Merge #4535

4535: Support Negative Keywords r=ManyTheFish a=Kerollmops

This PR fixes #4422 by supporting `-` before any word in the query.

The minus symbol `-`, from the ASCII table, is not the only character that can be considered the negative operator. You can see the two other matching characters under the `Based on "-" (U+002D)` section on [this unicode reference website](https://www.compart.com/en/unicode/U+002D).

It's important to notice the strange behavior when a query includes and excludes the same word; only the derivative ( synonyms and split) will be kept:
 - If you input `progamer -progamer`, the engine will still search for `pro gamer`.
 - If you have the synonym `like = love` and you input `like -like`, it will still search for `love`.

## TODO
 - [x] Add analytics
 - [x] Add support to the `-` operator
 - [x] Make sure to support spaces around `-` well
 - [x] Support phrase negation
 - [x] Add tests


Co-authored-by: Clément Renault <clement@meilisearch.com>

milli/src/index.rs

@@ -2435,6 +2435,7 @@ pub(crate) mod tests {
             document_scores: _,
             mut documents_ids,
             degraded: _,
+            used_negative_operator: _,
         } = search.execute().unwrap();
         let primary_key_id = index.fields_ids_map(&rtxn).unwrap().id("primary_key").unwrap();
         documents_ids.sort_unstable();

milli/src/search/hybrid.rs

@@ -11,6 +11,7 @@ struct ScoreWithRatioResult {
     candidates: RoaringBitmap,
     document_scores: Vec<(u32, ScoreWithRatio)>,
     degraded: bool,
+    used_negative_operator: bool,
 }
 
 type ScoreWithRatio = (Vec<ScoreDetails>, f32);
@@ -78,6 +79,7 @@ impl ScoreWithRatioResult {
             candidates: results.candidates,
             document_scores,
             degraded: results.degraded,
+            used_negative_operator: results.used_negative_operator,
         }
     }
 
@@ -113,6 +115,7 @@ impl ScoreWithRatioResult {
             documents_ids,
             document_scores,
             degraded: left.degraded | right.degraded,
+            used_negative_operator: left.used_negative_operator | right.used_negative_operator,
         }
     }
 }

milli/src/search/mod.rs

@@ -183,6 +183,7 @@ impl<'a> Search<'a> {
             documents_ids,
             document_scores,
             degraded,
+            used_negative_operator,
         } = match self.vector.as_ref() {
             Some(vector) => execute_vector_search(
                 &mut ctx,
@@ -221,7 +222,14 @@ impl<'a> Search<'a> {
             None => MatchingWords::default(),
         };
 
-        Ok(SearchResult { matching_words, candidates, document_scores, documents_ids, degraded })
+        Ok(SearchResult {
+            matching_words,
+            candidates,
+            document_scores,
+            documents_ids,
+            degraded,
+            used_negative_operator,
+        })
     }
 }
 
@@ -272,6 +280,7 @@ pub struct SearchResult {
     pub documents_ids: Vec<DocumentId>,
     pub document_scores: Vec<Vec<ScoreDetails>>,
     pub degraded: bool,
+    pub used_negative_operator: bool,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]

milli/src/search/new/matches/matching_words.rs

@@ -240,6 +240,7 @@ pub(crate) mod tests {
     use super::super::super::located_query_terms_from_tokens;
     use super::*;
     use crate::index::tests::TempIndex;
+    use crate::search::new::query_term::ExtractedTokens;
 
     pub(crate) fn temp_index_with_documents() -> TempIndex {
         let temp_index = TempIndex::new();
@@ -261,7 +262,8 @@ pub(crate) mod tests {
         let mut builder = TokenizerBuilder::default();
         let tokenizer = builder.build();
         let tokens = tokenizer.tokenize("split this world");
-        let query_terms = located_query_terms_from_tokens(&mut ctx, tokens, None).unwrap();
+        let ExtractedTokens { query_terms, .. } =
+            located_query_terms_from_tokens(&mut ctx, tokens, None).unwrap();
         let matching_words = MatchingWords::new(ctx, query_terms);
 
         assert_eq!(

milli/src/search/new/mod.rs

@@ -33,7 +33,9 @@ use interner::{DedupInterner, Interner};
 pub use logger::visual::VisualSearchLogger;
 pub use logger::{DefaultSearchLogger, SearchLogger};
 use query_graph::{QueryGraph, QueryNode};
-use query_term::{located_query_terms_from_tokens, LocatedQueryTerm, Phrase, QueryTerm};
+use query_term::{
+    located_query_terms_from_tokens, ExtractedTokens, LocatedQueryTerm, Phrase, QueryTerm,
+};
 use ranking_rules::{
     BoxRankingRule, PlaceholderQuery, RankingRule, RankingRuleOutput, RankingRuleQueryTrait,
 };
@@ -209,6 +211,35 @@ fn resolve_universe(
     )
 }
 
+#[tracing::instrument(level = "trace", skip_all, target = "search")]
+fn resolve_negative_words(
+    ctx: &mut SearchContext,
+    negative_words: &[Word],
+) -> Result<RoaringBitmap> {
+    let mut negative_bitmap = RoaringBitmap::new();
+    for &word in negative_words {
+        if let Some(bitmap) = ctx.word_docids(word)? {
+            negative_bitmap |= bitmap;
+        }
+    }
+    Ok(negative_bitmap)
+}
+
+#[tracing::instrument(level = "trace", skip_all, target = "search")]
+fn resolve_negative_phrases(
+    ctx: &mut SearchContext,
+    negative_phrases: &[LocatedQueryTerm],
+) -> Result<RoaringBitmap> {
+    let mut negative_bitmap = RoaringBitmap::new();
+    for term in negative_phrases {
+        let query_term = ctx.term_interner.get(term.value);
+        if let Some(phrase) = query_term.original_phrase() {
+            negative_bitmap |= ctx.get_phrase_docids(phrase)?;
+        }
+    }
+    Ok(negative_bitmap)
+}
+
 /// Return the list of initialised ranking rules to be used for a placeholder search.
 fn get_ranking_rules_for_placeholder_search<'ctx>(
     ctx: &SearchContext<'ctx>,
@@ -557,6 +588,7 @@ pub fn execute_vector_search(
         documents_ids: docids,
         located_query_terms: None,
         degraded,
+        used_negative_operator: false,
     })
 }
 
@@ -580,6 +612,7 @@ pub fn execute_search(
 ) -> Result<PartialSearchResult> {
     check_sort_criteria(ctx, sort_criteria.as_ref())?;
 
+    let mut used_negative_operator = false;
     let mut located_query_terms = None;
     let query_terms = if let Some(query) = query {
         let span = tracing::trace_span!(target: "search::tokens", "tokenizer_builder");
@@ -620,7 +653,16 @@ pub fn execute_search(
         let tokens = tokenizer.tokenize(query);
         drop(entered);
 
-        let query_terms = located_query_terms_from_tokens(ctx, tokens, words_limit)?;
+        let ExtractedTokens { query_terms, negative_words, negative_phrases } =
+            located_query_terms_from_tokens(ctx, tokens, words_limit)?;
+        used_negative_operator = !negative_words.is_empty() || !negative_phrases.is_empty();
+
+        let ignored_documents = resolve_negative_words(ctx, &negative_words)?;
+        let ignored_phrases = resolve_negative_phrases(ctx, &negative_phrases)?;
+
+        universe -= ignored_documents;
+        universe -= ignored_phrases;
+
         if query_terms.is_empty() {
             // Do a placeholder search instead
             None
@@ -630,6 +672,7 @@ pub fn execute_search(
     } else {
         None
     };
+
     let bucket_sort_output = if let Some(query_terms) = query_terms {
         let (graph, new_located_query_terms) = QueryGraph::from_query(ctx, &query_terms)?;
         located_query_terms = Some(new_located_query_terms);
@@ -690,6 +733,7 @@ pub fn execute_search(
         documents_ids: docids,
         located_query_terms,
         degraded,
+        used_negative_operator,
     })
 }
 
@@ -752,4 +796,5 @@ pub struct PartialSearchResult {
     pub document_scores: Vec<Vec<ScoreDetails>>,
 
     pub degraded: bool,
+    pub used_negative_operator: bool,
 }

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

@@ -9,7 +9,9 @@ use std::ops::RangeInclusive;
 
 use either::Either;
 pub use ntypo_subset::NTypoTermSubset;
-pub use parse_query::{located_query_terms_from_tokens, make_ngram, number_of_typos_allowed};
+pub use parse_query::{
+    located_query_terms_from_tokens, make_ngram, number_of_typos_allowed, ExtractedTokens,
+};
 pub use phrase::Phrase;
 
 use super::interner::{DedupInterner, Interned};
@@ -478,6 +480,11 @@ impl QueryTerm {
     pub fn original_word(&self, ctx: &SearchContext) -> String {
         ctx.word_interner.get(self.original).clone()
     }
+
+    pub fn original_phrase(&self) -> Option<Interned<Phrase>> {
+        self.zero_typo.phrase
+    }
+
     pub fn all_computed_derivations(&self) -> (Vec<Interned<String>>, Vec<Interned<Phrase>>) {
         let mut words = BTreeSet::new();
         let mut phrases = BTreeSet::new();

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

@@ -6,20 +6,37 @@ use charabia::{SeparatorKind, TokenKind};
 use super::compute_derivations::partially_initialized_term_from_word;
 use super::{LocatedQueryTerm, ZeroTypoTerm};
 use crate::search::new::query_term::{Lazy, Phrase, QueryTerm};
+use crate::search::new::Word;
 use crate::{Result, SearchContext, MAX_WORD_LENGTH};
 
+#[derive(Clone)]
+/// Extraction of the content of a query.
+pub struct ExtractedTokens {
+    /// The terms to search for in the database.
+    pub query_terms: Vec<LocatedQueryTerm>,
+    /// The words that must not appear in the results.
+    pub negative_words: Vec<Word>,
+    /// The phrases that must not appear in the results.
+    pub negative_phrases: Vec<LocatedQueryTerm>,
+}
+
 /// Convert the tokenised search query into a list of located query terms.
 #[tracing::instrument(level = "trace", skip_all, target = "search::query")]
 pub fn located_query_terms_from_tokens(
     ctx: &mut SearchContext,
     query: NormalizedTokenIter,
     words_limit: Option<usize>,
-) -> Result<Vec<LocatedQueryTerm>> {
+) -> Result<ExtractedTokens> {
     let nbr_typos = number_of_typos_allowed(ctx)?;
 
-    let mut located_terms = Vec::new();
+    let mut query_terms = Vec::new();
 
+    let mut negative_phrase = false;
     let mut phrase: Option<PhraseBuilder> = None;
+    let mut encountered_whitespace = true;
+    let mut negative_next_token = false;
+    let mut negative_words = Vec::new();
+    let mut negative_phrases = Vec::new();
 
     let parts_limit = words_limit.unwrap_or(usize::MAX);
 
@@ -31,9 +48,10 @@ pub fn located_query_terms_from_tokens(
         if token.lemma().is_empty() {
             continue;
         }
+
         // early return if word limit is exceeded
-        if located_terms.len() >= parts_limit {
-            return Ok(located_terms);
+        if query_terms.len() >= parts_limit {
+            return Ok(ExtractedTokens { query_terms, negative_words, negative_phrases });
         }
 
         match token.kind {
@@ -46,6 +64,11 @@ pub fn located_query_terms_from_tokens(
                 // 3. if the word is the last token of the query we push it as a prefix word.
                 if let Some(phrase) = &mut phrase {
                     phrase.push_word(ctx, &token, position)
+                } else if negative_next_token {
+                    let word = token.lemma().to_string();
+                    let word = Word::Original(ctx.word_interner.insert(word));
+                    negative_words.push(word);
+                    negative_next_token = false;
                 } else if peekable.peek().is_some() {
                     match token.kind {
                         TokenKind::Word => {
@@ -61,9 +84,9 @@ pub fn located_query_terms_from_tokens(
                                 value: ctx.term_interner.push(term),
                                 positions: position..=position,
                             };
-                            located_terms.push(located_term);
+                            query_terms.push(located_term);
                         }
-                        TokenKind::StopWord | TokenKind::Separator(_) | TokenKind::Unknown => {}
+                        TokenKind::StopWord | TokenKind::Separator(_) | TokenKind::Unknown => (),
                     }
                 } else {
                     let word = token.lemma();
@@ -78,7 +101,7 @@ pub fn located_query_terms_from_tokens(
                         value: ctx.term_interner.push(term),
                         positions: position..=position,
                     };
-                    located_terms.push(located_term);
+                    query_terms.push(located_term);
                 }
             }
             TokenKind::Separator(separator_kind) => {
@@ -94,7 +117,14 @@ pub fn located_query_terms_from_tokens(
                     let phrase = if separator_kind == SeparatorKind::Hard {
                         if let Some(phrase) = phrase {
                             if let Some(located_query_term) = phrase.build(ctx) {
-                                located_terms.push(located_query_term)
+                                // as we are evaluating a negative operator we put the phrase
+                                // in the negative one *but* we don't reset the negative operator
+                                // as we are immediatly starting a new negative phrase.
+                                if negative_phrase {
+                                    negative_phrases.push(located_query_term);
+                                } else {
+                                    query_terms.push(located_query_term);
+                                }
                             }
                             Some(PhraseBuilder::empty())
                         } else {
@@ -115,26 +145,49 @@ pub fn located_query_terms_from_tokens(
                         // Per the check above, quote_count > 0
                         quote_count -= 1;
                         if let Some(located_query_term) = phrase.build(ctx) {
-                            located_terms.push(located_query_term)
+                            // we were evaluating a negative operator so we
+                            // put the phrase in the negative phrases
+                            if negative_phrase {
+                                negative_phrases.push(located_query_term);
+                                negative_phrase = false;
+                            } else {
+                                query_terms.push(located_query_term);
+                            }
                         }
                     }
 
                     // Start new phrase if the token ends with an opening quote
-                    (quote_count % 2 == 1).then_some(PhraseBuilder::empty())
+                    if quote_count % 2 == 1 {
+                        negative_phrase = negative_next_token;
+                        Some(PhraseBuilder::empty())
+                    } else {
+                        None
+                    }
                 };
+
+                negative_next_token =
+                    phrase.is_none() && token.lemma() == "-" && encountered_whitespace;
             }
             _ => (),
         }
+
+        encountered_whitespace =
+            token.lemma().chars().last().filter(|c| c.is_whitespace()).is_some();
     }
 
     // If a quote is never closed, we consider all of the end of the query as a phrase.
     if let Some(phrase) = phrase.take() {
         if let Some(located_query_term) = phrase.build(ctx) {
-            located_terms.push(located_query_term);
+            // put the phrase in the negative set if we are evaluating a negative operator.
+            if negative_phrase {
+                negative_phrases.push(located_query_term);
+            } else {
+                query_terms.push(located_query_term);
+            }
         }
     }
 
-    Ok(located_terms)
+    Ok(ExtractedTokens { query_terms, negative_words, negative_phrases })
 }
 
 pub fn number_of_typos_allowed<'ctx>(
@@ -315,8 +368,10 @@ mod tests {
         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());
+        let ExtractedTokens { query_terms, .. } =
+            located_query_terms_from_tokens(&mut ctx, tokens, None)?;
+        assert!(query_terms.is_empty());
+
         Ok(())
     }
 }