Merge #5418

5418: Cache embeddings in search r=Kerollmops a=dureuill

# Pull Request

## Related issue
TBD

## What does this PR do?
- Adds a cache for embeddings produced in search
- The cache is disabled by default, and can be enabled following the instructions [here](https://github.com/orgs/meilisearch/discussions/818).
- Had to accommodate the `timeout` test for openai that uses a mock that simulates a timeout on subsequent responses: since the test was reusing the same query, the cache would kick-in and no request would be made to the mock, meaning no timeout any longer and so a failing test 😅 
- `Embedder::embed_search` now accepts a reference instead of an owned `String`.

## Manual testing

- I created 4 indexes on a fresh DB with the same settings (one embedder from openai)
- I sent 1/4 of movies.json to each index
- I sent a federated search request against all 4 indexes, with the same query for each index, using the embedder of each index.

Results:

- The first call took 400ms to 1s. Before this change, it took in the 3s range.
- Any repeated call with the same query took in the range of 25ms.
- Looking at the details at trace log level, I can see that the first index that needs the embedding is taking most of the 400ms in `embed_one`. The other indexes report that the query text is found in the cache and they each take a few µs.

Co-authored-by: Louis Dureuil <louis@meilisearch.com>

Cargo.lock

@@ -3498,6 +3498,15 @@ version = "0.4.26"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "30bde2b3dc3671ae49d8e2e9f044c7c005836e7a023ee57cffa25ab82764bb9e"
 
+[[package]]
+name = "lru"
+version = "0.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "227748d55f2f0ab4735d87fd623798cb6b664512fe979705f829c9f81c934465"
+dependencies = [
+ "hashbrown 0.15.2",
+]
+
 [[package]]
 name = "lzma-rs"
 version = "0.3.0"
@@ -3778,6 +3787,7 @@ dependencies = [
  "json-depth-checker",
  "levenshtein_automata",
  "liquid",
+ "lru",
  "maplit",
  "md5",
  "meili-snap",

crates/index-scheduler/src/lib.rs

@@ -125,6 +125,10 @@ pub struct IndexSchedulerOptions {
     pub instance_features: InstanceTogglableFeatures,
     /// The experimental features enabled for this instance.
     pub auto_upgrade: bool,
+    /// The maximal number of entries in the search query cache of an embedder.
+    ///
+    /// 0 disables the cache.
+    pub embedding_cache_cap: usize,
 }
 
 /// Structure which holds meilisearch's indexes and schedules the tasks
@@ -156,6 +160,11 @@ pub struct IndexScheduler {
     /// The Authorization header to send to the webhook URL.
     pub(crate) webhook_authorization_header: Option<String>,
 
+    /// A map to retrieve the runtime representation of an embedder depending on its configuration.
+    ///
+    /// This map may return the same embedder object for two different indexes or embedder settings,
+    /// but it will only do this if the embedder configuration options are the same, leading
+    /// to the same embeddings for the same input text.
     embedders: Arc<RwLock<HashMap<EmbedderOptions, Arc<Embedder>>>>,
 
     // ================= test
@@ -818,7 +827,7 @@ impl IndexScheduler {
 
                     // add missing embedder
                     let embedder = Arc::new(
-                        Embedder::new(embedder_options.clone())
+                        Embedder::new(embedder_options.clone(), self.scheduler.embedding_cache_cap)
                             .map_err(meilisearch_types::milli::vector::Error::from)
                             .map_err(|err| {
                                 Error::from_milli(err.into(), Some(index_uid.clone()))

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

@@ -76,6 +76,11 @@ pub struct Scheduler {
 
     /// The path to the version file of Meilisearch.
     pub(crate) version_file_path: PathBuf,
+
+    /// The maximal number of entries in the search query cache of an embedder.
+    ///
+    /// 0 disables the cache.
+    pub(crate) embedding_cache_cap: usize,
 }
 
 impl Scheduler {
@@ -90,6 +95,7 @@ impl Scheduler {
             snapshots_path: self.snapshots_path.clone(),
             auth_env: self.auth_env.clone(),
             version_file_path: self.version_file_path.clone(),
+            embedding_cache_cap: self.embedding_cache_cap,
         }
     }
 
@@ -105,6 +111,7 @@ impl Scheduler {
             snapshots_path: options.snapshots_path.clone(),
             auth_env,
             version_file_path: options.version_file_path.clone(),
+            embedding_cache_cap: options.embedding_cache_cap,
         }
     }
 }

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

@@ -104,10 +104,9 @@ fn import_vectors() {
 
         let configs = index_scheduler.embedders("doggos".to_string(), configs).unwrap();
         let (hf_embedder, _, _) = configs.get(&simple_hf_name).unwrap();
-        let beagle_embed =
-            hf_embedder.embed_search(S("Intel the beagle best doggo"), None).unwrap();
-        let lab_embed = hf_embedder.embed_search(S("Max the lab best doggo"), None).unwrap();
-        let patou_embed = hf_embedder.embed_search(S("kefir the patou best doggo"), None).unwrap();
+        let beagle_embed = hf_embedder.embed_search("Intel the beagle best doggo", None).unwrap();
+        let lab_embed = hf_embedder.embed_search("Max the lab best doggo", None).unwrap();
+        let patou_embed = hf_embedder.embed_search("kefir the patou best doggo", None).unwrap();
         (fakerest_name, simple_hf_name, beagle_embed, lab_embed, patou_embed)
     };
 

crates/index-scheduler/src/test_utils.rs

@@ -112,6 +112,7 @@ impl IndexScheduler {
             batched_tasks_size_limit: u64::MAX,
             instance_features: Default::default(),
             auto_upgrade: true, // Don't cost much and will ensure the happy path works
+            embedding_cache_cap: 10,
         };
         let version = configuration(&mut options).unwrap_or_else(|| {
             (

crates/meilisearch/src/analytics/segment_analytics.rs

@@ -199,6 +199,7 @@ struct Infos {
     experimental_network: bool,
     experimental_get_task_documents_route: bool,
     experimental_composite_embedders: bool,
+    experimental_embedding_cache_entries: usize,
     gpu_enabled: bool,
     db_path: bool,
     import_dump: bool,
@@ -246,6 +247,7 @@ impl Infos {
             experimental_reduce_indexing_memory_usage,
             experimental_max_number_of_batched_tasks,
             experimental_limit_batched_tasks_total_size,
+            experimental_embedding_cache_entries,
             http_addr,
             master_key: _,
             env,
@@ -312,6 +314,7 @@ impl Infos {
             experimental_network: network,
             experimental_get_task_documents_route: get_task_documents_route,
             experimental_composite_embedders: composite_embedders,
+            experimental_embedding_cache_entries,
             gpu_enabled: meilisearch_types::milli::vector::is_cuda_enabled(),
             db_path: db_path != PathBuf::from("./data.ms"),
             import_dump: import_dump.is_some(),

crates/meilisearch/src/lib.rs

@@ -233,6 +233,7 @@ pub fn setup_meilisearch(opt: &Opt) -> anyhow::Result<(Arc<IndexScheduler>, Arc<
         index_count: DEFAULT_INDEX_COUNT,
         instance_features: opt.to_instance_features(),
         auto_upgrade: opt.experimental_dumpless_upgrade,
+        embedding_cache_cap: opt.experimental_embedding_cache_entries,
     };
     let bin_major: u32 = VERSION_MAJOR.parse().unwrap();
     let bin_minor: u32 = VERSION_MINOR.parse().unwrap();

crates/meilisearch/src/option.rs

@@ -63,7 +63,8 @@ const MEILI_EXPERIMENTAL_MAX_NUMBER_OF_BATCHED_TASKS: &str =
     "MEILI_EXPERIMENTAL_MAX_NUMBER_OF_BATCHED_TASKS";
 const MEILI_EXPERIMENTAL_LIMIT_BATCHED_TASKS_TOTAL_SIZE: &str =
     "MEILI_EXPERIMENTAL_LIMIT_BATCHED_TASKS_SIZE";
-
+const MEILI_EXPERIMENTAL_EMBEDDING_CACHE_ENTRIES: &str =
+    "MEILI_EXPERIMENTAL_EMBEDDING_CACHE_ENTRIES";
 const DEFAULT_CONFIG_FILE_PATH: &str = "./config.toml";
 const DEFAULT_DB_PATH: &str = "./data.ms";
 const DEFAULT_HTTP_ADDR: &str = "localhost:7700";
@@ -446,6 +447,14 @@ pub struct Opt {
     #[serde(default = "default_limit_batched_tasks_total_size")]
     pub experimental_limit_batched_tasks_total_size: u64,
 
+    /// Enables experimental caching of search query embeddings. The value represents the maximal number of entries in the cache of each
+    /// distinct embedder.
+    ///
+    /// For more information, see <https://github.com/orgs/meilisearch/discussions/818>.
+    #[clap(long, env = MEILI_EXPERIMENTAL_EMBEDDING_CACHE_ENTRIES, default_value_t = default_embedding_cache_entries())]
+    #[serde(default = "default_embedding_cache_entries")]
+    pub experimental_embedding_cache_entries: usize,
+
     #[serde(flatten)]
     #[clap(flatten)]
     pub indexer_options: IndexerOpts,
@@ -549,6 +558,7 @@ impl Opt {
             experimental_reduce_indexing_memory_usage,
             experimental_max_number_of_batched_tasks,
             experimental_limit_batched_tasks_total_size,
+            experimental_embedding_cache_entries,
         } = self;
         export_to_env_if_not_present(MEILI_DB_PATH, db_path);
         export_to_env_if_not_present(MEILI_HTTP_ADDR, http_addr);
@@ -641,6 +651,10 @@ impl Opt {
             MEILI_EXPERIMENTAL_LIMIT_BATCHED_TASKS_TOTAL_SIZE,
             experimental_limit_batched_tasks_total_size.to_string(),
         );
+        export_to_env_if_not_present(
+            MEILI_EXPERIMENTAL_EMBEDDING_CACHE_ENTRIES,
+            experimental_embedding_cache_entries.to_string(),
+        );
         indexer_options.export_to_env();
     }
 
@@ -948,6 +962,10 @@ fn default_limit_batched_tasks_total_size() -> u64 {
     u64::MAX
 }
 
+fn default_embedding_cache_entries() -> usize {
+    0
+}
+
 fn default_snapshot_dir() -> PathBuf {
     PathBuf::from(DEFAULT_SNAPSHOT_DIR)
 }

crates/meilisearch/src/search/mod.rs

@@ -916,7 +916,7 @@ fn prepare_search<'t>(
                     let deadline = std::time::Instant::now() + std::time::Duration::from_secs(10);
 
                     embedder
-                        .embed_search(query.q.clone().unwrap(), Some(deadline))
+                        .embed_search(query.q.as_ref().unwrap(), Some(deadline))
                         .map_err(milli::vector::Error::from)
                         .map_err(milli::Error::from)?
                 }

crates/meilisearch/tests/vector/openai.rs

@@ -1995,7 +1995,7 @@ async fn timeout() {
 
     let (response, code) = index
         .search_post(json!({
-            "q": "grand chien de berger des montagnes",
+            "q": "grand chien de berger des montagnes foil the cache",
             "hybrid": {"semanticRatio": 0.99, "embedder": "default"}
         }))
         .await;

crates/milli/Cargo.toml

@@ -110,6 +110,7 @@ utoipa = { version = "5.3.1", features = [
     "time",
     "openapi_extensions",
 ] }
+lru = "0.13.0"
 
 [dev-dependencies]
 mimalloc = { version = "0.1.43", default-features = false }

crates/milli/src/search/hybrid.rs

@@ -203,7 +203,7 @@ impl<'a> Search<'a> {
 
                 let deadline = std::time::Instant::now() + std::time::Duration::from_secs(3);
 
-                match embedder.embed_search(query, Some(deadline)) {
+                match embedder.embed_search(&query, Some(deadline)) {
                     Ok(embedding) => embedding,
                     Err(error) => {
                         tracing::error!(error=%error, "Embedding failed");

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

@@ -2806,8 +2806,9 @@ mod tests {
             embedding_configs.pop().unwrap();
         insta::assert_snapshot!(embedder_name, @"manual");
         insta::assert_debug_snapshot!(user_provided, @"RoaringBitmap<[0, 1, 2]>");
-        let embedder =
-            std::sync::Arc::new(crate::vector::Embedder::new(embedder.embedder_options).unwrap());
+        let embedder = std::sync::Arc::new(
+            crate::vector::Embedder::new(embedder.embedder_options, 0).unwrap(),
+        );
         let res = index
             .search(&rtxn)
             .semantic(embedder_name, embedder, false, Some([0.0, 1.0, 2.0].to_vec()))

crates/milli/src/update/settings.rs

@@ -1628,7 +1628,8 @@ fn embedders(embedding_configs: Vec<IndexEmbeddingConfig>) -> Result<EmbeddingCo
                 let prompt = Arc::new(prompt.try_into().map_err(crate::Error::from)?);
 
                 let embedder = Arc::new(
-                    Embedder::new(embedder_options.clone())
+                    // cache_cap: no cache needed for indexing purposes
+                    Embedder::new(embedder_options.clone(), 0)
                         .map_err(crate::vector::Error::from)
                         .map_err(crate::Error::from)?,
                 );

crates/milli/src/vector/composite.rs

@@ -4,7 +4,8 @@ use arroy::Distance;
 
 use super::error::CompositeEmbedderContainsHuggingFace;
 use super::{
-    hf, manual, ollama, openai, rest, DistributionShift, EmbedError, Embedding, NewEmbedderError,
+    hf, manual, ollama, openai, rest, DistributionShift, EmbedError, Embedding, EmbeddingCache,
+    NewEmbedderError,
 };
 use crate::ThreadPoolNoAbort;
 
@@ -58,9 +59,11 @@ pub struct EmbedderOptions {
 impl Embedder {
     pub fn new(
         EmbedderOptions { search, index }: EmbedderOptions,
+        cache_cap: usize,
     ) -> Result<Self, NewEmbedderError> {
-        let search = SubEmbedder::new(search)?;
-        let index = SubEmbedder::new(index)?;
+        let search = SubEmbedder::new(search, cache_cap)?;
+        // cache is only used at search
+        let index = SubEmbedder::new(index, 0)?;
 
         // check dimensions
         if search.dimensions() != index.dimensions() {
@@ -118,19 +121,28 @@ impl Embedder {
 }
 
 impl SubEmbedder {
-    pub fn new(options: SubEmbedderOptions) -> std::result::Result<Self, NewEmbedderError> {
+    pub fn new(
+        options: SubEmbedderOptions,
+        cache_cap: usize,
+    ) -> std::result::Result<Self, NewEmbedderError> {
         Ok(match options {
             SubEmbedderOptions::HuggingFace(options) => {
-                Self::HuggingFace(hf::Embedder::new(options)?)
+                Self::HuggingFace(hf::Embedder::new(options, cache_cap)?)
+            }
+            SubEmbedderOptions::OpenAi(options) => {
+                Self::OpenAi(openai::Embedder::new(options, cache_cap)?)
+            }
+            SubEmbedderOptions::Ollama(options) => {
+                Self::Ollama(ollama::Embedder::new(options, cache_cap)?)
             }
-            SubEmbedderOptions::OpenAi(options) => Self::OpenAi(openai::Embedder::new(options)?),
-            SubEmbedderOptions::Ollama(options) => Self::Ollama(ollama::Embedder::new(options)?),
             SubEmbedderOptions::UserProvided(options) => {
                 Self::UserProvided(manual::Embedder::new(options))
             }
-            SubEmbedderOptions::Rest(options) => {
-                Self::Rest(rest::Embedder::new(options, rest::ConfigurationSource::User)?)
-            }
+            SubEmbedderOptions::Rest(options) => Self::Rest(rest::Embedder::new(
+                options,
+                cache_cap,
+                rest::ConfigurationSource::User,
+            )?),
         })
     }
 
@@ -148,6 +160,27 @@ impl SubEmbedder {
         }
     }
 
+    pub fn embed_one(
+        &self,
+        text: &str,
+        deadline: Option<Instant>,
+    ) -> std::result::Result<Embedding, EmbedError> {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => embedder.embed_one(text),
+            SubEmbedder::OpenAi(embedder) => {
+                embedder.embed(&[text], deadline)?.pop().ok_or_else(EmbedError::missing_embedding)
+            }
+            SubEmbedder::Ollama(embedder) => {
+                embedder.embed(&[text], deadline)?.pop().ok_or_else(EmbedError::missing_embedding)
+            }
+            SubEmbedder::UserProvided(embedder) => embedder.embed_one(text),
+            SubEmbedder::Rest(embedder) => embedder
+                .embed_ref(&[text], deadline)?
+                .pop()
+                .ok_or_else(EmbedError::missing_embedding),
+        }
+    }
+
     /// Embed multiple chunks of texts.
     ///
     /// Each chunk is composed of one or multiple texts.
@@ -233,6 +266,16 @@ impl SubEmbedder {
             SubEmbedder::Rest(embedder) => embedder.distribution(),
         }
     }
+
+    pub(super) fn cache(&self) -> Option<&EmbeddingCache> {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => Some(embedder.cache()),
+            SubEmbedder::OpenAi(embedder) => Some(embedder.cache()),
+            SubEmbedder::UserProvided(_) => None,
+            SubEmbedder::Ollama(embedder) => Some(embedder.cache()),
+            SubEmbedder::Rest(embedder) => Some(embedder.cache()),
+        }
+    }
 }
 
 fn check_similarity(

crates/milli/src/vector/hf.rs

@@ -7,7 +7,7 @@ use hf_hub::{Repo, RepoType};
 use tokenizers::{PaddingParams, Tokenizer};
 
 pub use super::error::{EmbedError, Error, NewEmbedderError};
-use super::{DistributionShift, Embedding};
+use super::{DistributionShift, Embedding, EmbeddingCache};
 
 #[derive(
     Debug,
@@ -84,6 +84,7 @@ pub struct Embedder {
     options: EmbedderOptions,
     dimensions: usize,
     pooling: Pooling,
+    cache: EmbeddingCache,
 }
 
 impl std::fmt::Debug for Embedder {
@@ -149,7 +150,10 @@ impl From<PoolingConfig> for Pooling {
 }
 
 impl Embedder {
-    pub fn new(options: EmbedderOptions) -> std::result::Result<Self, NewEmbedderError> {
+    pub fn new(
+        options: EmbedderOptions,
+        cache_cap: usize,
+    ) -> std::result::Result<Self, NewEmbedderError> {
         let device = match candle_core::Device::cuda_if_available(0) {
             Ok(device) => device,
             Err(error) => {
@@ -245,7 +249,14 @@ impl Embedder {
             tokenizer.with_padding(Some(pp));
         }
 
-        let mut this = Self { model, tokenizer, options, dimensions: 0, pooling };
+        let mut this = Self {
+            model,
+            tokenizer,
+            options,
+            dimensions: 0,
+            pooling,
+            cache: EmbeddingCache::new(cache_cap),
+        };
 
         let embeddings = this
             .embed(vec!["test".into()])
@@ -355,4 +366,8 @@ impl Embedder {
     pub(crate) fn embed_index_ref(&self, texts: &[&str]) -> Result<Vec<Embedding>, EmbedError> {
         texts.iter().map(|text| self.embed_one(text)).collect()
     }
+
+    pub(super) fn cache(&self) -> &EmbeddingCache {
+        &self.cache
+    }
 }

crates/milli/src/vector/mod.rs

@@ -1,5 +1,6 @@
 use std::collections::HashMap;
-use std::sync::Arc;
+use std::num::NonZeroUsize;
+use std::sync::{Arc, Mutex};
 use std::time::Instant;
 
 use arroy::distances::{BinaryQuantizedCosine, Cosine};
@@ -551,6 +552,46 @@ pub enum Embedder {
     Composite(composite::Embedder),
 }
 
+#[derive(Debug)]
+struct EmbeddingCache {
+    data: Option<Mutex<lru::LruCache<String, Embedding>>>,
+}
+
+impl EmbeddingCache {
+    const MAX_TEXT_LEN: usize = 2000;
+
+    pub fn new(cap: usize) -> Self {
+        let data = NonZeroUsize::new(cap).map(lru::LruCache::new).map(Mutex::new);
+        Self { data }
+    }
+
+    /// Get the embedding corresponding to `text`, if any is present in the cache.
+    pub fn get(&self, text: &str) -> Option<Embedding> {
+        let data = self.data.as_ref()?;
+        if text.len() > Self::MAX_TEXT_LEN {
+            return None;
+        }
+        let mut cache = data.lock().unwrap();
+
+        cache.get(text).cloned()
+    }
+
+    /// Puts a new embedding for the specified `text`
+    pub fn put(&self, text: String, embedding: Embedding) {
+        let Some(data) = self.data.as_ref() else {
+            return;
+        };
+        if text.len() > Self::MAX_TEXT_LEN {
+            return;
+        }
+        tracing::trace!(text, "embedding added to cache");
+
+        let mut cache = data.lock().unwrap();
+
+        cache.put(text, embedding);
+    }
+}
+
 /// Configuration for an embedder.
 #[derive(Debug, Clone, Default, serde::Deserialize, serde::Serialize)]
 pub struct EmbeddingConfig {
@@ -629,19 +670,30 @@ impl Default for EmbedderOptions {
 
 impl Embedder {
     /// Spawns a new embedder built from its options.
-    pub fn new(options: EmbedderOptions) -> std::result::Result<Self, NewEmbedderError> {
+    pub fn new(
+        options: EmbedderOptions,
+        cache_cap: usize,
+    ) -> std::result::Result<Self, NewEmbedderError> {
         Ok(match options {
-            EmbedderOptions::HuggingFace(options) => Self::HuggingFace(hf::Embedder::new(options)?),
-            EmbedderOptions::OpenAi(options) => Self::OpenAi(openai::Embedder::new(options)?),
-            EmbedderOptions::Ollama(options) => Self::Ollama(ollama::Embedder::new(options)?),
+            EmbedderOptions::HuggingFace(options) => {
+                Self::HuggingFace(hf::Embedder::new(options, cache_cap)?)
+            }
+            EmbedderOptions::OpenAi(options) => {
+                Self::OpenAi(openai::Embedder::new(options, cache_cap)?)
+            }
+            EmbedderOptions::Ollama(options) => {
+                Self::Ollama(ollama::Embedder::new(options, cache_cap)?)
+            }
             EmbedderOptions::UserProvided(options) => {
                 Self::UserProvided(manual::Embedder::new(options))
             }
-            EmbedderOptions::Rest(options) => {
-                Self::Rest(rest::Embedder::new(options, rest::ConfigurationSource::User)?)
-            }
+            EmbedderOptions::Rest(options) => Self::Rest(rest::Embedder::new(
+                options,
+                cache_cap,
+                rest::ConfigurationSource::User,
+            )?),
             EmbedderOptions::Composite(options) => {
-                Self::Composite(composite::Embedder::new(options)?)
+                Self::Composite(composite::Embedder::new(options, cache_cap)?)
             }
         })
     }
@@ -651,19 +703,35 @@ impl Embedder {
     #[tracing::instrument(level = "debug", skip_all, target = "search")]
     pub fn embed_search(
         &self,
-        text: String,
+        text: &str,
         deadline: Option<Instant>,
     ) -> std::result::Result<Embedding, EmbedError> {
-        let texts = vec![text];
-        let mut embedding = match self {
-            Embedder::HuggingFace(embedder) => embedder.embed(texts),
-            Embedder::OpenAi(embedder) => embedder.embed(&texts, deadline),
-            Embedder::Ollama(embedder) => embedder.embed(&texts, deadline),
-            Embedder::UserProvided(embedder) => embedder.embed(&texts),
-            Embedder::Rest(embedder) => embedder.embed(texts, deadline),
-            Embedder::Composite(embedder) => embedder.search.embed(texts, deadline),
+        if let Some(cache) = self.cache() {
+            if let Some(embedding) = cache.get(text) {
+                tracing::trace!(text, "embedding found in cache");
+                return Ok(embedding);
+            }
+        }
+        let embedding = match self {
+            Embedder::HuggingFace(embedder) => embedder.embed_one(text),
+            Embedder::OpenAi(embedder) => {
+                embedder.embed(&[text], deadline)?.pop().ok_or_else(EmbedError::missing_embedding)
+            }
+            Embedder::Ollama(embedder) => {
+                embedder.embed(&[text], deadline)?.pop().ok_or_else(EmbedError::missing_embedding)
+            }
+            Embedder::UserProvided(embedder) => embedder.embed_one(text),
+            Embedder::Rest(embedder) => embedder
+                .embed_ref(&[text], deadline)?
+                .pop()
+                .ok_or_else(EmbedError::missing_embedding),
+            Embedder::Composite(embedder) => embedder.search.embed_one(text, deadline),
         }?;
-        let embedding = embedding.pop().ok_or_else(EmbedError::missing_embedding)?;
+
+        if let Some(cache) = self.cache() {
+            cache.put(text.to_owned(), embedding.clone());
+        }
+
         Ok(embedding)
     }
 
@@ -759,6 +827,17 @@ impl Embedder {
             Embedder::Composite(embedder) => embedder.index.uses_document_template(),
         }
     }
+
+    fn cache(&self) -> Option<&EmbeddingCache> {
+        match self {
+            Embedder::HuggingFace(embedder) => Some(embedder.cache()),
+            Embedder::OpenAi(embedder) => Some(embedder.cache()),
+            Embedder::UserProvided(_) => None,
+            Embedder::Ollama(embedder) => Some(embedder.cache()),
+            Embedder::Rest(embedder) => Some(embedder.cache()),
+            Embedder::Composite(embedder) => embedder.search.cache(),
+        }
+    }
 }
 
 /// Describes the mean and sigma of distribution of embedding similarity in the embedding space.

crates/milli/src/vector/ollama.rs

@@ -5,7 +5,7 @@ use rayon::slice::ParallelSlice as _;
 
 use super::error::{EmbedError, EmbedErrorKind, NewEmbedderError, NewEmbedderErrorKind};
 use super::rest::{Embedder as RestEmbedder, EmbedderOptions as RestEmbedderOptions};
-use super::{DistributionShift, REQUEST_PARALLELISM};
+use super::{DistributionShift, EmbeddingCache, REQUEST_PARALLELISM};
 use crate::error::FaultSource;
 use crate::vector::Embedding;
 use crate::ThreadPoolNoAbort;
@@ -75,9 +75,10 @@ impl EmbedderOptions {
 }
 
 impl Embedder {
-    pub fn new(options: EmbedderOptions) -> Result<Self, NewEmbedderError> {
+    pub fn new(options: EmbedderOptions, cache_cap: usize) -> Result<Self, NewEmbedderError> {
         let rest_embedder = match RestEmbedder::new(
             options.into_rest_embedder_config()?,
+            cache_cap,
             super::rest::ConfigurationSource::Ollama,
         ) {
             Ok(embedder) => embedder,
@@ -182,6 +183,10 @@ impl Embedder {
     pub fn distribution(&self) -> Option<DistributionShift> {
         self.rest_embedder.distribution()
     }
+
+    pub(super) fn cache(&self) -> &EmbeddingCache {
+        self.rest_embedder.cache()
+    }
 }
 
 fn get_ollama_path() -> String {

crates/milli/src/vector/openai.rs

@@ -7,7 +7,7 @@ use rayon::slice::ParallelSlice as _;
 
 use super::error::{EmbedError, NewEmbedderError};
 use super::rest::{Embedder as RestEmbedder, EmbedderOptions as RestEmbedderOptions};
-use super::{DistributionShift, REQUEST_PARALLELISM};
+use super::{DistributionShift, EmbeddingCache, REQUEST_PARALLELISM};
 use crate::error::FaultSource;
 use crate::vector::error::EmbedErrorKind;
 use crate::vector::Embedding;
@@ -176,7 +176,7 @@ pub struct Embedder {
 }
 
 impl Embedder {
-    pub fn new(options: EmbedderOptions) -> Result<Self, NewEmbedderError> {
+    pub fn new(options: EmbedderOptions, cache_cap: usize) -> Result<Self, NewEmbedderError> {
         let mut inferred_api_key = Default::default();
         let api_key = options.api_key.as_ref().unwrap_or_else(|| {
             inferred_api_key = infer_api_key();
@@ -201,6 +201,7 @@ impl Embedder {
                 }),
                 headers: Default::default(),
             },
+            cache_cap,
             super::rest::ConfigurationSource::OpenAi,
         )?;
 
@@ -318,6 +319,10 @@ impl Embedder {
     pub fn distribution(&self) -> Option<DistributionShift> {
         self.options.distribution()
     }
+
+    pub(super) fn cache(&self) -> &EmbeddingCache {
+        self.rest_embedder.cache()
+    }
 }
 
 impl fmt::Debug for Embedder {

crates/milli/src/vector/rest.rs

@@ -9,7 +9,9 @@ use serde::{Deserialize, Serialize};
 
 use super::error::EmbedErrorKind;
 use super::json_template::ValueTemplate;
-use super::{DistributionShift, EmbedError, Embedding, NewEmbedderError, REQUEST_PARALLELISM};
+use super::{
+    DistributionShift, EmbedError, Embedding, EmbeddingCache, NewEmbedderError, REQUEST_PARALLELISM,
+};
 use crate::error::FaultSource;
 use crate::ThreadPoolNoAbort;
 
@@ -75,6 +77,7 @@ pub struct Embedder {
     data: EmbedderData,
     dimensions: usize,
     distribution: Option<DistributionShift>,
+    cache: EmbeddingCache,
 }
 
 /// All data needed to perform requests and parse responses
@@ -123,6 +126,7 @@ enum InputType {
 impl Embedder {
     pub fn new(
         options: EmbedderOptions,
+        cache_cap: usize,
         configuration_source: ConfigurationSource,
     ) -> Result<Self, NewEmbedderError> {
         let bearer = options.api_key.as_deref().map(|api_key| format!("Bearer {api_key}"));
@@ -152,7 +156,12 @@ impl Embedder {
             infer_dimensions(&data)?
         };
 
-        Ok(Self { data, dimensions, distribution: options.distribution })
+        Ok(Self {
+            data,
+            dimensions,
+            distribution: options.distribution,
+            cache: EmbeddingCache::new(cache_cap),
+        })
     }
 
     pub fn embed(
@@ -256,6 +265,10 @@ impl Embedder {
     pub fn distribution(&self) -> Option<DistributionShift> {
         self.distribution
     }
+
+    pub(super) fn cache(&self) -> &EmbeddingCache {
+        &self.cache
+    }
 }
 
 fn infer_dimensions(data: &EmbedderData) -> Result<usize, NewEmbedderError> {