Small commit to add hybrid search and autoembedding

meilisearch/src/analytics/segment_analytics.rs

@@ -686,7 +686,7 @@ impl SearchAggregator {
             ret.max_terms_number = q.split_whitespace().count();
         }
 
-        if let Some(ref vector) = vector {
+        if let Some(meilisearch_types::milli::VectorQuery::Vector(ref vector)) = vector {
             ret.max_vector_size = vector.len();
         }
 

meilisearch/src/main.rs

@@ -19,7 +19,11 @@ static ALLOC: mimalloc::MiMalloc = mimalloc::MiMalloc;
 /// does all the setup before meilisearch is launched
 fn setup(opt: &Opt) -> anyhow::Result<()> {
     let mut log_builder = env_logger::Builder::new();
-    log_builder.parse_filters(&opt.log_level.to_string());
+    let log_filters = format!(
+        "{},h2=warn,hyper=warn,tokio_util=warn,tracing=warn,rustls=warn,mio=warn,reqwest=warn",
+        opt.log_level
+    );
+    log_builder.parse_filters(&log_filters);
 
     log_builder.init();
 

meilisearch/src/routes/indexes/facet_search.rs

@@ -7,6 +7,7 @@ use meilisearch_types::deserr::DeserrJsonError;
 use meilisearch_types::error::deserr_codes::*;
 use meilisearch_types::error::ResponseError;
 use meilisearch_types::index_uid::IndexUid;
+use meilisearch_types::milli::VectorQuery;
 use serde_json::Value;
 
 use crate::analytics::{Analytics, FacetSearchAggregator};
@@ -117,7 +118,7 @@ impl From<FacetSearchQuery> for SearchQuery {
             highlight_post_tag: DEFAULT_HIGHLIGHT_POST_TAG(),
             crop_marker: DEFAULT_CROP_MARKER(),
             matching_strategy,
-            vector,
+            vector: vector.map(VectorQuery::Vector),
             attributes_to_search_on,
         }
     }

meilisearch/src/routes/indexes/search.rs

@@ -2,12 +2,13 @@ use actix_web::web::Data;
 use actix_web::{web, HttpRequest, HttpResponse};
 use deserr::actix_web::{AwebJson, AwebQueryParameter};
 use index_scheduler::IndexScheduler;
-use log::debug;
+use log::{debug, warn};
 use meilisearch_types::deserr::query_params::Param;
 use meilisearch_types::deserr::{DeserrJsonError, DeserrQueryParamError};
 use meilisearch_types::error::deserr_codes::*;
 use meilisearch_types::error::ResponseError;
 use meilisearch_types::index_uid::IndexUid;
+use meilisearch_types::milli::VectorQuery;
 use meilisearch_types::serde_cs::vec::CS;
 use serde_json::Value;
 
@@ -88,7 +89,7 @@ impl From<SearchQueryGet> for SearchQuery {
 
         Self {
             q: other.q,
-            vector: other.vector.map(CS::into_inner),
+            vector: other.vector.map(CS::into_inner).map(VectorQuery::Vector),
             offset: other.offset.0,
             limit: other.limit.0,
             page: other.page.as_deref().copied(),
@@ -193,6 +194,9 @@ pub async fn search_with_post(
     let index = index_scheduler.index(&index_uid)?;
 
     let features = index_scheduler.features();
+
+    embed(&mut query, index_scheduler.get_ref(), &index).await?;
+
     let search_result =
         tokio::task::spawn_blocking(move || perform_search(&index, query, features)).await?;
     if let Ok(ref search_result) = search_result {
@@ -206,6 +210,38 @@ pub async fn search_with_post(
     Ok(HttpResponse::Ok().json(search_result))
 }
 
+pub async fn embed(
+    query: &mut SearchQuery,
+    index_scheduler: &IndexScheduler,
+    index: &meilisearch_types::milli::Index,
+) -> Result<(), ResponseError> {
+    if let Some(VectorQuery::String(prompt)) = query.vector.take() {
+        let embedder_configs = index.embedding_configs(&index.read_txn()?)?;
+        let embedder = index_scheduler.embedders(embedder_configs)?;
+
+        /// FIXME: add error if no embedder, remove unwrap, support multiple embedders
+        let embeddings = embedder
+            .get("default")
+            .unwrap()
+            .0
+            .embed(vec![prompt])
+            .await
+            .map_err(meilisearch_types::milli::vector::Error::from)
+            .map_err(meilisearch_types::milli::UserError::from)
+            .map_err(meilisearch_types::milli::Error::from)?
+            .pop()
+            .expect("No vector returned from embedding");
+
+        if embeddings.iter().nth(1).is_some() {
+            warn!("Ignoring embeddings past the first one in long search query");
+            query.vector = Some(VectorQuery::Vector(embeddings.iter().next().unwrap().to_vec()));
+        } else {
+            query.vector = Some(VectorQuery::Vector(embeddings.into_inner()));
+        }
+    };
+    Ok(())
+}
+
 #[cfg(test)]
 mod test {
     use super::*;

meilisearch/src/routes/multi_search.rs

@@ -13,6 +13,7 @@ use crate::analytics::{Analytics, MultiSearchAggregator};
 use crate::extractors::authentication::policies::ActionPolicy;
 use crate::extractors::authentication::{AuthenticationError, GuardedData};
 use crate::extractors::sequential_extractor::SeqHandler;
+use crate::routes::indexes::search::embed;
 use crate::search::{
     add_search_rules, perform_search, SearchQueryWithIndex, SearchResultWithIndex,
 };
@@ -74,6 +75,8 @@ pub async fn multi_search_with_post(
                 })
                 .with_index(query_index)?;
 
+            embed(&mut query, index_scheduler.get_ref(), &index).await.with_index(query_index)?;
+
             let search_result =
                 tokio::task::spawn_blocking(move || perform_search(&index, query, features))
                     .await

meilisearch/src/search.rs

@@ -16,6 +16,7 @@ use meilisearch_types::index_uid::IndexUid;
 use meilisearch_types::milli::score_details::{ScoreDetails, ScoringStrategy};
 use meilisearch_types::milli::{
     dot_product_similarity, FacetValueHit, InternalError, OrderBy, SearchForFacetValues,
+    VectorQuery,
 };
 use meilisearch_types::settings::DEFAULT_PAGINATION_MAX_TOTAL_HITS;
 use meilisearch_types::{milli, Document};
@@ -46,7 +47,7 @@ pub struct SearchQuery {
     #[deserr(default, error = DeserrJsonError<InvalidSearchQ>)]
     pub q: Option<String>,
     #[deserr(default, error = DeserrJsonError<InvalidSearchVector>)]
-    pub vector: Option<Vec<f32>>,
+    pub vector: Option<milli::VectorQuery>,
     #[deserr(default = DEFAULT_SEARCH_OFFSET(), error = DeserrJsonError<InvalidSearchOffset>)]
     pub offset: usize,
     #[deserr(default = DEFAULT_SEARCH_LIMIT(), error = DeserrJsonError<InvalidSearchLimit>)]
@@ -105,7 +106,7 @@ pub struct SearchQueryWithIndex {
     #[deserr(default, error = DeserrJsonError<InvalidSearchQ>)]
     pub q: Option<String>,
     #[deserr(default, error = DeserrJsonError<InvalidSearchQ>)]
-    pub vector: Option<Vec<f32>>,
+    pub vector: Option<VectorQuery>,
     #[deserr(default = DEFAULT_SEARCH_OFFSET(), error = DeserrJsonError<InvalidSearchOffset>)]
     pub offset: usize,
     #[deserr(default = DEFAULT_SEARCH_LIMIT(), error = DeserrJsonError<InvalidSearchLimit>)]
@@ -339,11 +340,18 @@ fn prepare_search<'t>(
     let mut search = index.search(rtxn);
 
     if query.vector.is_some() && query.q.is_some() {
-        warn!("Ignoring the query string `q` when used with the `vector` parameter.");
+        warn!("Attempting hybrid search");
     }
 
     if let Some(ref vector) = query.vector {
-        search.vector(vector.clone());
+        match vector {
+            VectorQuery::Vector(vector) => {
+                search.vector(vector.clone());
+            }
+            VectorQuery::String(_) => {
+                panic!("Failed while preparing search; caller did not generate embedding for query")
+            }
+        }
     }
 
     if let Some(ref query) = query.q {
@@ -375,7 +383,7 @@ fn prepare_search<'t>(
     }
 
     if query.vector.is_some() {
-        features.check_vector()?;
+        features.check_vector("Passing `vector` as a query parameter")?;
     }
 
     // compute the offset on the limit depending on the pagination mode.
@@ -429,7 +437,11 @@ pub fn perform_search(
         prepare_search(index, &rtxn, &query, features)?;
 
     let milli::SearchResult { documents_ids, matching_words, candidates, document_scores, .. } =
-        search.execute()?;
+        if query.q.is_some() && query.vector.is_some() {
+            search.execute_hybrid()?
+        } else {
+            search.execute()?
+        };
 
     let fields_ids_map = index.fields_ids_map(&rtxn).unwrap();
 
@@ -538,13 +550,13 @@ pub fn perform_search(
             insert_geo_distance(sort, &mut document);
         }
 
-        let semantic_score = match query.vector.as_ref() {
+        let semantic_score = /*match query.vector.as_ref() {
             Some(vector) => match extract_field("_vectors", &fields_ids_map, obkv)? {
                 Some(vectors) => compute_semantic_score(vector, vectors)?,
                 None => None,
             },
             None => None,
-        };
+        };*/ None;
 
         let ranking_score =
             query.show_ranking_score.then(|| ScoreDetails::global_score(score.iter()));
@@ -629,7 +641,8 @@ pub fn perform_search(
         hits: documents,
         hits_info,
         query: query.q.unwrap_or_default(),
-        vector: query.vector,
+        // FIXME: display input vector
+        vector: None,
         processing_time_ms: before_search.elapsed().as_millis(),
         facet_distribution,
         facet_stats,