlwaysRtree(i)
+            | GeoSortStrategy::Dynamic(i) => *i,
+        }
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+pub fn fill_cache(
+    index: &Index,
+    txn: &RoTxn<heed::AnyTls>,
+    strategy: GeoSortStrategy,
+    ascending: bool,
+    target_point: [f64; 2],
+    field_ids: &Option<[u16; 2]>,
+    rtree: &mut Option<RTree<GeoPoint>>,
+    geo_candidates: &RoaringBitmap,
+    cached_sorted_docids: &mut VecDeque<(u32, [f64; 2])>,
+) -> crate::Result<()> {
+    debug_assert!(cached_sorted_docids.is_empty());
+
+    // lazily initialize the rtree if needed by the strategy, and cache it in `self.rtree`
+    let rtree = if strategy.use_rtree(geo_candidates.len() as usize) {
+        if let Some(rtree) = rtree.as_ref() {
+            // get rtree from cache
+            Some(rtree)
+        } else {
+            let rtree2 = index.geo_rtree(txn)?.expect("geo candidates but no rtree");
+            // insert rtree in cache and returns it.
+            // Can't use `get_or_insert_with` because getting the rtree from the DB is a fallible operation.
+            Some(&*rtree.insert(rtree2))
+        }
+    } else {
+        None
+    };
+
+    let cache_size = strategy.cache_size();
+    if let Some(rtree) = rtree {
+        if ascending {
+            let point = lat_lng_to_xyz(&target_point);
+            for point in rtree.nearest_neighbor_iter(&point) {
+                if geo_candidates.contains(point.data.0) {
+                    cached_sorted_docids.push_back(point.data);
+                    if cached_sorted_docids.len() >= cache_size {
+                        break;
+                    }
+                }
+            }
+        } else {
+            // in the case of the desc geo sort we look for the closest point to the opposite of the queried point
+            // and we insert the points in reverse order they get reversed when emptying the cache later on
+            let point = lat_lng_to_xyz(&opposite_of(target_point));
+            for point in rtree.nearest_neighbor_iter(&point) {
+                if geo_candidates.contains(point.data.0) {
+                    cached_sorted_docids.push_front(point.data);
+                    if cached_sorted_docids.len() >= cache_size {
+                        break;
+                    }
+                }
+            }
+        }
+    } else {
+        // the iterative version
+        let [lat, lng] = field_ids.expect("fill_buffer can't be called without the lat&lng");
+
+        let mut documents = geo_candidates
+            .iter()
+            .map(|id| -> crate::Result<_> { Ok((id, geo_value(id, lat, lng, index, txn)?)) })
+            .collect::<crate::Result<Vec<(u32, [f64; 2])>>>()?;
+        // computing the distance between two points is expensive thus we cache the result
+        documents
+            .sort_by_cached_key(|(_, p)| distance_between_two_points(&target_point, p) as usize);
+        cached_sorted_docids.extend(documents);
+    };
+
+    Ok(())
+}
+
+#[allow(clippy::too_many_arguments)]
+pub fn next_bucket(
+    index: &Index,
+    txn: &RoTxn<heed::AnyTls>,
+    universe: &RoaringBitmap,
+    ascending: bool,
+    target_point: [f64; 2],
+    field_ids: &Option<[u16; 2]>,
+    rtree: &mut Option<RTree<GeoPoint>>,
+    cached_sorted_docids: &mut VecDeque<(u32, [f64; 2])>,
+    geo_candidates: &RoaringBitmap,
+    parameter: GeoSortParameter,
+) -> crate::Result<Option<(RoaringBitmap, Option<[f64; 2]>)>> {
+    let mut geo_candidates = geo_candidates & universe;
+
+    if geo_candidates.is_empty() {
+        return Ok(Some((universe.clone(), None)));
+    }
+
+    let next = |cache: &mut VecDeque<_>| {
+        if ascending {
+            cache.pop_front()
+        } else {
+            cache.pop_back()
+        }
+    };
+    let put_back = |cache: &mut VecDeque<_>, x: _| {
+        if ascending {
+            cache.push_front(x)
+        } else {
+            cache.push_back(x)
+        }
+    };
+
+    let mut current_bucket = RoaringBitmap::new();
+    // current_distance stores the first point and distance in current bucket
+    let mut current_distance: Option<([f64; 2], f64)> = None;
+    loop {
+        // The loop will only exit when we have found all points with equal distance or have exhausted the candidates.
+        if let Some((id, point)) = next(cached_sorted_docids) {
+            if geo_candidates.contains(id) {
+                let distance = distance_between_two_points(&target_point, &point);
+                if let Some((point0, bucket_distance)) = current_distance.as_ref() {
+                    if (bucket_distance - distance).abs() > parameter.distance_error_margin {
+                        // different distance, point belongs to next bucket
+                        put_back(cached_sorted_docids, (id, point));
+                        return Ok(Some((current_bucket, Some(point0.to_owned()))));
+                    } else {
+                        // same distance, point belongs to current bucket
+                        current_bucket.insert(id);
+                        // remove from candidates to prevent it from being added to the cache again
+                        geo_candidates.remove(id);
+                        // current bucket size reaches limit, force return
+                        if current_bucket.len() == parameter.max_bucket_size {
+                            return Ok(Some((current_bucket, Some(point0.to_owned()))));
+                        }
+                    }
+                } else {
+                    // first doc in current bucket
+                    current_distance = Some((point, distance));
+                    current_bucket.insert(id);
+                    geo_candidates.remove(id);
+                    // current bucket size reaches limit, force return
+                    if current_bucket.len() == parameter.max_bucket_size {
+                        return Ok(Some((current_bucket, Some(point.to_owned()))));
+                    }
+                }
+            }
+        } else {
+            // cache exhausted, we need to refill it
+            fill_cache(
+                index,
+                txn,
+                parameter.strategy,
+                ascending,
+                target_point,
+                field_ids,
+                rtree,
+                &geo_candidates,
+                cached_sorted_docids,
+            )?;
+
+            if cached_sorted_docids.is_empty() {
+                // candidates exhausted, exit
+                if let Some((point0, _)) = current_distance.as_ref() {
+                    return Ok(Some((current_bucket, Some(point0.to_owned()))));
+                } else {
+                    return Ok(Some((universe.clone(), None)));
+                }
+            }
+        }
+    }
+}
+
+/// Return an iterator over each number value in the given field of the given document.
+fn facet_number_values<'a>(
+    docid: u32,
+    field_id: u16,
+    index: &Index,
+    txn: &'a RoTxn<'a>,
+) -> crate::Result<RoPrefix<'a, FieldDocIdFacetCodec<OrderedF64Codec>, Unit>> {
+    let key = facet_values_prefix_key(field_id, docid);
+
+    let iter = index
+        .field_id_docid_facet_f64s
+        .remap_key_type::<Bytes>()
+        .prefix_iter(txn, &key)?
+        .remap_key_type();
+
+    Ok(iter)
+}
+
+/// Extracts the lat and long values from a single document.
+///
+/// If it is not able to find it in the facet number index it will extract it
+/// from the facet string index and parse it as f64 (as the geo extraction behaves).
+pub(crate) fn geo_value(
+    docid: u32,
+    field_lat: u16,
+    field_lng: u16,
+    index: &Index,
+    rtxn: &RoTxn<'_>,
+) -> crate::Result<[f64; 2]> {
+    let extract_geo = |geo_field: u16| -> crate::Result<f64> {
+        match facet_number_values(docid, geo_field, index, rtxn)?.next() {
+            Some(Ok(((_, _, geo), ()))) => Ok(geo),
+            Some(Err(e)) => Err(e.into()),
+            None => match facet_string_values(docid, geo_field, index, rtxn)?.next() {
+                Some(Ok((_, geo))) => {
+                    Ok(geo.parse::<f64>().expect("cannot parse geo field as f64"))
+                }
+                Some(Err(e)) => Err(e.into()),
+                None => panic!("A geo faceted document doesn't contain any lat or lng"),
+            },
+        }
+    };
+
+    let lat = extract_geo(field_lat)?;
+    let lng = extract_geo(field_lng)?;
+
+    Ok([lat, lng])
+}
+
+/// Compute the antipodal coordinate of `coord`
+pub(crate) fn opposite_of(mut coord: [f64; 2]) -> [f64; 2] {
+    coord[0] *= -1.;
+    // in the case of x,0 we want to return x,180
+    if coord[1] > 0. {
+        coord[1] -= 180.;
+    } else {
+        coord[1] += 180.;
+    }
+
+    coord
+}

crates/milli/src/documents/mod.rs

@@ -1,8 +1,10 @@
 mod builder;
 mod enriched;
+pub mod geo_sort;
 mod primary_key;
 mod reader;
 mod serde_impl;
+pub mod sort;
 
 use std::fmt::Debug;
 use std::io;
@@ -19,6 +21,7 @@ pub use primary_key::{
 pub use reader::{DocumentsBatchCursor, DocumentsBatchCursorError, DocumentsBatchReader};
 use serde::{Deserialize, Serialize};
 
+pub use self::geo_sort::{GeoSortParameter, GeoSortStrategy};
 use crate::error::{FieldIdMapMissingEntry, InternalError};
 use crate::{FieldId, Object, Result};
 

crates/milli/src/documents/sort.rs

@@ -0,0 +1,444 @@
+use std::collections::{BTreeSet, VecDeque};
+
+use crate::{
+    constants::RESERVED_GEO_FIELD_NAME,
+    documents::{geo_sort::next_bucket, GeoSortParameter},
+    heed_codec::{
+        facet::{FacetGroupKeyCodec, FacetGroupValueCodec},
+        BytesRefCodec,
+    },
+    is_faceted,
+    search::facet::{ascending_facet_sort, descending_facet_sort},
+    AscDesc, DocumentId, Member, UserError,
+};
+use heed::Database;
+use roaring::RoaringBitmap;
+
+#[derive(Debug, Clone, Copy)]
+enum AscDescId {
+    Facet { field_id: u16, ascending: bool },
+    Geo { field_ids: [u16; 2], target_point: [f64; 2], ascending: bool },
+}
+
+/// A [`SortedDocumentsIterator`] allows efficient access to a continuous range of sorted documents.
+/// This is ideal in the context of paginated queries in which only a small number of documents are needed at a time.
+/// Search operations will only be performed upon access.
+pub enum SortedDocumentsIterator<'ctx> {
+    Leaf {
+        /// The exact number of documents remaining
+        size: usize,
+        values: Box<dyn Iterator<Item = DocumentId> + 'ctx>,
+    },
+    Branch {
+        /// The current child, got from the children iterator
+        current_child: Option<Box<SortedDocumentsIterator<'ctx>>>,
+        /// The exact number of documents remaining, excluding documents in the current child
+        next_children_size: usize,
+        /// Iterators to become the current child once it is exhausted
+        next_children:
+            Box<dyn Iterator<Item = crate::Result<SortedDocumentsIteratorBuilder<'ctx>>> + 'ctx>,
+    },
+}
+
+impl SortedDocumentsIterator<'_> {
+    /// Takes care of updating the current child if it is `None`, and also updates the size
+    fn update_current<'ctx>(
+        current_child: &mut Option<Box<SortedDocumentsIterator<'ctx>>>,
+        next_children_size: &mut usize,
+        next_children: &mut Box<
+            dyn Iterator<Item = crate::Result<SortedDocumentsIteratorBuilder<'ctx>>> + 'ctx,
+        >,
+    ) -> crate::Result<()> {
+        if current_child.is_none() {
+            *current_child = match next_children.next() {
+                Some(Ok(builder)) => {
+                    let next_child = Box::new(builder.build()?);
+                    *next_children_size -= next_child.size_hint().0;
+                    Some(next_child)
+                }
+                Some(Err(e)) => return Err(e),
+                None => return Ok(()),
+            };
+        }
+        Ok(())
+    }
+}
+
+impl Iterator for SortedDocumentsIterator<'_> {
+    type Item = crate::Result<DocumentId>;
+
+    /// Implementing the `nth` method allows for efficient access to the nth document in the sorted order.
+    /// It's used by `skip` internally.
+    /// The default implementation of `nth` would iterate over all children, which is inefficient for large datasets.
+    /// This implementation will jump over whole chunks of children until it gets close.
+    fn nth(&mut self, n: usize) -> Option<Self::Item> {
+        if n == 0 {
+            return self.next();
+        }
+
+        // If it's at the leaf level, just forward the call to the values iterator
+        let (current_child, next_children, next_children_size) = match self {
+            SortedDocumentsIterator::Leaf { values, size } => {
+                *size = size.saturating_sub(n);
+                return values.nth(n).map(Ok);
+            }
+            SortedDocumentsIterator::Branch {
+                current_child,
+                next_children,
+                next_children_size,
+            } => (current_child, next_children, next_children_size),
+        };
+
+        // Otherwise don't directly iterate over children, skip them if we know we will go further
+        let mut to_skip = n - 1;
+        while to_skip > 0 {
+            if let Err(e) = SortedDocumentsIterator::update_current(
+                current_child,
+                next_children_size,
+                next_children,
+            ) {
+                return Some(Err(e));
+            }
+            let Some(inner) = current_child else {
+                return None; // No more inner iterators, everything has been consumed.
+            };
+
+            if to_skip >= inner.size_hint().0 {
+                // The current child isn't large enough to contain the nth element.
+                // Skip it and continue with the next one.
+                to_skip -= inner.size_hint().0;
+                *current_child = None;
+                continue;
+            } else {
+                // The current iterator is large enough, so we can forward the call to it.
+                return inner.nth(to_skip + 1);
+            }
+        }
+
+        self.next()
+    }
+
+    /// Iterators need to keep track of their size so that they can be skipped efficiently by the `nth` method.
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        let size = match self {
+            SortedDocumentsIterator::Leaf { size, .. } => *size,
+            SortedDocumentsIterator::Branch {
+                next_children_size,
+                current_child: Some(current_child),
+                ..
+            } => current_child.size_hint().0 + next_children_size,
+            SortedDocumentsIterator::Branch { next_children_size, current_child: None, .. } => {
+                *next_children_size
+            }
+        };
+
+        (size, Some(size))
+    }
+
+    fn next(&mut self) -> Option<Self::Item> {
+        match self {
+            SortedDocumentsIterator::Leaf { values, size } => {
+                let result = values.next().map(Ok);
+                if result.is_some() {
+                    *size -= 1;
+                }
+                result
+            }
+            SortedDocumentsIterator::Branch {
+                current_child,
+                next_children_size,
+                next_children,
+            } => {
+                let mut result = None;
+                while result.is_none() {
+                    // Ensure we have selected an iterator to work with
+                    if let Err(e) = SortedDocumentsIterator::update_current(
+                        current_child,
+                        next_children_size,
+                        next_children,
+                    ) {
+                        return Some(Err(e));
+                    }
+                    let Some(inner) = current_child else {
+                        return None;
+                    };
+
+                    result = inner.next();
+
+                    // If the current iterator is exhausted, we need to try the next one
+                    if result.is_none() {
+                        *current_child = None;
+                    }
+                }
+                result
+            }
+        }
+    }
+}
+
+/// Builder for a [`SortedDocumentsIterator`].
+/// Most builders won't ever be built, because pagination will skip them.
+pub struct SortedDocumentsIteratorBuilder<'ctx> {
+    index: &'ctx crate::Index,
+    rtxn: &'ctx heed::RoTxn<'ctx>,
+    number_db: Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
+    string_db: Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
+    fields: &'ctx [AscDescId],
+    candidates: RoaringBitmap,
+    geo_candidates: &'ctx RoaringBitmap,
+}
+
+impl<'ctx> SortedDocumentsIteratorBuilder<'ctx> {
+    /// Performs the sort and builds a [`SortedDocumentsIterator`].
+    fn build(self) -> crate::Result<SortedDocumentsIterator<'ctx>> {
+        let size = self.candidates.len() as usize;
+
+        match self.fields {
+            [] => Ok(SortedDocumentsIterator::Leaf {
+                size,
+                values: Box::new(self.candidates.into_iter()),
+            }),
+            [AscDescId::Facet { field_id, ascending }, next_fields @ ..] => {
+                SortedDocumentsIteratorBuilder::build_facet(
+                    self.index,
+                    self.rtxn,
+                    self.number_db,
+                    self.string_db,
+                    next_fields,
+                    self.candidates,
+                    self.geo_candidates,
+                    *field_id,
+                    *ascending,
+                )
+            }
+            [AscDescId::Geo { field_ids, target_point, ascending }, next_fields @ ..] => {
+                SortedDocumentsIteratorBuilder::build_geo(
+                    self.index,
+                    self.rtxn,
+                    self.number_db,
+                    self.string_db,
+                    next_fields,
+                    self.candidates,
+                    self.geo_candidates,
+                    *field_ids,
+                    *target_point,
+                    *ascending,
+                )
+            }
+        }
+    }
+
+    /// Builds a [`SortedDocumentsIterator`] based on the results of a facet sort.
+    #[allow(clippy::too_many_arguments)]
+    fn build_facet(
+        index: &'ctx crate::Index,
+        rtxn: &'ctx heed::RoTxn<'ctx>,
+        number_db: Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
+        string_db: Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
+        next_fields: &'ctx [AscDescId],
+        candidates: RoaringBitmap,
+        geo_candidates: &'ctx RoaringBitmap,
+        field_id: u16,
+        ascending: bool,
+    ) -> crate::Result<SortedDocumentsIterator<'ctx>> {
+        let size = candidates.len() as usize;
+
+        // Perform the sort on the first field
+        let (number_iter, string_iter) = if ascending {
+            let number_iter = ascending_facet_sort(rtxn, number_db, field_id, candidates.clone())?;
+            let string_iter = ascending_facet_sort(rtxn, string_db, field_id, candidates)?;
+
+            (itertools::Either::Left(number_iter), itertools::Either::Left(string_iter))
+        } else {
+            let number_iter = descending_facet_sort(rtxn, number_db, field_id, candidates.clone())?;
+            let string_iter = descending_facet_sort(rtxn, string_db, field_id, candidates)?;
+
+            (itertools::Either::Right(number_iter), itertools::Either::Right(string_iter))
+        };
+
+        // Create builders for the next level of the tree
+        let number_iter = number_iter.map(|r| r.map(|(d, _)| d));
+        let string_iter = string_iter.map(|r| r.map(|(d, _)| d));
+        let next_children = number_iter.chain(string_iter).map(move |r| {
+            Ok(SortedDocumentsIteratorBuilder {
+                index,
+                rtxn,
+                number_db,
+                string_db,
+                fields: next_fields,
+                candidates: r?,
+                geo_candidates,
+            })
+        });
+
+        Ok(SortedDocumentsIterator::Branch {
+            current_child: None,
+            next_children_size: size,
+            next_children: Box::new(next_children),
+        })
+    }
+
+    /// Builds a [`SortedDocumentsIterator`] based on the (lazy) results of a geo sort.
+    #[allow(clippy::too_many_arguments)]
+    fn build_geo(
+        index: &'ctx crate::Index,
+        rtxn: &'ctx heed::RoTxn<'ctx>,
+        number_db: Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
+        string_db: Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
+        next_fields: &'ctx [AscDescId],
+        candidates: RoaringBitmap,
+        geo_candidates: &'ctx RoaringBitmap,
+        field_ids: [u16; 2],
+        target_point: [f64; 2],
+        ascending: bool,
+    ) -> crate::Result<SortedDocumentsIterator<'ctx>> {
+        let mut cache = VecDeque::new();
+        let mut rtree = None;
+        let size = candidates.len() as usize;
+        let not_geo_candidates = candidates.clone() - geo_candidates;
+        let mut geo_remaining = size - not_geo_candidates.len() as usize;
+        let mut not_geo_candidates = Some(not_geo_candidates);
+
+        let next_children = std::iter::from_fn(move || {
+            // Find the next bucket of geo-sorted documents.
+            // next_bucket loops and will go back to the beginning so we use a variable to track how many are left.
+            if geo_remaining > 0 {
+                if let Ok(Some((docids, _point))) = next_bucket(
+                    index,
+                    rtxn,
+                    &candidates,
+                    ascending,
+                    target_point,
+                    &Some(field_ids),
+                    &mut rtree,
+                    &mut cache,
+                    geo_candidates,
+                    GeoSortParameter::default(),
+                ) {
+                    geo_remaining -= docids.len() as usize;
+                    return Some(Ok(SortedDocumentsIteratorBuilder {
+                        index,
+                        rtxn,
+                        number_db,
+                        string_db,
+                        fields: next_fields,
+                        candidates: docids,
+                        geo_candidates,
+                    }));
+                }
+            }
+
+            // Once all geo candidates have been processed, we can return the others
+            if let Some(not_geo_candidates) = not_geo_candidates.take() {
+                if !not_geo_candidates.is_empty() {
+                    return Some(Ok(SortedDocumentsIteratorBuilder {
+                        index,
+                        rtxn,
+                        number_db,
+                        string_db,
+                        fields: next_fields,
+                        candidates: not_geo_candidates,
+                        geo_candidates,
+                    }));
+                }
+            }
+
+            None
+        });
+
+        Ok(SortedDocumentsIterator::Branch {
+            current_child: None,
+            next_children_size: size,
+            next_children: Box::new(next_children),
+        })
+    }
+}
+
+/// A structure owning the data needed during the lifetime of a [`SortedDocumentsIterator`].
+pub struct SortedDocuments<'ctx> {
+    index: &'ctx crate::Index,
+    rtxn: &'ctx heed::RoTxn<'ctx>,
+    fields: Vec<AscDescId>,
+    number_db: Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
+    string_db: Database<FacetGroupKeyCodec<BytesRefCodec>, FacetGroupValueCodec>,
+    candidates: &'ctx RoaringBitmap,
+    geo_candidates: RoaringBitmap,
+}
+
+impl<'ctx> SortedDocuments<'ctx> {
+    pub fn iter(&'ctx self) -> crate::Result<SortedDocumentsIterator<'ctx>> {
+        let builder = SortedDocumentsIteratorBuilder {
+            index: self.index,
+            rtxn: self.rtxn,
+            number_db: self.number_db,
+            string_db: self.string_db,
+            fields: &self.fields,
+            candidates: self.candidates.clone(),
+            geo_candidates: &self.geo_candidates,
+        };
+        builder.build()
+    }
+}
+
+pub fn recursive_sort<'ctx>(
+    index: &'ctx crate::Index,
+    rtxn: &'ctx heed::RoTxn<'ctx>,
+    sort: Vec<AscDesc>,
+    candidates: &'ctx RoaringBitmap,
+) -> crate::Result<SortedDocuments<'ctx>> {
+    let sortable_fields: BTreeSet<_> = index.sortable_fields(rtxn)?.into_iter().collect();
+    let fields_ids_map = index.fields_ids_map(rtxn)?;
+
+    // Retrieve the field ids that are used for sorting
+    let mut fields = Vec::new();
+    let mut need_geo_candidates = false;
+    for asc_desc in sort {
+        let (field, geofield) = match asc_desc {
+            AscDesc::Asc(Member::Field(field)) => (Some((field, true)), None),
+            AscDesc::Desc(Member::Field(field)) => (Some((field, false)), None),
+            AscDesc::Asc(Member::Geo(target_point)) => (None, Some((target_point, true))),
+            AscDesc::Desc(Member::Geo(target_point)) => (None, Some((target_point, false))),
+        };
+        if let Some((field, ascending)) = field {
+            if is_faceted(&field, &sortable_fields) {
+                if let Some(field_id) = fields_ids_map.id(&field) {
+                    fields.push(AscDescId::Facet { field_id, ascending });
+                    continue;
+                }
+            }
+            return Err(UserError::InvalidDocumentSortableAttribute {
+                field: field.to_string(),
+                sortable_fields: sortable_fields.clone(),
+            }
+            .into());
+        }
+        if let Some((target_point, ascending)) = geofield {
+            if sortable_fields.contains(RESERVED_GEO_FIELD_NAME) {
+                if let (Some(lat), Some(lng)) =
+                    (fields_ids_map.id("_geo.lat"), fields_ids_map.id("_geo.lng"))
+                {
+                    need_geo_candidates = true;
+                    fields.push(AscDescId::Geo { field_ids: [lat, lng], target_point, ascending });
+                    continue;
+                }
+            }
+            return Err(UserError::InvalidDocumentSortableAttribute {
+                field: RESERVED_GEO_FIELD_NAME.to_string(),
+                sortable_fields: sortable_fields.clone(),
+            }
+            .into());
+        }
+    }
+
+    let geo_candidates = if need_geo_candidates {
+        index.geo_faceted_documents_ids(rtxn)?
+    } else {
+        RoaringBitmap::new()
+    };
+
+    let number_db = index.facet_id_f64_docids.remap_key_type::<FacetGroupKeyCodec<BytesRefCodec>>();
+    let string_db =
+        index.facet_id_string_docids.remap_key_type::<FacetGroupKeyCodec<BytesRefCodec>>();
+
+    Ok(SortedDocuments { index, rtxn, fields, number_db, string_db, candidates, geo_candidates })
+}

crates/milli/src/error.rs

@@ -191,7 +191,21 @@ and can not be more than 511 bytes.", .document_id.to_string()
                 ),
         }
     )]
-    InvalidSortableAttribute { field: String, valid_fields: BTreeSet<String>, hidden_fields: bool },
+    InvalidSearchSortableAttribute {
+        field: String,
+        valid_fields: BTreeSet<String>,
+        hidden_fields: bool,
+    },
+    #[error("Attribute `{}` is not sortable. {}",
+        .field,
+        match .sortable_fields.is_empty() {
+            true => "This index does not have configured sortable attributes.".to_string(),
+            false => format!("Available sortable attributes are: `{}`.",
+                    sortable_fields.iter().map(AsRef::as_ref).collect::<Vec<&str>>().join(", ")
+                ),
+        }
+    )]
+    InvalidDocumentSortableAttribute { field: String, sortable_fields: BTreeSet<String> },
     #[error("Attribute `{}` is not filterable and thus, cannot be used as distinct attribute. {}",
         .field,
         match (.valid_patterns.is_empty(), .matching_rule_index) {
@@ -272,8 +286,8 @@ and can not be more than 511 bytes.", .document_id.to_string()
     PrimaryKeyCannotBeChanged(String),
     #[error(transparent)]
     SerdeJson(serde_json::Error),
-    #[error(transparent)]
-    SortError(#[from] SortError),
+    #[error("{error}")]
+    SortError { error: SortError, search: bool },
     #[error("An unknown internal document id have been used: `{document_id}`.")]
     UnknownInternalDocumentId { document_id: DocumentId },
     #[error("`minWordSizeForTypos` setting is invalid. `oneTypo` and `twoTypos` fields should be between `0` and `255`, and `twoTypos` should be greater or equals to `oneTypo` but found `oneTypo: {0}` and twoTypos: {1}`.")]
@@ -616,7 +630,7 @@ fn conditionally_lookup_for_error_message() {
     ];
 
     for (list, suffix) in messages {
-        let err = UserError::InvalidSortableAttribute {
+        let err = UserError::InvalidSearchSortableAttribute {
             field: "name".to_string(),
             valid_fields: list,
             hidden_fields: false,

crates/milli/src/filterable_attributes_rules.rs

@@ -111,7 +111,7 @@ impl FilterableAttributesFeatures {
         self.filter.is_filterable_null()
     }
 
-    /// Check if `IS EXISTS` is allowed
+    /// Check if `EXISTS` is allowed
     pub fn is_filterable_exists(&self) -> bool {
         self.filter.is_filterable_exists()
     }

crates/milli/src/lib.rs

@@ -43,12 +43,13 @@ use std::fmt;
 use std::hash::BuildHasherDefault;
 
 use charabia::normalizer::{CharNormalizer, CompatibilityDecompositionNormalizer};
+pub use documents::GeoSortStrategy;
 pub use filter_parser::{Condition, FilterCondition, Span, Token};
 use fxhash::{FxHasher32, FxHasher64};
 pub use grenad::CompressionType;
 pub use search::new::{
-    execute_search, filtered_universe, DefaultSearchLogger, GeoSortStrategy, SearchContext,
-    SearchLogger, VisualSearchLogger,
+    execute_search, filtered_universe, DefaultSearchLogger, SearchContext, SearchLogger,
+    VisualSearchLogger,
 };
 use serde_json::Value;
 pub use thread_pool_no_abort::{PanicCatched, ThreadPoolNoAbort, ThreadPoolNoAbortBuilder};

crates/milli/src/search/facet/filter.rs

@@ -10,7 +10,7 @@ use memchr::memmem::Finder;
 use roaring::{MultiOps, RoaringBitmap};
 use serde_json::Value;
 
-use super::facet_range_search;
+use super::{facet_range_search, filter_vector::VectorFilter};
 use crate::constants::RESERVED_GEO_FIELD_NAME;
 use crate::error::{Error, UserError};
 use crate::filterable_attributes_rules::{filtered_matching_patterns, matching_features};
@@ -228,6 +228,10 @@ impl<'a> Filter<'a> {
     pub fn use_contains_operator(&self) -> Option<&Token> {
         self.condition.use_contains_operator()
     }
+
+    pub fn use_vector_filter(&self) -> Option<&Token> {
+        self.condition.use_vector_filter()
+    }
 }
 
 impl<'a> Filter<'a> {
@@ -235,10 +239,12 @@ impl<'a> Filter<'a> {
         // to avoid doing this for each recursive call we're going to do it ONCE ahead of time
         let fields_ids_map = index.fields_ids_map(rtxn)?;
         let filterable_attributes_rules = index.filterable_attributes_rules(rtxn)?;
+
         for fid in self.condition.fids(MAX_FILTER_DEPTH) {
             let attribute = fid.value();
             if matching_features(attribute, &filterable_attributes_rules)
                 .is_some_and(|(_, features)| features.is_filterable())
+                || VectorFilter::matches(attribute)
             {
                 continue;
             }
@@ -542,7 +548,18 @@ impl<'a> Filter<'a> {
                     .union()
             }
             FilterCondition::Condition { fid, op } => {
-                let Some(field_id) = field_ids_map.id(fid.value()) else {
+                let value = fid.value();
+                if VectorFilter::matches(value) {
+                    if !matches!(op, Condition::Exists) {
+                        return Err(Error::UserError(UserError::InvalidFilter(String::from(
+                            "Vector filter can only be used with the `exists` operator",
+                        ))));
+                    }
+                    let vector_filter = VectorFilter::parse(fid)?;
+                    return vector_filter.evaluate(rtxn, index, universe);
+                }
+
+                let Some(field_id) = field_ids_map.id(value) else {
                     return Ok(RoaringBitmap::new());
                 };
                 let Some((rule_index, features)) =

crates/milli/src/search/facet/filter_vector.rs

@@ -0,0 +1,276 @@
+use filter_parser::Token;
+use roaring::{MultiOps, RoaringBitmap};
+
+use crate::error::{Error, UserError};
+use crate::vector::db::IndexEmbeddingConfig;
+use crate::vector::{ArroyStats, ArroyWrapper};
+use crate::Index;
+
+#[derive(Debug)]
+enum VectorFilterInner<'a> {
+    Fragment { embedder_token: Token<'a>, fragment_token: Token<'a> },
+    DocumentTemplate { embedder_token: Token<'a> },
+    UserProvided { embedder_token: Token<'a> },
+    FullEmbedder { embedder_token: Token<'a> },
+}
+
+#[derive(Debug)]
+pub(super) struct VectorFilter<'a> {
+    inner: Option<VectorFilterInner<'a>>,
+    regenerate: bool,
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum VectorFilterError<'a> {
+    #[error("Vector filter cannot be empty.")]
+    EmptyFilter,
+
+    #[error("Vector filter must start with `_vectors` but found `{}`.", _0.value())]
+    InvalidPrefix(Token<'a>),
+
+    #[error("Vector filter is inconsistent: either specify a fragment name or remove the `fragments` part.")]
+    MissingFragmentName(Token<'a>),
+
+    #[error("Vector filter cannot have both `{}` and `{}`.", _0.0.value(), _0.1.value())]
+    ExclusiveOptions(Box<(Token<'a>, Token<'a>)>),
+
+    #[error("Vector filter has leftover token: `{}`.", _0.value())]
+    LeftoverToken(Token<'a>),
+
+    #[error("The embedder `{}` does not exist. {}", embedder.value(), {
+        if available.is_empty() {
+            String::from("This index does not have any configured embedders.")
+        } else {
+            let mut available = available.clone();
+            available.sort_unstable();
+            format!("Available embedders are: {}.", available.iter().map(|e| format!("`{e}`")).collect::<Vec<_>>().join(", "))
+        }
+    })]
+    EmbedderDoesNotExist { embedder: &'a Token<'a>, available: Vec<String> },
+
+    #[error("The fragment `{}` does not exist on embedder `{}`. {}", fragment.value(), embedder.value(), {
+        if available.is_empty() {
+            String::from("This embedder does not have any configured fragments.")
+        } else {
+            let mut available = available.clone();
+            available.sort_unstable();
+            format!("Available fragments on this embedder are: {}.", available.iter().map(|f| format!("`{f}`")).collect::<Vec<_>>().join(", "))
+        }
+    })]
+    FragmentDoesNotExist {
+        embedder: &'a Token<'a>,
+        fragment: &'a Token<'a>,
+        available: Vec<String>,
+    },
+}
+
+use VectorFilterError::*;
+
+impl<'a> From<VectorFilterError<'a>> for Error {
+    fn from(err: VectorFilterError<'a>) -> Self {
+        match &err {
+            EmptyFilter => Error::UserError(UserError::InvalidFilter(err.to_string())),
+            InvalidPrefix(token) | MissingFragmentName(token) | LeftoverToken(token) => {
+                token.clone().as_external_error(err).into()
+            }
+            ExclusiveOptions(tokens) => tokens.1.clone().as_external_error(err).into(),
+            EmbedderDoesNotExist { embedder: token, .. }
+            | FragmentDoesNotExist { fragment: token, .. } => token.as_external_error(err).into(),
+        }
+    }
+}
+
+impl<'a> VectorFilter<'a> {
+    pub(super) fn matches(value: &str) -> bool {
+        value.starts_with("_vectors.") || value == "_vectors"
+    }
+
+    /// Parses a vector filter string.
+    ///
+    /// Valid formats:
+    /// - `_vectors`
+    /// - `_vectors.{embedder_name}`
+    /// - `_vectors.{embedder_name}.regenerate`
+    /// - `_vectors.{embedder_name}.userProvided`
+    /// - `_vectors.{embedder_name}.userProvided.regenerate`
+    /// - `_vectors.{embedder_name}.documentTemplate`
+    /// - `_vectors.{embedder_name}.documentTemplate.regenerate`
+    /// - `_vectors.{embedder_name}.fragments.{fragment_name}`
+    /// - `_vectors.{embedder_name}.fragments.{fragment_name}.regenerate`
+    pub(super) fn parse(s: &'a Token<'a>) -> Result<Self, VectorFilterError<'a>> {
+        let mut split = s.split(".").peekable();
+
+        match split.next() {
+            Some(token) if token.value() == "_vectors" => (),
+            Some(token) => return Err(InvalidPrefix(token)),
+            None => return Err(EmptyFilter),
+        }
+
+        let embedder_name = split.next();
+
+        let mut fragment_name = None;
+        if split.peek().map(|t| t.value()) == Some("fragments") {
+            let token = split.next().expect("it was peeked before");
+
+            fragment_name = Some(split.next().ok_or(MissingFragmentName(token))?);
+        }
+
+        let mut user_provided_token = None;
+        if split.peek().map(|t| t.value()) == Some("userProvided") {
+            user_provided_token = split.next();
+        }
+
+        let mut document_template_token = None;
+        if split.peek().map(|t| t.value()) == Some("documentTemplate") {
+            document_template_token = split.next();
+        }
+
+        let mut regenerate_token = None;
+        if split.peek().map(|t| t.value()) == Some("regenerate") {
+            regenerate_token = split.next();
+        }
+
+        let inner = match (fragment_name, user_provided_token, document_template_token) {
+            (Some(fragment_name), None, None) => Some(VectorFilterInner::Fragment {
+                embedder_token: embedder_name
+                    .expect("embedder name comes before fragment so it's always Some"),
+                fragment_token: fragment_name,
+            }),
+            (None, Some(_), None) => Some(VectorFilterInner::UserProvided {
+                embedder_token: embedder_name
+                    .expect("embedder name comes before userProvided so it's always Some"),
+            }),
+            (None, None, Some(_)) => Some(VectorFilterInner::DocumentTemplate {
+                embedder_token: embedder_name
+                    .expect("embedder name comes before documentTemplate so it's always Some"),
+            }),
+            (Some(a), Some(b), _) | (_, Some(a), Some(b)) | (Some(a), None, Some(b)) => {
+                return Err(ExclusiveOptions(Box::new((a, b))));
+            }
+            (None, None, None) => embedder_name
+                .map(|embedder_token| VectorFilterInner::FullEmbedder { embedder_token }),
+        };
+
+        if let Some(next) = split.next() {
+            return Err(LeftoverToken(next))?;
+        }
+
+        Ok(Self { inner, regenerate: regenerate_token.is_some() })
+    }
+
+    pub(super) fn evaluate(
+        self,
+        rtxn: &heed::RoTxn<'_>,
+        index: &Index,
+        universe: Option<&RoaringBitmap>,
+    ) -> crate::Result<RoaringBitmap> {
+        let index_embedding_configs = index.embedding_configs();
+        let embedding_configs = index_embedding_configs.embedding_configs(rtxn)?;
+
+        let inners = match self.inner {
+            Some(inner) => vec![inner],
+            None => embedding_configs
+                .iter()
+                .map(|config| VectorFilterInner::FullEmbedder {
+                    embedder_token: Token::from(config.name.as_str()),
+                })
+                .collect(),
+        };
+
+        let mut docids = inners
+            .iter()
+            .map(|i| i.evaluate_inner(rtxn, index, &embedding_configs, self.regenerate))
+            .union()?;
+
+        if let Some(universe) = universe {
+            docids &= universe;
+        }
+
+        Ok(docids)
+    }
+}
+
+impl VectorFilterInner<'_> {
+    fn evaluate_inner(
+        &self,
+        rtxn: &heed::RoTxn<'_>,
+        index: &Index,
+        embedding_configs: &[IndexEmbeddingConfig],
+        regenerate: bool,
+    ) -> crate::Result<RoaringBitmap> {
+        let embedder = match self {
+            VectorFilterInner::Fragment { embedder_token, .. } => embedder_token,
+            VectorFilterInner::DocumentTemplate { embedder_token } => embedder_token,
+            VectorFilterInner::UserProvided { embedder_token } => embedder_token,
+            VectorFilterInner::FullEmbedder { embedder_token } => embedder_token,
+        };
+        let embedder_name = embedder.value();
+        let available_embedders =
+            || embedding_configs.iter().map(|c| c.name.clone()).collect::<Vec<_>>();
+
+        let embedding_config = embedding_configs
+            .iter()
+            .find(|config| config.name == embedder_name)
+            .ok_or_else(|| EmbedderDoesNotExist { embedder, available: available_embedders() })?;
+
+        let embedder_info = index
+            .embedding_configs()
+            .embedder_info(rtxn, embedder_name)?
+            .ok_or_else(|| EmbedderDoesNotExist { embedder, available: available_embedders() })?;
+
+        let arroy_wrapper = ArroyWrapper::new(
+            index.vector_arroy,
+            embedder_info.embedder_id,
+            embedding_config.config.quantized(),
+        );
+
+        let mut docids = match self {
+            VectorFilterInner::Fragment { embedder_token: embedder, fragment_token: fragment } => {
+                let fragment_name = fragment.value();
+                let fragment_config = embedding_config
+                    .fragments
+                    .as_slice()
+                    .iter()
+                    .find(|fragment| fragment.name == fragment_name)
+                    .ok_or_else(|| FragmentDoesNotExist {
+                        embedder,
+                        fragment,
+                        available: embedding_config
+                            .fragments
+                            .as_slice()
+                            .iter()
+                            .map(|f| f.name.clone())
+                            .collect(),
+                    })?;
+
+                arroy_wrapper.items_in_store(rtxn, fragment_config.id, |bitmap| bitmap.clone())?
+            }
+            VectorFilterInner::DocumentTemplate { .. } => {
+                if !embedding_config.fragments.as_slice().is_empty() {
+                    return Ok(RoaringBitmap::new());
+                }
+
+                let user_provided_docsids = embedder_info.embedding_status.user_provided_docids();
+                let mut stats = ArroyStats::default();
+                arroy_wrapper.aggregate_stats(rtxn, &mut stats)?;
+                stats.documents - user_provided_docsids.clone()
+            }
+            VectorFilterInner::UserProvided { .. } => {
+                let user_provided_docsids = embedder_info.embedding_status.user_provided_docids();
+                user_provided_docsids.clone()
+            }
+            VectorFilterInner::FullEmbedder { .. } => {
+                let mut stats = ArroyStats::default();
+                arroy_wrapper.aggregate_stats(rtxn, &mut stats)?;
+                stats.documents
+            }
+        };
+
+        if regenerate {
+            let skip_regenerate = embedder_info.embedding_status.skip_regenerate_docids();
+            docids -= skip_regenerate;
+        }
+
+        Ok(docids)
+    }
+}

crates/milli/src/search/facet/mod.rs

@@ -17,6 +17,7 @@ mod facet_range_search;
 mod facet_sort_ascending;
 mod facet_sort_descending;
 mod filter;
+mod filter_vector;
 mod search;
 
 fn facet_extreme_value<'t>(

crates/milli/src/search/hybrid.rs

@@ -210,6 +210,7 @@ impl Search<'_> {
             scoring_strategy: ScoringStrategy::Detailed,
             words_limit: self.words_limit,
             exhaustive_number_hits: self.exhaustive_number_hits,
+            max_total_hits: self.max_total_hits,
             rtxn: self.rtxn,
             index: self.index,
             semantic: self.semantic.clone(),

crates/milli/src/search/mod.rs

@@ -9,6 +9,7 @@ use roaring::bitmap::RoaringBitmap;
 pub use self::facet::{FacetDistribution, Filter, OrderBy, DEFAULT_VALUES_PER_FACET};
 pub use self::new::matches::{FormatOptions, MatchBounds, MatcherBuilder, MatchingWords};
 use self::new::{execute_vector_search, PartialSearchResult, VectorStoreStats};
+use crate::documents::GeoSortParameter;
 use crate::filterable_attributes_rules::{filtered_matching_patterns, matching_features};
 use crate::index::MatchingStrategy;
 use crate::score_details::{ScoreDetails, ScoringStrategy};
@@ -47,11 +48,12 @@ pub struct Search<'a> {
     sort_criteria: Option<Vec<AscDesc>>,
     distinct: Option<String>,
     searchable_attributes: Option<&'a [String]>,
-    geo_param: new::GeoSortParameter,
+    geo_param: GeoSortParameter,
     terms_matching_strategy: TermsMatchingStrategy,
     scoring_strategy: ScoringStrategy,
     words_limit: usize,
     exhaustive_number_hits: bool,
+    max_total_hits: Option<usize>,
     rtxn: &'a heed::RoTxn<'a>,
     index: &'a Index,
     semantic: Option<SemanticSearch>,
@@ -70,10 +72,11 @@ impl<'a> Search<'a> {
             sort_criteria: None,
             distinct: None,
             searchable_attributes: None,
-            geo_param: new::GeoSortParameter::default(),
+            geo_param: GeoSortParameter::default(),
             terms_matching_strategy: TermsMatchingStrategy::default(),
             scoring_strategy: Default::default(),
             exhaustive_number_hits: false,
+            max_total_hits: None,
             words_limit: 10,
             rtxn,
             index,
@@ -147,7 +150,7 @@ impl<'a> Search<'a> {
     }
 
     #[cfg(test)]
-    pub fn geo_sort_strategy(&mut self, strategy: new::GeoSortStrategy) -> &mut Search<'a> {
+    pub fn geo_sort_strategy(&mut self, strategy: crate::GeoSortStrategy) -> &mut Search<'a> {
         self.geo_param.strategy = strategy;
         self
     }
@@ -165,6 +168,11 @@ impl<'a> Search<'a> {
         self
     }
 
+    pub fn max_total_hits(&mut self, max_total_hits: Option<usize>) -> &mut Search<'a> {
+        self.max_total_hits = max_total_hits;
+        self
+    }
+
     pub fn time_budget(&mut self, time_budget: TimeBudget) -> &mut Search<'a> {
         self.time_budget = time_budget;
         self
@@ -243,6 +251,8 @@ impl<'a> Search<'a> {
                 &mut ctx,
                 vector,
                 self.scoring_strategy,
+                self.exhaustive_number_hits,
+                self.max_total_hits,
                 universe,
                 &self.sort_criteria,
                 &self.distinct,
@@ -261,6 +271,7 @@ impl<'a> Search<'a> {
                 self.terms_matching_strategy,
                 self.scoring_strategy,
                 self.exhaustive_number_hits,
+                self.max_total_hits,
                 universe,
                 &self.sort_criteria,
                 &self.distinct,
@@ -314,6 +325,7 @@ impl fmt::Debug for Search<'_> {
             scoring_strategy,
             words_limit,
             exhaustive_number_hits,
+            max_total_hits,
             rtxn: _,
             index: _,
             semantic,
@@ -333,6 +345,7 @@ impl fmt::Debug for Search<'_> {
             .field("terms_matching_strategy", terms_matching_strategy)
             .field("scoring_strategy", scoring_strategy)
             .field("exhaustive_number_hits", exhaustive_number_hits)
+            .field("max_total_hits", max_total_hits)
             .field("words_limit", words_limit)
             .field(
                 "semantic.embedder_name",

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

@@ -32,6 +32,8 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
     logger: &mut dyn SearchLogger<Q>,
     time_budget: TimeBudget,
     ranking_score_threshold: Option<f64>,
+    exhaustive_number_hits: bool,
+    max_total_hits: Option<usize>,
 ) -> Result<BucketSortOutput> {
     logger.initial_query(query);
     logger.ranking_rules(&ranking_rules);
@@ -159,7 +161,13 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
         };
     }
 
-    while valid_docids.len() < length {
+    let max_len_to_evaluate =
+        match (max_total_hits, exhaustive_number_hits && ranking_score_threshold.is_some()) {
+            (Some(max_total_hits), true) => max_total_hits,
+            _ => length,
+        };
+
+    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]);

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

@@ -82,7 +82,7 @@ fn facet_value_docids(
 }
 
 /// Return an iterator over each number value in the given field of the given document.
-fn facet_number_values<'a>(
+pub(crate) fn facet_number_values<'a>(
     docid: u32,
     field_id: u16,
     index: &Index,
@@ -118,7 +118,7 @@ pub fn facet_string_values<'a>(
 }
 
 #[allow(clippy::drop_non_drop)]
-fn facet_values_prefix_key(distinct: u16, id: u32) -> [u8; FID_SIZE + DOCID_SIZE] {
+pub(crate) fn facet_values_prefix_key(distinct: u16, id: u32) -> [u8; FID_SIZE + DOCID_SIZE] {
     concat_arrays::concat_arrays!(distinct.to_be_bytes(), id.to_be_bytes())
 }
 

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

@@ -1,96 +1,18 @@
 use std::collections::VecDeque;
 
-use heed::types::{Bytes, Unit};
-use heed::{RoPrefix, RoTxn};
 use roaring::RoaringBitmap;
 use rstar::RTree;
 
-use super::facet_string_values;
 use super::ranking_rules::{RankingRule, RankingRuleOutput, RankingRuleQueryTrait};
-use crate::heed_codec::facet::{FieldDocIdFacetCodec, OrderedF64Codec};
+use crate::documents::geo_sort::{fill_cache, next_bucket};
+use crate::documents::{GeoSortParameter, GeoSortStrategy};
 use crate::score_details::{self, ScoreDetails};
-use crate::{
-    distance_between_two_points, lat_lng_to_xyz, GeoPoint, Index, Result, SearchContext,
-    SearchLogger,
-};
-
-const FID_SIZE: usize = 2;
-const DOCID_SIZE: usize = 4;
-
-#[allow(clippy::drop_non_drop)]
-fn facet_values_prefix_key(distinct: u16, id: u32) -> [u8; FID_SIZE + DOCID_SIZE] {
-    concat_arrays::concat_arrays!(distinct.to_be_bytes(), id.to_be_bytes())
-}
-
-/// Return an iterator over each number value in the given field of the given document.
-fn facet_number_values<'a>(
-    docid: u32,
-    field_id: u16,
-    index: &Index,
-    txn: &'a RoTxn<'a>,
-) -> Result<RoPrefix<'a, FieldDocIdFacetCodec<OrderedF64Codec>, Unit>> {
-    let key = facet_values_prefix_key(field_id, docid);
-
-    let iter = index
-        .field_id_docid_facet_f64s
-        .remap_key_type::<Bytes>()
-        .prefix_iter(txn, &key)?
-        .remap_key_type();
-
-    Ok(iter)
-}
-
-#[derive(Debug, Clone, Copy)]
-pub struct Parameter {
-    // Define the strategy used by the geo sort
-    pub strategy: Strategy,
-    // Limit the number of docs in a single bucket to avoid unexpectedly large overhead
-    pub max_bucket_size: u64,
-    // Considering the errors of GPS and geographical calculations, distances less than distance_error_margin will be treated as equal
-    pub distance_error_margin: f64,
-}
-
-impl Default for Parameter {
-    fn default() -> Self {
-        Self { strategy: Strategy::default(), max_bucket_size: 1000, distance_error_margin: 1.0 }
-    }
-}
-/// Define the strategy used by the geo sort.
-/// The parameter represents the cache size, and, in the case of the Dynamic strategy,
-/// the point where we move from using the iterative strategy to the rtree.
-#[derive(Debug, Clone, Copy)]
-pub enum Strategy {
-    AlwaysIterative(usize),
-    AlwaysRtree(usize),
-    Dynamic(usize),
-}
-
-impl Default for Strategy {
-    fn default() -> Self {
-        Strategy::Dynamic(1000)
-    }
-}
-
-impl Strategy {
-    pub fn use_rtree(&self, candidates: usize) -> bool {
-        match self {
-            Strategy::AlwaysIterative(_) => false,
-            Strategy::AlwaysRtree(_) => true,
-            Strategy::Dynamic(i) => candidates >= *i,
-        }
-    }
-
-    pub fn cache_size(&self) -> usize {
-        match self {
-            Strategy::AlwaysIterative(i) | Strategy::AlwaysRtree(i) | Strategy::Dynamic(i) => *i,
-        }
-    }
-}
+use crate::{GeoPoint, Result, SearchContext, SearchLogger};
 
 pub struct GeoSort<Q: RankingRuleQueryTrait> {
     query: Option<Q>,
 
-    strategy: Strategy,
+    strategy: GeoSortStrategy,
     ascending: bool,
     point: [f64; 2],
     field_ids: Option<[u16; 2]>,
@@ -107,12 +29,12 @@ pub struct GeoSort<Q: RankingRuleQueryTrait> {
 
 impl<Q: RankingRuleQueryTrait> GeoSort<Q> {
     pub fn new(
-        parameter: Parameter,
+        parameter: GeoSortParameter,
         geo_faceted_docids: RoaringBitmap,
         point: [f64; 2],
         ascending: bool,
     ) -> Result<Self> {
-        let Parameter { strategy, max_bucket_size, distance_error_margin } = parameter;
+        let GeoSortParameter { strategy, max_bucket_size, distance_error_margin } = parameter;
         Ok(Self {
             query: None,
             strategy,
@@ -134,98 +56,22 @@ impl<Q: RankingRuleQueryTrait> GeoSort<Q> {
         ctx: &mut SearchContext<'_>,
         geo_candidates: &RoaringBitmap,
     ) -> Result<()> {
-        debug_assert!(self.field_ids.is_some(), "fill_buffer can't be called without the lat&lng");
-        debug_assert!(self.cached_sorted_docids.is_empty());
-
-        // lazily initialize the rtree if needed by the strategy, and cache it in `self.rtree`
-        let rtree = if self.strategy.use_rtree(geo_candidates.len() as usize) {
-            if let Some(rtree) = self.rtree.as_ref() {
-                // get rtree from cache
-                Some(rtree)
-            } else {
-                let rtree = ctx.index.geo_rtree(ctx.txn)?.expect("geo candidates but no rtree");
-                // insert rtree in cache and returns it.
-                // Can't use `get_or_insert_with` because getting the rtree from the DB is a fallible operation.
-                Some(&*self.rtree.insert(rtree))
-            }
-        } else {
-            None
-        };
-
-        let cache_size = self.strategy.cache_size();
-        if let Some(rtree) = rtree {
-            if self.ascending {
-                let point = lat_lng_to_xyz(&self.point);
-                for point in rtree.nearest_neighbor_iter(&point) {
-                    if geo_candidates.contains(point.data.0) {
-                        self.cached_sorted_docids.push_back(point.data);
-                        if self.cached_sorted_docids.len() >= cache_size {
-                            break;
-                        }
-                    }
-                }
-            } else {
-                // in the case of the desc geo sort we look for the closest point to the opposite of the queried point
-                // and we insert the points in reverse order they get reversed when emptying the cache later on
-                let point = lat_lng_to_xyz(&opposite_of(self.point));
-                for point in rtree.nearest_neighbor_iter(&point) {
-                    if geo_candidates.contains(point.data.0) {
-                        self.cached_sorted_docids.push_front(point.data);
-                        if self.cached_sorted_docids.len() >= cache_size {
-                            break;
-                        }
-                    }
-                }
-            }
-        } else {
-            // the iterative version
-            let [lat, lng] = self.field_ids.unwrap();
-
-            let mut documents = geo_candidates
-                .iter()
-                .map(|id| -> Result<_> { Ok((id, geo_value(id, lat, lng, ctx.index, ctx.txn)?)) })
-                .collect::<Result<Vec<(u32, [f64; 2])>>>()?;
-            // computing the distance between two points is expensive thus we cache the result
-            documents
-                .sort_by_cached_key(|(_, p)| distance_between_two_points(&self.point, p) as usize);
-            self.cached_sorted_docids.extend(documents);
-        };
+        fill_cache(
+            ctx.index,
+            ctx.txn,
+            self.strategy,
+            self.ascending,
+            self.point,
+            &self.field_ids,
+            &mut self.rtree,
+            geo_candidates,
+            &mut self.cached_sorted_docids,
+        )?;
 
         Ok(())
     }
 }
 
-/// Extracts the lat and long values from a single document.
-///
-/// If it is not able to find it in the facet number index it will extract it
-/// from the facet string index and parse it as f64 (as the geo extraction behaves).
-fn geo_value(
-    docid: u32,
-    field_lat: u16,
-    field_lng: u16,
-    index: &Index,
-    rtxn: &RoTxn<'_>,
-) -> Result<[f64; 2]> {
-    let extract_geo = |geo_field: u16| -> Result<f64> {
-        match facet_number_values(docid, geo_field, index, rtxn)?.next() {
-            Some(Ok(((_, _, geo), ()))) => Ok(geo),
-            Some(Err(e)) => Err(e.into()),
-            None => match facet_string_values(docid, geo_field, index, rtxn)?.next() {
-                Some(Ok((_, geo))) => {
-                    Ok(geo.parse::<f64>().expect("cannot parse geo field as f64"))
-                }
-                Some(Err(e)) => Err(e.into()),
-                None => panic!("A geo faceted document doesn't contain any lat or lng"),
-            },
-        }
-    };
-
-    let lat = extract_geo(field_lat)?;
-    let lng = extract_geo(field_lng)?;
-
-    Ok([lat, lng])
-}
-
 impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for GeoSort<Q> {
     fn id(&self) -> String {
         "geo_sort".to_owned()
@@ -267,124 +113,33 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for GeoSort<Q> {
     ) -> Result<Option<RankingRuleOutput<Q>>> {
         let query = self.query.as_ref().unwrap().clone();
 
-        let mut geo_candidates = &self.geo_candidates & universe;
-
-        if geo_candidates.is_empty() {
-            return Ok(Some(RankingRuleOutput {
+        next_bucket(
+            ctx.index,
+            ctx.txn,
+            universe,
+            self.ascending,
+            self.point,
+            &self.field_ids,
+            &mut self.rtree,
+            &mut self.cached_sorted_docids,
+            &self.geo_candidates,
+            GeoSortParameter {
+                strategy: self.strategy,
+                max_bucket_size: self.max_bucket_size,
+                distance_error_margin: self.distance_error_margin,
+            },
+        )
+        .map(|o| {
+            o.map(|(candidates, point)| RankingRuleOutput {
                 query,
-                candidates: universe.clone(),
+                candidates,
                 score: ScoreDetails::GeoSort(score_details::GeoSort {
                     target_point: self.point,
                     ascending: self.ascending,
-                    value: None,
+                    value: point,
                 }),
-            }));
-        }
-
-        let ascending = self.ascending;
-        let next = |cache: &mut VecDeque<_>| {
-            if ascending {
-                cache.pop_front()
-            } else {
-                cache.pop_back()
-            }
-        };
-        let put_back = |cache: &mut VecDeque<_>, x: _| {
-            if ascending {
-                cache.push_front(x)
-            } else {
-                cache.push_back(x)
-            }
-        };
-
-        let mut current_bucket = RoaringBitmap::new();
-        // current_distance stores the first point and distance in current bucket
-        let mut current_distance: Option<([f64; 2], f64)> = None;
-        loop {
-            // The loop will only exit when we have found all points with equal distance or have exhausted the candidates.
-            if let Some((id, point)) = next(&mut self.cached_sorted_docids) {
-                if geo_candidates.contains(id) {
-                    let distance = distance_between_two_points(&self.point, &point);
-                    if let Some((point0, bucket_distance)) = current_distance.as_ref() {
-                        if (bucket_distance - distance).abs() > self.distance_error_margin {
-                            // different distance, point belongs to next bucket
-                            put_back(&mut self.cached_sorted_docids, (id, point));
-                            return Ok(Some(RankingRuleOutput {
-                                query,
-                                candidates: current_bucket,
-                                score: ScoreDetails::GeoSort(score_details::GeoSort {
-                                    target_point: self.point,
-                                    ascending: self.ascending,
-                                    value: Some(point0.to_owned()),
-                                }),
-                            }));
-                        } else {
-                            // same distance, point belongs to current bucket
-                            current_bucket.insert(id);
-                            // remove from cadidates to prevent it from being added to the cache again
-                            geo_candidates.remove(id);
-                            // current bucket size reaches limit, force return
-                            if current_bucket.len() == self.max_bucket_size {
-                                return Ok(Some(RankingRuleOutput {
-                                    query,
-                                    candidates: current_bucket,
-                                    score: ScoreDetails::GeoSort(score_details::GeoSort {
-                                        target_point: self.point,
-                                        ascending: self.ascending,
-                                        value: Some(point0.to_owned()),
-                                    }),
-                                }));
-                            }
-                        }
-                    } else {
-                        // first doc in current bucket
-                        current_distance = Some((point, distance));
-                        current_bucket.insert(id);
-                        geo_candidates.remove(id);
-                        // current bucket size reaches limit, force return
-                        if current_bucket.len() == self.max_bucket_size {
-                            return Ok(Some(RankingRuleOutput {
-                                query,
-                                candidates: current_bucket,
-                                score: ScoreDetails::GeoSort(score_details::GeoSort {
-                                    target_point: self.point,
-                                    ascending: self.ascending,
-                                    value: Some(point.to_owned()),
-                                }),
-                            }));
-                        }
-                    }
-                }
-            } else {
-                // cache exhausted, we need to refill it
-                self.fill_buffer(ctx, &geo_candidates)?;
-
-                if self.cached_sorted_docids.is_empty() {
-                    // candidates exhausted, exit
-                    if let Some((point0, _)) = current_distance.as_ref() {
-                        return Ok(Some(RankingRuleOutput {
-                            query,
-                            candidates: current_bucket,
-                            score: ScoreDetails::GeoSort(score_details::GeoSort {
-                                target_point: self.point,
-                                ascending: self.ascending,
-                                value: Some(point0.to_owned()),
-                            }),
-                        }));
-                    } else {
-                        return Ok(Some(RankingRuleOutput {
-                            query,
-                            candidates: universe.clone(),
-                            score: ScoreDetails::GeoSort(score_details::GeoSort {
-                                target_point: self.point,
-                                ascending: self.ascending,
-                                value: None,
-                            }),
-                        }));
-                    }
-                }
-            }
-        }
+            })
+        })
     }
 
     #[tracing::instrument(level = "trace", skip_all, target = "search::geo_sort")]
@@ -394,16 +149,3 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for GeoSort<Q> {
         self.cached_sorted_docids.clear();
     }
 }
-
-/// Compute the antipodal coordinate of `coord`
-fn opposite_of(mut coord: [f64; 2]) -> [f64; 2] {
-    coord[0] *= -1.;
-    // in the case of x,0 we want to return x,180
-    if coord[1] > 0. {
-        coord[1] -= 180.;
-    } else {
-        coord[1] += 180.;
-    }
-
-    coord
-}

crates/milli/src/search/new/matches/mod.rs

@@ -510,6 +510,7 @@ mod tests {
                 crate::TermsMatchingStrategy::default(),
                 crate::score_details::ScoringStrategy::Skip,
                 false,
+                None,
                 universe,
                 &None,
                 &None,

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

@@ -1,7 +1,7 @@
 mod bucket_sort;
 mod db_cache;
 mod distinct;
-mod geo_sort;
+pub(crate) mod geo_sort;
 mod graph_based_ranking_rule;
 mod interner;
 mod limits;
@@ -46,14 +46,14 @@ use resolve_query_graph::{compute_query_graph_docids, PhraseDocIdsCache};
 use roaring::RoaringBitmap;
 use sort::Sort;
 
-use self::distinct::facet_string_values;
+pub(crate) use self::distinct::{facet_string_values, facet_values_prefix_key};
 use self::geo_sort::GeoSort;
-pub use self::geo_sort::{Parameter as GeoSortParameter, Strategy as GeoSortStrategy};
 use self::graph_based_ranking_rule::Words;
 use self::interner::Interned;
 use self::vector_sort::VectorSort;
 use crate::attribute_patterns::{match_pattern, PatternMatch};
 use crate::constants::RESERVED_GEO_FIELD_NAME;
+use crate::documents::GeoSortParameter;
 use crate::index::PrefixSearch;
 use crate::localized_attributes_rules::LocalizedFieldIds;
 use crate::score_details::{ScoreDetails, ScoringStrategy};
@@ -319,7 +319,7 @@ fn resolve_negative_phrases(
 fn get_ranking_rules_for_placeholder_search<'ctx>(
     ctx: &SearchContext<'ctx>,
     sort_criteria: &Option<Vec<AscDesc>>,
-    geo_param: geo_sort::Parameter,
+    geo_param: GeoSortParameter,
 ) -> Result<Vec<BoxRankingRule<'ctx, PlaceholderQuery>>> {
     let mut sort = false;
     let mut sorted_fields = HashSet::new();
@@ -371,7 +371,7 @@ fn get_ranking_rules_for_placeholder_search<'ctx>(
 fn get_ranking_rules_for_vector<'ctx>(
     ctx: &SearchContext<'ctx>,
     sort_criteria: &Option<Vec<AscDesc>>,
-    geo_param: geo_sort::Parameter,
+    geo_param: GeoSortParameter,
     limit_plus_offset: usize,
     target: &[f32],
     embedder_name: &str,
@@ -448,7 +448,7 @@ fn get_ranking_rules_for_vector<'ctx>(
 fn get_ranking_rules_for_query_graph_search<'ctx>(
     ctx: &SearchContext<'ctx>,
     sort_criteria: &Option<Vec<AscDesc>>,
-    geo_param: geo_sort::Parameter,
+    geo_param: GeoSortParameter,
     terms_matching_strategy: TermsMatchingStrategy,
 ) -> Result<Vec<BoxRankingRule<'ctx, QueryGraph>>> {
     // query graph search
@@ -559,7 +559,7 @@ fn resolve_sort_criteria<'ctx, Query: RankingRuleQueryTrait>(
     ranking_rules: &mut Vec<BoxRankingRule<'ctx, Query>>,
     sorted_fields: &mut HashSet<String>,
     geo_sorted: &mut bool,
-    geo_param: geo_sort::Parameter,
+    geo_param: GeoSortParameter,
 ) -> Result<()> {
     let sort_criteria = sort_criteria.clone().unwrap_or_default();
     ranking_rules.reserve(sort_criteria.len());
@@ -626,10 +626,12 @@ pub fn execute_vector_search(
     ctx: &mut SearchContext<'_>,
     vector: &[f32],
     scoring_strategy: ScoringStrategy,
+    exhaustive_number_hits: bool,
+    max_total_hits: Option<usize>,
     universe: RoaringBitmap,
     sort_criteria: &Option<Vec<AscDesc>>,
     distinct: &Option<String>,
-    geo_param: geo_sort::Parameter,
+    geo_param: GeoSortParameter,
     from: usize,
     length: usize,
     embedder_name: &str,
@@ -669,6 +671,8 @@ pub fn execute_vector_search(
         placeholder_search_logger,
         time_budget,
         ranking_score_threshold,
+        exhaustive_number_hits,
+        max_total_hits,
     )?;
 
     Ok(PartialSearchResult {
@@ -689,10 +693,11 @@ pub fn execute_search(
     terms_matching_strategy: TermsMatchingStrategy,
     scoring_strategy: ScoringStrategy,
     exhaustive_number_hits: bool,
+    max_total_hits: Option<usize>,
     mut universe: RoaringBitmap,
     sort_criteria: &Option<Vec<AscDesc>>,
     distinct: &Option<String>,
-    geo_param: geo_sort::Parameter,
+    geo_param: GeoSortParameter,
     from: usize,
     length: usize,
     words_limit: Option<usize>,
@@ -825,6 +830,8 @@ pub fn execute_search(
             query_graph_logger,
             time_budget,
             ranking_score_threshold,
+            exhaustive_number_hits,
+            max_total_hits,
         )?
     } else {
         let ranking_rules =
@@ -841,6 +848,8 @@ pub fn execute_search(
             placeholder_search_logger,
             time_budget,
             ranking_score_threshold,
+            exhaustive_number_hits,
+            max_total_hits,
         )?
     };
 
@@ -872,7 +881,7 @@ pub fn execute_search(
     })
 }
 
-fn check_sort_criteria(
+pub(crate) fn check_sort_criteria(
     ctx: &SearchContext<'_>,
     sort_criteria: Option<&Vec<AscDesc>>,
 ) -> Result<()> {
@@ -902,7 +911,7 @@ fn check_sort_criteria(
                 let (valid_fields, hidden_fields) =
                     ctx.index.remove_hidden_fields(ctx.txn, sortable_fields)?;
 
-                return Err(UserError::InvalidSortableAttribute {
+                return Err(UserError::InvalidSearchSortableAttribute {
                     field: field.to_string(),
                     valid_fields,
                     hidden_fields,
@@ -913,7 +922,7 @@ fn check_sort_criteria(
                 let (valid_fields, hidden_fields) =
                     ctx.index.remove_hidden_fields(ctx.txn, sortable_fields)?;
 
-                return Err(UserError::InvalidSortableAttribute {
+                return Err(UserError::InvalidSearchSortableAttribute {
                     field: RESERVED_GEO_FIELD_NAME.to_string(),
                     valid_fields,
                     hidden_fields,

crates/milli/src/test_index.rs

@@ -19,7 +19,9 @@ use crate::update::{
 };
 use crate::vector::settings::{EmbedderSource, EmbeddingSettings};
 use crate::vector::RuntimeEmbedders;
-use crate::{db_snap, obkv_to_json, Filter, FilterableAttributesRule, Index, Search, SearchResult};
+use crate::{
+    db_snap, obkv_to_json, Filter, FilterableAttributesRule, Index, Search, SearchResult, UserError,
+};
 
 pub(crate) struct TempIndex {
     pub inner: Index,
@@ -1341,8 +1343,8 @@ fn vectors_are_never_indexed_as_searchable_or_filterable() {
     let results = search
         .filter(Filter::from_str("_vectors.doggo = 6789").unwrap().unwrap())
         .execute()
-        .unwrap();
-    assert!(results.candidates.is_empty());
+        .unwrap_err();
+    assert!(matches!(results, Error::UserError(UserError::InvalidFilter(_))));
 
     index
         .update_settings(|settings| {
@@ -1373,6 +1375,6 @@ fn vectors_are_never_indexed_as_searchable_or_filterable() {
     let results = search
         .filter(Filter::from_str("_vectors.doggo = 6789").unwrap().unwrap())
         .execute()
-        .unwrap();
-    assert!(results.candidates.is_empty());
+        .unwrap_err();
+    assert!(matches!(results, Error::UserError(UserError::InvalidFilter(_))));
 }

crates/milli/src/update/clear_documents.rs

@@ -2,7 +2,7 @@ use heed::RwTxn;
 use roaring::RoaringBitmap;
 use time::OffsetDateTime;
 
-use crate::{FieldDistribution, Index, Result};
+use crate::{database_stats::DatabaseStats, FieldDistribution, Index, Result};
 
 pub struct ClearDocuments<'t, 'i> {
     wtxn: &'t mut RwTxn<'i>,
@@ -92,6 +92,10 @@ impl<'t, 'i> ClearDocuments<'t, 'i> {
 
         documents.clear(self.wtxn)?;
 
+        // Update the stats of the documents database after clearing all documents.
+        let stats = DatabaseStats::new(self.index.documents.remap_data_type(), self.wtxn)?;
+        self.index.put_documents_stats(self.wtxn, stats)?;
+
         Ok(number_of_documents)
     }
 }
@@ -122,6 +126,9 @@ mod tests {
 
         let rtxn = index.read_txn().unwrap();
 
+        // Variables for statistics verification
+        let stats = index.documents_stats(&rtxn).unwrap().unwrap();
+
         // the value is 7 because there is `[id, name, age, country, _geo, _geo.lng, _geo.lat]`
         assert_eq!(index.fields_ids_map(&rtxn).unwrap().len(), 7);
 
@@ -142,5 +149,9 @@ mod tests {
         assert!(index.field_id_docid_facet_f64s.is_empty(&rtxn).unwrap());
         assert!(index.field_id_docid_facet_strings.is_empty(&rtxn).unwrap());
         assert!(index.documents.is_empty(&rtxn).unwrap());
+
+        // Verify that the statistics are correctly updated after clearing documents
+        assert_eq!(index.number_of_documents(&rtxn).unwrap(), 0);
+        assert_eq!(stats.number_of_entries(), 0);
     }
 }

crates/milli/src/update/settings.rs

@@ -101,6 +101,10 @@ impl<T> Setting<T> {
         matches!(self, Self::NotSet)
     }
 
+    pub const fn is_reset(&self) -> bool {
+        matches!(self, Self::Reset)
+    }
+
     /// If `Self` is `Reset`, then map self to `Set` with the provided `val`.
     pub fn or_reset(self, val: T) -> Self {
         match self {
@@ -554,10 +558,10 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
         match self.searchable_fields {
             Setting::Set(ref fields) => {
                 // Check to see if the searchable fields changed before doing anything else
-                let old_fields = self.index.searchable_fields(self.wtxn)?;
+                let old_fields = self.index.user_defined_searchable_fields(self.wtxn)?;
                 let did_change = {
                     let new_fields = fields.iter().map(String::as_str).collect::<Vec<_>>();
-                    new_fields != old_fields
+                    old_fields.is_none_or(|old| new_fields != old)
                 };
                 if !did_change {
                     return Ok(false);
@@ -1213,6 +1217,10 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
                 // new config
                 EitherOrBoth::Right((name, mut setting)) => {
                     tracing::debug!(embedder = name, "new embedder");
+                    // if we are asked to reset an embedder that doesn't exist, just ignore it
+                    if setting.is_reset() {
+                        continue;
+                    }
                     // apply the default source in case the source was not set so that it gets validated
                     crate::vector::settings::EmbeddingSettings::apply_default_source(&mut setting);
                     crate::vector::settings::EmbeddingSettings::apply_default_openai_model(

crates/milli/src/vector/composite.rs

@@ -59,12 +59,24 @@ pub struct EmbedderOptions {
 
 impl Embedder {
     pub fn new(
-        EmbedderOptions { search, index }: EmbedderOptions,
+        EmbedderOptions { search: search_options, index: index_options }: EmbedderOptions,
         cache_cap: usize,
     ) -> Result<Self, NewEmbedderError> {
-        let search = SubEmbedder::new(search, cache_cap)?;
+        // don't check similarity if one child is a rest embedder with fragments
+        // FIXME: skipping the check isn't ideal but we are unsure how to handle fragments in this context
+        let mut skip_similarity_check = false;
+        for options in [&search_options, &index_options] {
+            if let SubEmbedderOptions::Rest(options) = &options {
+                if !options.search_fragments.is_empty() || !options.indexing_fragments.is_empty() {
+                    skip_similarity_check = true;
+                    break;
+                }
+            }
+        }
+
+        let search = SubEmbedder::new(search_options, cache_cap)?;
         // cache is only used at search
-        let index = SubEmbedder::new(index, 0)?;
+        let index = SubEmbedder::new(index_options, 0)?;
 
         // check dimensions
         if search.dimensions() != index.dimensions() {
@@ -73,7 +85,12 @@ impl Embedder {
                 index.dimensions(),
             ));
         }
+
         // check similarity
+        if skip_similarity_check {
+            return Ok(Self { search, index });
+        }
+
         let search_embeddings = search
             .embed(
                 vec![

crates/milli/src/vector/db.rs

@@ -128,6 +128,7 @@ impl EmbeddingStatus {
     pub fn is_user_provided(&self, docid: DocumentId) -> bool {
         self.user_provided.contains(docid)
     }
+
     /// Whether vectors should be regenerated for that document and that embedder.
     pub fn must_regenerate(&self, docid: DocumentId) -> bool {
         let invert = self.skip_regenerate_different_from_user_provided.contains(docid);