TMP: remove optimization where later ranking rules are not applied on buckets of a single document

3786: Consistently use wrapping add to avoid overflow in debug when query s… r=dureuill a=dureuill

# Pull Request

## Related issue
Fixes https://github.com/meilisearch/meilisearch/issues/3785

## What does this PR do?
- Some of the code paths would erroneously use the default addition operator that has the semantics that "overflow is an error, checked at runtime in debug" instead of the intended "overflow is expected" semantics that this code use (this code is using `u16::MAX` as a sentinel). This PR makes it so the wrapping add operator is used everywhere.

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

meilisearch-types/src/deserr/mod.rs

@@ -150,6 +150,7 @@ make_missing_field_convenience_builder!(MissingApiKeyActions, missing_api_key_ac
 make_missing_field_convenience_builder!(MissingApiKeyExpiresAt, missing_api_key_expires_at);
 make_missing_field_convenience_builder!(MissingApiKeyIndexes, missing_api_key_indexes);
 make_missing_field_convenience_builder!(MissingSwapIndexes, missing_swap_indexes);
+make_missing_field_convenience_builder!(MissingDocumentFilter, missing_document_filter);
 
 // Integrate a sub-error into a [`DeserrError`] by taking its error message but using
 // the default error code (C) from `Self`

meilisearch-types/src/error.rs

@@ -214,12 +214,12 @@ InvalidApiKeyUid                      , InvalidRequest       , BAD_REQUEST ;
 InvalidContentType                    , InvalidRequest       , UNSUPPORTED_MEDIA_TYPE ;
 InvalidDocumentCsvDelimiter           , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentFields                 , InvalidRequest       , BAD_REQUEST ;
+MissingDocumentFilter                 , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentFilter                 , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentGeoField               , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentId                     , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentLimit                  , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentOffset                 , InvalidRequest       , BAD_REQUEST ;
-InvalidDocumentDeleteFilter           , InvalidRequest       , BAD_REQUEST ;
 InvalidIndexLimit                     , InvalidRequest       , BAD_REQUEST ;
 InvalidIndexOffset                    , InvalidRequest       , BAD_REQUEST ;
 InvalidIndexPrimaryKey                , InvalidRequest       , BAD_REQUEST ;

meilisearch/src/error.rs

@@ -61,7 +61,7 @@ impl ErrorCode for MeilisearchHttpError {
             MeilisearchHttpError::MissingPayload(_) => Code::MissingPayload,
             MeilisearchHttpError::InvalidContentType(_, _) => Code::InvalidContentType,
             MeilisearchHttpError::DocumentNotFound(_) => Code::DocumentNotFound,
-            MeilisearchHttpError::EmptyFilter => Code::InvalidDocumentDeleteFilter,
+            MeilisearchHttpError::EmptyFilter => Code::InvalidDocumentFilter,
             MeilisearchHttpError::InvalidExpression(_, _) => Code::InvalidSearchFilter,
             MeilisearchHttpError::PayloadTooLarge(_) => Code::PayloadTooLarge,
             MeilisearchHttpError::SwapIndexPayloadWrongLength(_) => Code::InvalidSwapIndexes,

meilisearch/src/routes/indexes/documents.rs

@@ -486,7 +486,7 @@ pub async fn delete_documents_batch(
 #[derive(Debug, Deserr)]
 #[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
 pub struct DocumentDeletionByFilter {
-    #[deserr(error = DeserrJsonError<InvalidDocumentDeleteFilter>)]
+    #[deserr(error = DeserrJsonError<InvalidDocumentFilter>, missing_field_error = DeserrJsonError::missing_document_filter)]
     filter: Value,
 }
 
@@ -508,8 +508,8 @@ pub async fn delete_documents_by_filter(
     || -> Result<_, ResponseError> {
         Ok(crate::search::parse_filter(&filter)?.ok_or(MeilisearchHttpError::EmptyFilter)?)
     }()
-    // and whatever was the error, the error code should always be an InvalidDocumentDeleteFilter
-    .map_err(|err| ResponseError::from_msg(err.message, Code::InvalidDocumentDeleteFilter))?;
+    // and whatever was the error, the error code should always be an InvalidDocumentFilter
+    .map_err(|err| ResponseError::from_msg(err.message, Code::InvalidDocumentFilter))?;
     let task = KindWithContent::DocumentDeletionByFilter { index_uid, filter_expr: filter };
 
     let task: SummarizedTaskView =

meilisearch/src/routes/tasks.rs

@@ -99,7 +99,7 @@ pub struct DetailsView {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub deleted_tasks: Option<Option<u64>>,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub original_filter: Option<String>,
+    pub original_filter: Option<Option<String>>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub dump_uid: Option<Option<String>>,
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -131,12 +131,13 @@ impl From<Details> for DetailsView {
             } => DetailsView {
                 provided_ids: Some(received_document_ids),
                 deleted_documents: Some(deleted_documents),
+                original_filter: Some(None),
                 ..DetailsView::default()
             },
             Details::DocumentDeletionByFilter { original_filter, deleted_documents } => {
                 DetailsView {
                     provided_ids: Some(0),
-                    original_filter: Some(original_filter),
+                    original_filter: Some(Some(original_filter)),
                     deleted_documents: Some(deleted_documents),
                     ..DetailsView::default()
                 }
@@ -148,7 +149,7 @@ impl From<Details> for DetailsView {
                 DetailsView {
                     matched_tasks: Some(matched_tasks),
                     canceled_tasks: Some(canceled_tasks),
-                    original_filter: Some(original_filter),
+                    original_filter: Some(Some(original_filter)),
                     ..DetailsView::default()
                 }
             }
@@ -156,7 +157,7 @@ impl From<Details> for DetailsView {
                 DetailsView {
                     matched_tasks: Some(matched_tasks),
                     deleted_tasks: Some(deleted_tasks),
-                    original_filter: Some(original_filter),
+                    original_filter: Some(Some(original_filter)),
                     ..DetailsView::default()
                 }
             }

meilisearch/tests/documents/errors.rs

@@ -547,9 +547,9 @@ async fn delete_document_by_filter() {
     snapshot!(json_string!(response), @r###"
     {
       "message": "Invalid syntax for the filter parameter: `expected String, Array, found: true`.",
-      "code": "invalid_document_delete_filter",
+      "code": "invalid_document_filter",
       "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_document_delete_filter"
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
     }
     "###);
 
@@ -559,9 +559,9 @@ async fn delete_document_by_filter() {
     snapshot!(json_string!(response), @r###"
     {
       "message": "Was expecting an operation `=`, `!=`, `>=`, `>`, `<=`, `<`, `IN`, `NOT IN`, `TO`, `EXISTS`, `NOT EXISTS`, `IS NULL`, `IS NOT NULL`, `IS EMPTY`, `IS NOT EMPTY`, `_geoRadius`, or `_geoBoundingBox` at `hello`.\n1:6 hello",
-      "code": "invalid_document_delete_filter",
+      "code": "invalid_document_filter",
       "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_document_delete_filter"
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
     }
     "###);
 
@@ -571,9 +571,21 @@ async fn delete_document_by_filter() {
     snapshot!(json_string!(response), @r###"
     {
       "message": "Sending an empty filter is forbidden.",
-      "code": "invalid_document_delete_filter",
+      "code": "invalid_document_filter",
       "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_document_delete_filter"
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
+    }
+    "###);
+
+    // do not send any filter
+    let (response, code) = index.delete_document_by_filter(json!({})).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Missing field `filter`",
+      "code": "missing_document_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#missing_document_filter"
     }
     "###);
 

meilisearch/tests/tasks/mod.rs

@@ -413,7 +413,7 @@ async fn test_summarized_document_addition_or_update() {
 }
 
 #[actix_web::test]
-async fn test_summarized_delete_batch() {
+async fn test_summarized_delete_documents_by_batch() {
     let server = Server::new().await;
     let index = server.index("test");
     index.delete_batch(vec![1, 2, 3]).await;
@@ -430,7 +430,8 @@ async fn test_summarized_delete_batch() {
       "canceledBy": null,
       "details": {
         "providedIds": 3,
-        "deletedDocuments": 0
+        "deletedDocuments": 0,
+        "originalFilter": null
       },
       "error": {
         "message": "Index `test` not found.",
@@ -460,7 +461,8 @@ async fn test_summarized_delete_batch() {
       "canceledBy": null,
       "details": {
         "providedIds": 1,
-        "deletedDocuments": 0
+        "deletedDocuments": 0,
+        "originalFilter": null
       },
       "error": null,
       "duration": "[duration]",
@@ -472,7 +474,100 @@ async fn test_summarized_delete_batch() {
 }
 
 #[actix_web::test]
-async fn test_summarized_delete_document() {
+async fn test_summarized_delete_documents_by_filter() {
+    let server = Server::new().await;
+    let index = server.index("test");
+
+    index.delete_document_by_filter(json!({ "filter": "doggo = bernese" })).await;
+    index.wait_task(0).await;
+    let (task, _) = index.get_task(0).await;
+    assert_json_snapshot!(task,
+        { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" },
+        @r###"
+    {
+      "uid": 0,
+      "indexUid": "test",
+      "status": "failed",
+      "type": "documentDeletion",
+      "canceledBy": null,
+      "details": {
+        "providedIds": 0,
+        "deletedDocuments": 0,
+        "originalFilter": "\"doggo = bernese\""
+      },
+      "error": {
+        "message": "Index `test` not found.",
+        "code": "index_not_found",
+        "type": "invalid_request",
+        "link": "https://docs.meilisearch.com/errors#index_not_found"
+      },
+      "duration": "[duration]",
+      "enqueuedAt": "[date]",
+      "startedAt": "[date]",
+      "finishedAt": "[date]"
+    }
+    "###);
+
+    index.create(None).await;
+    index.delete_document_by_filter(json!({ "filter": "doggo = bernese" })).await;
+    index.wait_task(2).await;
+    let (task, _) = index.get_task(2).await;
+    assert_json_snapshot!(task,
+        { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" },
+        @r###"
+    {
+      "uid": 2,
+      "indexUid": "test",
+      "status": "failed",
+      "type": "documentDeletion",
+      "canceledBy": null,
+      "details": {
+        "providedIds": 0,
+        "deletedDocuments": 0,
+        "originalFilter": "\"doggo = bernese\""
+      },
+      "error": {
+        "message": "Attribute `doggo` is not filterable. This index does not have configured filterable attributes.\n1:6 doggo = bernese",
+        "code": "invalid_document_filter",
+        "type": "invalid_request",
+        "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
+      },
+      "duration": "[duration]",
+      "enqueuedAt": "[date]",
+      "startedAt": "[date]",
+      "finishedAt": "[date]"
+    }
+    "###);
+
+    index.update_settings(json!({ "filterableAttributes": ["doggo"] })).await;
+    index.delete_document_by_filter(json!({ "filter": "doggo = bernese" })).await;
+    index.wait_task(4).await;
+    let (task, _) = index.get_task(4).await;
+    assert_json_snapshot!(task,
+        { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" },
+        @r###"
+    {
+      "uid": 4,
+      "indexUid": "test",
+      "status": "succeeded",
+      "type": "documentDeletion",
+      "canceledBy": null,
+      "details": {
+        "providedIds": 0,
+        "deletedDocuments": 0,
+        "originalFilter": "\"doggo = bernese\""
+      },
+      "error": null,
+      "duration": "[duration]",
+      "enqueuedAt": "[date]",
+      "startedAt": "[date]",
+      "finishedAt": "[date]"
+    }
+    "###);
+}
+
+#[actix_web::test]
+async fn test_summarized_delete_document_by_id() {
     let server = Server::new().await;
     let index = server.index("test");
     index.delete_document(1).await;
@@ -489,7 +584,8 @@ async fn test_summarized_delete_document() {
       "canceledBy": null,
       "details": {
         "providedIds": 1,
-        "deletedDocuments": 0
+        "deletedDocuments": 0,
+        "originalFilter": null
       },
       "error": {
         "message": "Index `test` not found.",
@@ -519,7 +615,8 @@ async fn test_summarized_delete_document() {
       "canceledBy": null,
       "details": {
         "providedIds": 1,
-        "deletedDocuments": 0
+        "deletedDocuments": 0,
+        "originalFilter": null
       },
       "error": null,
       "duration": "[duration]",

milli/src/search/new/bucket_sort.rs

@@ -116,16 +116,15 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
     }
 
     while valid_docids.len() < length {
-        // The universe for this bucket is zero or one element, so we don't need to sort
-        // anything, just extend the results and go back to the parent ranking rule.
-        if ranking_rule_universes[cur_ranking_rule_index].len() <= 1 {
-            let bucket = std::mem::take(&mut ranking_rule_universes[cur_ranking_rule_index]);
-            maybe_add_to_results!(bucket);
+        // The universe for this bucket is zero element, so we don't need to sort
+        // anything, just go back to the parent ranking rule.
+        if ranking_rule_universes[cur_ranking_rule_index].is_empty() {
             back!();
             continue;
         }
 
-        let Some(next_bucket) = ranking_rules[cur_ranking_rule_index].next_bucket(ctx, logger, &ranking_rule_universes[cur_ranking_rule_index])? else {
+        let Some(next_bucket) = ranking_rules[cur_ranking_rule_index].next_bucket(ctx, logger, &ranking_rule_universes[cur_ranking_rule_index])?
+        else {
             back!();
             continue;
         };

milli/src/search/new/graph_based_ranking_rule.rs

@@ -46,7 +46,7 @@ use super::logger::SearchLogger;
 use super::query_graph::QueryNode;
 use super::ranking_rule_graph::{
     ConditionDocIdsCache, DeadEndsCache, ExactnessGraph, FidGraph, PositionGraph, ProximityGraph,
-    RankingRuleGraph, RankingRuleGraphTrait, TypoGraph,
+    RankingRuleGraph, RankingRuleGraphTrait, TypoGraph, WordsGraph,
 };
 use super::small_bitmap::SmallBitmap;
 use super::{QueryGraph, RankingRule, RankingRuleOutput, SearchContext};
@@ -54,6 +54,12 @@ use crate::search::new::query_term::LocatedQueryTermSubset;
 use crate::search::new::ranking_rule_graph::PathVisitor;
 use crate::{Result, TermsMatchingStrategy};
 
+pub type Words = GraphBasedRankingRule<WordsGraph>;
+impl GraphBasedRankingRule<WordsGraph> {
+    pub fn new(terms_matching_strategy: TermsMatchingStrategy) -> Self {
+        Self::new_with_id("words".to_owned(), Some(terms_matching_strategy))
+    }
+}
 pub type Proximity = GraphBasedRankingRule<ProximityGraph>;
 impl GraphBasedRankingRule<ProximityGraph> {
     pub fn new(terms_matching_strategy: Option<TermsMatchingStrategy>) -> Self {
@@ -175,9 +181,6 @@ impl<'ctx, G: RankingRuleGraphTrait> RankingRule<'ctx, QueryGraph> for GraphBase
         logger: &mut dyn SearchLogger<QueryGraph>,
         universe: &RoaringBitmap,
     ) -> Result<Option<RankingRuleOutput<QueryGraph>>> {
-        // If universe.len() <= 1, the bucket sort algorithm
-        // should not have called this function.
-        assert!(universe.len() > 1);
         // Will crash if `next_bucket` is called before `start_iteration` or after `end_iteration`,
         // should never happen
         let mut state = self.state.take().unwrap();

milli/src/search/new/logger/visual.rs

@@ -4,7 +4,6 @@ use std::io::{BufWriter, Write};
 use std::path::{Path, PathBuf};
 use std::time::Instant;
 
-// use rand::random;
 use roaring::RoaringBitmap;
 
 use crate::search::new::interner::Interned;
@@ -13,6 +12,7 @@ use crate::search::new::query_term::LocatedQueryTermSubset;
 use crate::search::new::ranking_rule_graph::{
     Edge, FidCondition, FidGraph, PositionCondition, PositionGraph, ProximityCondition,
     ProximityGraph, RankingRuleGraph, RankingRuleGraphTrait, TypoCondition, TypoGraph,
+    WordsCondition, WordsGraph,
 };
 use crate::search::new::ranking_rules::BoxRankingRule;
 use crate::search::new::{QueryGraph, QueryNode, RankingRule, SearchContext, SearchLogger};
@@ -24,11 +24,12 @@ pub enum SearchEvents {
     RankingRuleSkipBucket { ranking_rule_idx: usize, bucket_len: u64 },
     RankingRuleEndIteration { ranking_rule_idx: usize, universe_len: u64 },
     ExtendResults { new: Vec<u32> },
-    WordsGraph { query_graph: QueryGraph },
     ProximityGraph { graph: RankingRuleGraph<ProximityGraph> },
     ProximityPaths { paths: Vec<Vec<Interned<ProximityCondition>>> },
     TypoGraph { graph: RankingRuleGraph<TypoGraph> },
     TypoPaths { paths: Vec<Vec<Interned<TypoCondition>>> },
+    WordsGraph { graph: RankingRuleGraph<WordsGraph> },
+    WordsPaths { paths: Vec<Vec<Interned<WordsCondition>>> },
     FidGraph { graph: RankingRuleGraph<FidGraph> },
     FidPaths { paths: Vec<Vec<Interned<FidCondition>>> },
     PositionGraph { graph: RankingRuleGraph<PositionGraph> },
@@ -139,8 +140,11 @@ impl SearchLogger<QueryGraph> for VisualSearchLogger {
         let Some(location) = self.location.last() else { return };
         match location {
             Location::Words => {
-                if let Some(query_graph) = state.downcast_ref::<QueryGraph>() {
-                    self.events.push(SearchEvents::WordsGraph { query_graph: query_graph.clone() });
+                if let Some(graph) = state.downcast_ref::<RankingRuleGraph<WordsGraph>>() {
+                    self.events.push(SearchEvents::WordsGraph { graph: graph.clone() });
+                }
+                if let Some(paths) = state.downcast_ref::<Vec<Vec<Interned<WordsCondition>>>>() {
+                    self.events.push(SearchEvents::WordsPaths { paths: paths.clone() });
                 }
             }
             Location::Typo => {
@@ -329,7 +333,6 @@ impl<'ctx> DetailedLoggerFinish<'ctx> {
             SearchEvents::ExtendResults { new } => {
                 self.write_extend_results(new)?;
             }
-            SearchEvents::WordsGraph { query_graph } => self.write_words_graph(query_graph)?,
             SearchEvents::ProximityGraph { graph } => self.write_rr_graph(&graph)?,
             SearchEvents::ProximityPaths { paths } => {
                 self.write_rr_graph_paths::<ProximityGraph>(paths)?;
@@ -338,6 +341,10 @@ impl<'ctx> DetailedLoggerFinish<'ctx> {
             SearchEvents::TypoPaths { paths } => {
                 self.write_rr_graph_paths::<TypoGraph>(paths)?;
             }
+            SearchEvents::WordsGraph { graph } => self.write_rr_graph(&graph)?,
+            SearchEvents::WordsPaths { paths } => {
+                self.write_rr_graph_paths::<WordsGraph>(paths)?;
+            }
             SearchEvents::FidGraph { graph } => self.write_rr_graph(&graph)?,
             SearchEvents::FidPaths { paths } => {
                 self.write_rr_graph_paths::<FidGraph>(paths)?;
@@ -455,7 +462,7 @@ fill: \"#B6E2D3\"
                 shape: class
                 max_nbr_typo: {}",
                     term_subset.description(ctx),
-                    term_subset.max_nbr_typos(ctx)
+                    term_subset.max_typo_cost(ctx)
                 )?;
 
                 for w in term_subset.all_single_words_except_prefix_db(ctx)? {
@@ -482,13 +489,6 @@ fill: \"#B6E2D3\"
         }
         Ok(())
     }
-    fn write_words_graph(&mut self, qg: QueryGraph) -> Result<()> {
-        self.make_new_file_for_internal_state_if_needed()?;
-
-        self.write_query_graph(&qg)?;
-
-        Ok(())
-    }
     fn write_rr_graph<R: RankingRuleGraphTrait>(
         &mut self,
         graph: &RankingRuleGraph<R>,

milli/src/search/new/mod.rs

@@ -15,11 +15,7 @@ mod resolve_query_graph;
 mod small_bitmap;
 
 mod exact_attribute;
-// TODO: documentation + comments
-// implementation is currently an adaptation of the previous implementation to fit with the new model
 mod sort;
-// TODO: documentation + comments
-mod words;
 
 #[cfg(test)]
 mod tests;
@@ -43,10 +39,10 @@ use ranking_rules::{
 use resolve_query_graph::{compute_query_graph_docids, PhraseDocIdsCache};
 use roaring::RoaringBitmap;
 use sort::Sort;
-use words::Words;
 
 use self::geo_sort::GeoSort;
 pub use self::geo_sort::Strategy as GeoSortStrategy;
+use self::graph_based_ranking_rule::Words;
 use self::interner::Interned;
 use crate::search::new::distinct::apply_distinct_rule;
 use crate::{AscDesc, DocumentId, Filter, Index, Member, Result, TermsMatchingStrategy, UserError};
@@ -202,6 +198,11 @@ fn get_ranking_rules_for_query_graph_search<'ctx>(
     let mut sorted_fields = HashSet::new();
     let mut geo_sorted = false;
 
+    // Don't add the `words` ranking rule if the term matching strategy is `All`
+    if matches!(terms_matching_strategy, TermsMatchingStrategy::All) {
+        words = true;
+    }
+
     let mut ranking_rules: Vec<BoxRankingRule<QueryGraph>> = vec![];
     let settings_ranking_rules = ctx.index.criteria(ctx.txn)?;
     for rr in settings_ranking_rules {

milli/src/search/new/query_term/compute_derivations.rs

@@ -28,14 +28,14 @@ pub enum ZeroOrOneTypo {
 impl Interned<QueryTerm> {
     pub fn compute_fully_if_needed(self, ctx: &mut SearchContext) -> Result<()> {
         let s = ctx.term_interner.get_mut(self);
-        if s.max_nbr_typos <= 1 && s.one_typo.is_uninit() {
+        if s.max_levenshtein_distance <= 1 && s.one_typo.is_uninit() {
             assert!(s.two_typo.is_uninit());
             // Initialize one_typo subterm even if max_nbr_typo is 0 because of split words
             self.initialize_one_typo_subterm(ctx)?;
             let s = ctx.term_interner.get_mut(self);
             assert!(s.one_typo.is_init());
             s.two_typo = Lazy::Init(TwoTypoTerm::default());
-        } else if s.max_nbr_typos > 1 && s.two_typo.is_uninit() {
+        } else if s.max_levenshtein_distance > 1 && s.two_typo.is_uninit() {
             assert!(s.two_typo.is_uninit());
             self.initialize_one_and_two_typo_subterm(ctx)?;
             let s = ctx.term_interner.get_mut(self);
@@ -185,7 +185,7 @@ pub fn partially_initialized_term_from_word(
                 original: ctx.word_interner.insert(word.to_owned()),
                 ngram_words: None,
                 is_prefix: false,
-                max_nbr_typos: 0,
+                max_levenshtein_distance: 0,
                 zero_typo: <_>::default(),
                 one_typo: Lazy::Init(<_>::default()),
                 two_typo: Lazy::Init(<_>::default()),
@@ -256,7 +256,7 @@ pub fn partially_initialized_term_from_word(
     Ok(QueryTerm {
         original: word_interned,
         ngram_words: None,
-        max_nbr_typos: max_typo,
+        max_levenshtein_distance: max_typo,
         is_prefix,
         zero_typo,
         one_typo: Lazy::Uninit,
@@ -275,7 +275,16 @@ fn find_split_words(ctx: &mut SearchContext, word: &str) -> Result<Option<Intern
 impl Interned<QueryTerm> {
     fn initialize_one_typo_subterm(self, ctx: &mut SearchContext) -> Result<()> {
         let self_mut = ctx.term_interner.get_mut(self);
-        let QueryTerm { original, is_prefix, one_typo, max_nbr_typos, .. } = self_mut;
+
+        let allows_split_words = self_mut.allows_split_words();
+        let QueryTerm {
+            original,
+            is_prefix,
+            one_typo,
+            max_levenshtein_distance: max_nbr_typos,
+            ..
+        } = self_mut;
+
         let original = *original;
         let is_prefix = *is_prefix;
         // let original_str = ctx.word_interner.get(*original).to_owned();
@@ -300,13 +309,17 @@ impl Interned<QueryTerm> {
             })?;
         }
 
-        let original_str = ctx.word_interner.get(original).to_owned();
-        let split_words = find_split_words(ctx, original_str.as_str())?;
+        let split_words = if allows_split_words {
+            let original_str = ctx.word_interner.get(original).to_owned();
+            find_split_words(ctx, original_str.as_str())?
+        } else {
+            None
+        };
 
         let self_mut = ctx.term_interner.get_mut(self);
 
         // Only add the split words to the derivations if:
-        // 1. the term is not an ngram; OR
+        // 1. the term is neither an ngram nor a phrase; OR
         // 2. the term is an ngram, but the split words are different from the ngram's component words
         let split_words = if let Some((ngram_words, split_words)) =
             self_mut.ngram_words.as_ref().zip(split_words.as_ref())
@@ -328,7 +341,13 @@ impl Interned<QueryTerm> {
     }
     fn initialize_one_and_two_typo_subterm(self, ctx: &mut SearchContext) -> Result<()> {
         let self_mut = ctx.term_interner.get_mut(self);
-        let QueryTerm { original, is_prefix, two_typo, max_nbr_typos, .. } = self_mut;
+        let QueryTerm {
+            original,
+            is_prefix,
+            two_typo,
+            max_levenshtein_distance: max_nbr_typos,
+            ..
+        } = self_mut;
         let original_str = ctx.word_interner.get(*original).to_owned();
         if two_typo.is_init() {
             return Ok(());

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

@@ -43,7 +43,7 @@ pub struct QueryTermSubset {
 pub struct QueryTerm {
     original: Interned<String>,
     ngram_words: Option<Vec<Interned<String>>>,
-    max_nbr_typos: u8,
+    max_levenshtein_distance: u8,
     is_prefix: bool,
     zero_typo: ZeroTypoTerm,
     // May not be computed yet
@@ -342,10 +342,16 @@ impl QueryTermSubset {
         }
         None
     }
-    pub fn max_nbr_typos(&self, ctx: &SearchContext) -> u8 {
+    pub fn max_typo_cost(&self, ctx: &SearchContext) -> u8 {
         let t = ctx.term_interner.get(self.original);
-        match t.max_nbr_typos {
-            0 => 0,
+        match t.max_levenshtein_distance {
+            0 => {
+                if t.allows_split_words() {
+                    1
+                } else {
+                    0
+                }
+            }
             1 => {
                 if self.one_typo_subset.is_empty() {
                     0
@@ -438,6 +444,9 @@ impl QueryTerm {
 
         self.zero_typo.is_empty() && one_typo.is_empty() && two_typo.is_empty()
     }
+    fn allows_split_words(&self) -> bool {
+        self.zero_typo.phrase.is_none()
+    }
 }
 
 impl Interned<QueryTerm> {

milli/src/search/new/query_term/parse_query.rs

@@ -77,13 +77,9 @@ pub fn located_query_terms_from_tokens(
                 }
             }
             TokenKind::Separator(separator_kind) => {
-                match separator_kind {
-                    SeparatorKind::Hard => {
-                        position += 1;
-                    }
-                    SeparatorKind::Soft => {
-                        position += 0;
-                    }
+                // add penalty for hard separators
+                if let SeparatorKind::Hard = separator_kind {
+                    position = position.wrapping_add(1);
                 }
 
                 phrase = 'phrase: {
@@ -217,7 +213,7 @@ pub fn make_ngram(
         original: ngram_str_interned,
         ngram_words: Some(words_interned),
         is_prefix,
-        max_nbr_typos,
+        max_levenshtein_distance: max_nbr_typos,
         zero_typo: term.zero_typo,
         one_typo: Lazy::Uninit,
         two_typo: Lazy::Uninit,
@@ -271,7 +267,7 @@ impl PhraseBuilder {
                 QueryTerm {
                     original: ctx.word_interner.insert(phrase_desc),
                     ngram_words: None,
-                    max_nbr_typos: 0,
+                    max_levenshtein_distance: 0,
                     is_prefix: false,
                     zero_typo: ZeroTypoTerm {
                         phrase: Some(phrase),
@@ -288,3 +284,36 @@ impl PhraseBuilder {
         })
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use charabia::TokenizerBuilder;
+
+    use super::*;
+    use crate::index::tests::TempIndex;
+
+    fn temp_index_with_documents() -> TempIndex {
+        let temp_index = TempIndex::new();
+        temp_index
+            .add_documents(documents!([
+                { "id": 1, "name": "split this world westfali westfalia the Ŵôřlḑôle" },
+                { "id": 2, "name": "Westfália" },
+                { "id": 3, "name": "Ŵôřlḑôle" },
+            ]))
+            .unwrap();
+        temp_index
+    }
+
+    #[test]
+    fn start_with_hard_separator() -> Result<()> {
+        let tokenizer = TokenizerBuilder::new().build();
+        let tokens = tokenizer.tokenize(".");
+        let index = temp_index_with_documents();
+        let rtxn = index.read_txn()?;
+        let mut ctx = SearchContext::new(&index, &rtxn);
+        // panics with `attempt to add with overflow` before <https://github.com/meilisearch/meilisearch/issues/3785>
+        let located_query_terms = located_query_terms_from_tokens(&mut ctx, tokens, None)?;
+        assert!(located_query_terms.is_empty());
+        Ok(())
+    }
+}

milli/src/search/new/ranking_rule_graph/cheapest_paths.rs

@@ -205,18 +205,12 @@ impl<G: RankingRuleGraphTrait> VisitorState<G> {
 impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
     pub fn find_all_costs_to_end(&self) -> MappedInterner<QueryNode, Vec<u64>> {
         let mut costs_to_end = self.query_graph.nodes.map(|_| vec![]);
-        let mut enqueued = SmallBitmap::new(self.query_graph.nodes.len());
 
-        let mut node_stack = VecDeque::new();
-
-        *costs_to_end.get_mut(self.query_graph.end_node) = vec![0];
-
-        for prev_node in self.query_graph.nodes.get(self.query_graph.end_node).predecessors.iter() {
-            node_stack.push_back(prev_node);
-            enqueued.insert(prev_node);
-        }
-
-        while let Some(cur_node) = node_stack.pop_front() {
+        self.traverse_breadth_first_backward(self.query_graph.end_node, |cur_node| {
+            if cur_node == self.query_graph.end_node {
+                *costs_to_end.get_mut(self.query_graph.end_node) = vec![0];
+                return;
+            }
             let mut self_costs = Vec::<u64>::new();
 
             let cur_node_edges = &self.edges_of_node.get(cur_node);
@@ -232,13 +226,7 @@ impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
             self_costs.dedup();
 
             *costs_to_end.get_mut(cur_node) = self_costs;
-            for prev_node in self.query_graph.nodes.get(cur_node).predecessors.iter() {
-                if !enqueued.contains(prev_node) {
-                    node_stack.push_back(prev_node);
-                    enqueued.insert(prev_node);
-                }
-            }
-        }
+        });
         costs_to_end
     }
 
@@ -247,17 +235,12 @@ impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
         node_with_removed_outgoing_conditions: Interned<QueryNode>,
         costs: &mut MappedInterner<QueryNode, Vec<u64>>,
     ) {
-        let mut enqueued = SmallBitmap::new(self.query_graph.nodes.len());
-        let mut node_stack = VecDeque::new();
-
-        enqueued.insert(node_with_removed_outgoing_conditions);
-        node_stack.push_back(node_with_removed_outgoing_conditions);
-
-        'main_loop: while let Some(cur_node) = node_stack.pop_front() {
+        // Traverse the graph backward from the target node, recomputing the cost for each of its predecessors.
+        // We first check that no other node is contributing the same total cost to a predecessor before removing
+        // the cost from the predecessor.
+        self.traverse_breadth_first_backward(node_with_removed_outgoing_conditions, |cur_node| {
             let mut costs_to_remove = FxHashSet::default();
-            for c in costs.get(cur_node) {
-                costs_to_remove.insert(*c);
-            }
+            costs_to_remove.extend(costs.get(cur_node).iter().copied());
 
             let cur_node_edges = &self.edges_of_node.get(cur_node);
             for edge_idx in cur_node_edges.iter() {
@@ -265,22 +248,75 @@ impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
                 for cost in costs.get(edge.dest_node).iter() {
                     costs_to_remove.remove(&(*cost + edge.cost as u64));
                     if costs_to_remove.is_empty() {
-                        continue 'main_loop;
+                        return;
                     }
                 }
             }
             if costs_to_remove.is_empty() {
-                continue 'main_loop;
+                return;
             }
             let mut new_costs = BTreeSet::from_iter(costs.get(cur_node).iter().copied());
             for c in costs_to_remove {
                 new_costs.remove(&c);
             }
             *costs.get_mut(cur_node) = new_costs.into_iter().collect();
+        });
+    }
 
+    /// Traverse the graph backwards from the given node such that every time
+    /// a node is visited, we are guaranteed that all its successors either:
+    /// 1. have already been visited; OR
+    /// 2. were not reachable from the given node
+    pub fn traverse_breadth_first_backward(
+        &self,
+        from: Interned<QueryNode>,
+        mut visit: impl FnMut(Interned<QueryNode>),
+    ) {
+        let mut reachable = SmallBitmap::for_interned_values_in(&self.query_graph.nodes);
+        {
+            // go backward to get the set of all reachable nodes from the given node
+            // the nodes that are not reachable will be set as `visited`
+            let mut stack = VecDeque::new();
+            let mut enqueued = SmallBitmap::for_interned_values_in(&self.query_graph.nodes);
+            enqueued.insert(from);
+            stack.push_back(from);
+            while let Some(n) = stack.pop_front() {
+                if reachable.contains(n) {
+                    continue;
+                }
+                reachable.insert(n);
+                for prev_node in self.query_graph.nodes.get(n).predecessors.iter() {
+                    if !enqueued.contains(prev_node) && !reachable.contains(prev_node) {
+                        stack.push_back(prev_node);
+                        enqueued.insert(prev_node);
+                    }
+                }
+            }
+        };
+        let mut unreachable_or_visited =
+            SmallBitmap::for_interned_values_in(&self.query_graph.nodes);
+        for (n, _) in self.query_graph.nodes.iter() {
+            if !reachable.contains(n) {
+                unreachable_or_visited.insert(n);
+            }
+        }
+
+        let mut enqueued = SmallBitmap::for_interned_values_in(&self.query_graph.nodes);
+        let mut stack = VecDeque::new();
+
+        enqueued.insert(from);
+        stack.push_back(from);
+
+        while let Some(cur_node) = stack.pop_front() {
+            if !self.query_graph.nodes.get(cur_node).successors.is_subset(&unreachable_or_visited) {
+                stack.push_back(cur_node);
+                continue;
+            }
+            unreachable_or_visited.insert(cur_node);
+            visit(cur_node);
             for prev_node in self.query_graph.nodes.get(cur_node).predecessors.iter() {
-                if !enqueued.contains(prev_node) {
-                    node_stack.push_back(prev_node);
+                if !enqueued.contains(prev_node) && !unreachable_or_visited.contains(prev_node) {
+                    stack.push_back(prev_node);
                     enqueued.insert(prev_node);
                 }
             }

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

@@ -20,6 +20,8 @@ mod position;
 mod proximity;
 /// Implementation of the `typo` ranking rule
 mod typo;
+/// Implementation of the `words` ranking rule
+mod words;
 
 use std::collections::BTreeSet;
 use std::hash::Hash;
@@ -33,6 +35,7 @@ pub use position::{PositionCondition, PositionGraph};
 pub use proximity::{ProximityCondition, ProximityGraph};
 use roaring::RoaringBitmap;
 pub use typo::{TypoCondition, TypoGraph};
+pub use words::{WordsCondition, WordsGraph};
 
 use super::interner::{DedupInterner, FixedSizeInterner, Interned, MappedInterner};
 use super::query_term::LocatedQueryTermSubset;

milli/src/search/new/ranking_rule_graph/typo/mod.rs

@@ -50,7 +50,7 @@ impl RankingRuleGraphTrait for TypoGraph {
         // 3-gram -> equivalent to 2 typos
         let base_cost = if term.term_ids.len() == 1 { 0 } else { term.term_ids.len() as u32 };
 
-        for nbr_typos in 0..=term.term_subset.max_nbr_typos(ctx) {
+        for nbr_typos in 0..=term.term_subset.max_typo_cost(ctx) {
             let mut term = term.clone();
             match nbr_typos {
                 0 => {

milli/src/search/new/ranking_rule_graph/words/mod.rs

@@ -0,0 +1,49 @@
+use roaring::RoaringBitmap;
+
+use super::{ComputedCondition, RankingRuleGraphTrait};
+use crate::search::new::interner::{DedupInterner, Interned};
+use crate::search::new::query_term::LocatedQueryTermSubset;
+use crate::search::new::resolve_query_graph::compute_query_term_subset_docids;
+use crate::search::new::SearchContext;
+use crate::Result;
+
+#[derive(Clone, PartialEq, Eq, Hash)]
+pub struct WordsCondition {
+    term: LocatedQueryTermSubset,
+}
+
+pub enum WordsGraph {}
+
+impl RankingRuleGraphTrait for WordsGraph {
+    type Condition = WordsCondition;
+
+    fn resolve_condition(
+        ctx: &mut SearchContext,
+        condition: &Self::Condition,
+        universe: &RoaringBitmap,
+    ) -> Result<ComputedCondition> {
+        let WordsCondition { term, .. } = condition;
+        // maybe compute_query_term_subset_docids should accept a universe as argument
+        let mut docids = compute_query_term_subset_docids(ctx, &term.term_subset)?;
+        docids &= universe;
+
+        Ok(ComputedCondition {
+            docids,
+            universe_len: universe.len(),
+            start_term_subset: None,
+            end_term_subset: term.clone(),
+        })
+    }
+
+    fn build_edges(
+        _ctx: &mut SearchContext,
+        conditions_interner: &mut DedupInterner<Self::Condition>,
+        _from: Option<&LocatedQueryTermSubset>,
+        to_term: &LocatedQueryTermSubset,
+    ) -> Result<Vec<(u32, Interned<Self::Condition>)>> {
+        Ok(vec![(
+            to_term.term_ids.len() as u32,
+            conditions_interner.insert(WordsCondition { term: to_term.clone() }),
+        )])
+    }
+}

milli/src/search/new/words.rs

@@ -1,87 +0,0 @@
-use roaring::RoaringBitmap;
-
-use super::logger::SearchLogger;
-use super::query_graph::QueryNode;
-use super::resolve_query_graph::compute_query_graph_docids;
-use super::small_bitmap::SmallBitmap;
-use super::{QueryGraph, RankingRule, RankingRuleOutput, SearchContext};
-use crate::{Result, TermsMatchingStrategy};
-
-pub struct Words {
-    exhausted: bool, // TODO: remove
-    query_graph: Option<QueryGraph>,
-    nodes_to_remove: Vec<SmallBitmap<QueryNode>>,
-    terms_matching_strategy: TermsMatchingStrategy,
-}
-impl Words {
-    pub fn new(terms_matching_strategy: TermsMatchingStrategy) -> Self {
-        Self {
-            exhausted: true,
-            query_graph: None,
-            nodes_to_remove: vec![],
-            terms_matching_strategy,
-        }
-    }
-}
-
-impl<'ctx> RankingRule<'ctx, QueryGraph> for Words {
-    fn id(&self) -> String {
-        "words".to_owned()
-    }
-    fn start_iteration(
-        &mut self,
-        ctx: &mut SearchContext<'ctx>,
-        _logger: &mut dyn SearchLogger<QueryGraph>,
-        _universe: &RoaringBitmap,
-        parent_query_graph: &QueryGraph,
-    ) -> Result<()> {
-        self.exhausted = false;
-        self.query_graph = Some(parent_query_graph.clone());
-        self.nodes_to_remove = match self.terms_matching_strategy {
-            TermsMatchingStrategy::Last => {
-                let mut ns = parent_query_graph.removal_order_for_terms_matching_strategy_last(ctx);
-                ns.reverse();
-                ns
-            }
-            TermsMatchingStrategy::All => {
-                vec![]
-            }
-        };
-        Ok(())
-    }
-
-    fn next_bucket(
-        &mut self,
-        ctx: &mut SearchContext<'ctx>,
-        logger: &mut dyn SearchLogger<QueryGraph>,
-        universe: &RoaringBitmap,
-    ) -> Result<Option<RankingRuleOutput<QueryGraph>>> {
-        if self.exhausted {
-            return Ok(None);
-        }
-        let Some(query_graph) = &mut self.query_graph else { panic!() };
-        logger.log_internal_state(query_graph);
-
-        let this_bucket = compute_query_graph_docids(ctx, query_graph, universe)?;
-
-        let child_query_graph = query_graph.clone();
-
-        if self.nodes_to_remove.is_empty() {
-            self.exhausted = true;
-        } else {
-            let nodes_to_remove = self.nodes_to_remove.pop().unwrap();
-            query_graph.remove_nodes_keep_edges(&nodes_to_remove.iter().collect::<Vec<_>>());
-        }
-        Ok(Some(RankingRuleOutput { query: child_query_graph, candidates: this_bucket }))
-    }
-
-    fn end_iteration(
-        &mut self,
-        _ctx: &mut SearchContext<'ctx>,
-        _logger: &mut dyn SearchLogger<QueryGraph>,
-    ) {
-        self.exhausted = true;
-        self.nodes_to_remove = vec![];
-        self.query_graph = None;
-    }
-}