Work in progress: Bad Typo detection

I have an issue where "speakers" is split into "speaker" and "s",
when I compute the distances for the Typo criterion,
it takes "s" into account and put a distance of zero in the bucket 0
(the "speakers" bucket), therefore it reports any document matching "s"
without typos as best results.

I need to make sure to ignore "s" when its associated part "speaker"
doesn't even exist in the document and is not in the place
it should be ("speaker" followed by "s").

This is hard to think that it will had much computation time to
the Typo criterion like in the previous algorithm where I computed
the real query/words indexes based and removed the invalid ones
before sending the documents to the bucket sort.

meilisearch-core/src/automaton/dfa.rs

@@ -46,3 +46,8 @@ pub fn build_prefix_dfa(query: &str) -> DFA {
 pub fn build_dfa(query: &str) -> DFA {
     build_dfa_with_setting(query, PrefixSetting::NoPrefix)
 }
+
+pub fn build_exact_dfa(query: &str) -> DFA {
+    let builder = LEVDIST0.get_or_init(|| LevBuilder::new(0, true));
+    builder.build_dfa(query)
+}

meilisearch-core/src/automaton/mod.rs

@@ -13,7 +13,7 @@ use crate::database::MainT;
 use crate::error::MResult;
 use crate::store;
 
-pub use self::dfa::{build_dfa, build_prefix_dfa};
+pub use self::dfa::{build_dfa, build_prefix_dfa, build_exact_dfa};
 pub use self::query_enhancer::QueryEnhancer;
 pub use self::query_enhancer::QueryEnhancerBuilder;
 

meilisearch-core/src/bucket_sort.rs

@@ -19,7 +19,7 @@ use slice_group_by::{GroupBy, GroupByMut};
 
 use crate::automaton::NGRAMS;
 use crate::automaton::{QueryEnhancer, QueryEnhancerBuilder};
-use crate::automaton::{build_dfa, build_prefix_dfa};
+use crate::automaton::{build_dfa, build_prefix_dfa, build_exact_dfa};
 use crate::automaton::{normalize_str, split_best_frequency};
 
 use crate::criterion2::*;
@@ -41,6 +41,8 @@ pub fn bucket_sort<'c>(
     let (automatons, query_enhancer) =
         construct_automatons2(reader, query, main_store, postings_lists_store, synonyms_store)?;
 
+    debug!("{:?}", query_enhancer);
+
     let before_postings_lists_fetching = Instant::now();
     mk_arena!(arena);
     let mut bare_matches = fetch_matches(reader, &automatons, &mut arena, main_store, postings_lists_store)?;
@@ -74,7 +76,7 @@ pub fn bucket_sort<'c>(
 
     let criteria = [
         Box::new(Typo) as Box<dyn Criterion>,
-        Box::new(Words),
+        Box::new(Words) as Box<dyn Criterion>,
         Box::new(Proximity),
         Box::new(Attribute),
         Box::new(WordsPosition),
@@ -88,7 +90,7 @@ pub fn bucket_sort<'c>(
 
         for mut group in tmp_groups {
             let before_criterion_preparation = Instant::now();
-            criterion.prepare(&mut group, &mut arena, &query_enhancer);
+            criterion.prepare(&mut group, &mut arena, &query_enhancer, &automatons);
             debug!("{:?} preparation took {:.02?}", criterion.name(), before_criterion_preparation.elapsed());
 
             let before_criterion_sort = Instant::now();
@@ -116,6 +118,7 @@ pub fn bucket_sort<'c>(
             let postings_list = &arena[sm.postings_list];
             let input = postings_list.input();
             let query = &automatons[sm.query_index as usize].query;
+            debug!("{:?} contains {:?}", d.raw_matches[0].document_id, query);
             postings_list.iter().map(move |m| {
                 let covered_area = if query.len() > input.len() {
                     input.len()
@@ -126,6 +129,8 @@ pub fn bucket_sort<'c>(
             })
         }).collect();
 
+        debug!("{:?} contains {:?}", d.raw_matches[0].document_id, d.processed_distances);
+
         Document {
             id: d.raw_matches[0].document_id,
             highlights,
@@ -233,7 +238,7 @@ fn fetch_matches<'txn, 'tag>(
     for (query_index, automaton) in automatons.iter().enumerate() {
         let before_dfa = Instant::now();
         let dfa = automaton.dfa();
-        let QueryWordAutomaton { query, is_exact, is_prefix } = automaton;
+        let QueryWordAutomaton { query, is_exact, is_prefix, .. } = automaton;
         dfa_time += before_dfa.elapsed();
 
         let mut number_of_words = 0;
@@ -294,28 +299,48 @@ fn fetch_matches<'txn, 'tag>(
 
 #[derive(Debug)]
 pub struct QueryWordAutomaton {
-    query: String,
+    pub query: String,
     /// Is it a word that must be considered exact
     /// or is it some derived word (i.e. a synonym)
-    is_exact: bool,
+    pub is_exact: bool,
-    is_prefix: bool,
+    pub is_prefix: bool,
+    /// If it's a phrase query and what is
+    /// its index an the length of the phrase
+    pub phrase_query: Option<(u16, u16)>,
 }
 
 impl QueryWordAutomaton {
     pub fn exact(query: &str) -> QueryWordAutomaton {
-        QueryWordAutomaton { query: query.to_string(), is_exact: true, is_prefix: false }
+        QueryWordAutomaton {
+            query: query.to_string(),
+            is_exact: true,
+            is_prefix: false,
+            phrase_query: None,
+        }
     }
 
     pub fn exact_prefix(query: &str) -> QueryWordAutomaton {
-        QueryWordAutomaton { query: query.to_string(), is_exact: true, is_prefix: true }
+        QueryWordAutomaton {
+            query: query.to_string(),
+            is_exact: true,
+            is_prefix: true,
+            phrase_query: None,
+        }
     }
 
     pub fn non_exact(query: &str) -> QueryWordAutomaton {
-        QueryWordAutomaton { query: query.to_string(), is_exact: false, is_prefix: false }
+        QueryWordAutomaton {
+            query: query.to_string(),
+            is_exact: false,
+            is_prefix: false,
+            phrase_query: None,
+        }
     }
 
     pub fn dfa(&self) -> DFA {
-        if self.is_prefix {
+        if self.phrase_query.is_some() {
+            build_exact_dfa(&self.query)
+        } else if self.is_prefix {
             build_prefix_dfa(&self.query)
         } else {
             build_dfa(&self.query)
@@ -411,16 +436,17 @@ fn construct_automatons2(
 
             if n == 1 {
                 if let Some((left, right)) = split_best_frequency(reader, &normalized, postings_lists_store)? {
-                    let left_automaton = QueryWordAutomaton::exact(left);
+                    let mut left_automaton = QueryWordAutomaton::exact(left);
+                    left_automaton.phrase_query = Some((0, 2));
                     enhancer_builder.declare(query_range.clone(), automaton_index, &[left]);
                     automaton_index += 1;
                     automatons.push(left_automaton);
 
-                    let right_automaton = QueryWordAutomaton::exact(right);
+                    let mut right_automaton = QueryWordAutomaton::exact(right);
+                    right_automaton.phrase_query = Some((1, 2));
                     enhancer_builder.declare(query_range.clone(), automaton_index, &[right]);
                     automaton_index += 1;
                     automatons.push(right_automaton);
-
                 }
             } else {
                 // automaton of concatenation of query words

meilisearch-core/src/criterion2.rs

@@ -5,9 +5,10 @@ use std::sync::atomic::{self, AtomicUsize};
 use slice_group_by::{GroupBy, GroupByMut};
 use compact_arena::SmallArena;
 use sdset::{Set, SetBuf};
+use log::debug;
 
 use crate::{DocIndex, DocumentId};
-use crate::bucket_sort::{BareMatch, SimpleMatch, RawDocument, PostingsListView};
+use crate::bucket_sort::{BareMatch, SimpleMatch, RawDocument, PostingsListView, QueryWordAutomaton};
 use crate::automaton::QueryEnhancer;
 
 type PostingsListsArena<'tag, 'txn> = SmallArena<'tag, PostingsListView<'txn>>;
@@ -20,6 +21,7 @@ pub trait Criterion {
         documents: &mut [RawDocument<'a, 'tag>],
         postings_lists: &mut PostingsListsArena<'tag, 'txn>,
         query_enhancer: &QueryEnhancer,
+        automatons: &[QueryWordAutomaton],
     );
 
     fn evaluate<'a, 'tag, 'txn>(
@@ -77,6 +79,7 @@ impl Criterion for Typo {
         documents: &mut [RawDocument],
         postings_lists: &mut PostingsListsArena,
         query_enhancer: &QueryEnhancer,
+        automatons: &[QueryWordAutomaton],
     ) {
         prepare_query_distances(documents, query_enhancer);
     }
@@ -134,6 +137,7 @@ impl Criterion for Words {
         documents: &mut [RawDocument],
         postings_lists: &mut PostingsListsArena,
         query_enhancer: &QueryEnhancer,
+        automatons: &[QueryWordAutomaton],
     ) {
         prepare_query_distances(documents, query_enhancer);
     }
@@ -161,6 +165,7 @@ fn prepare_raw_matches<'a, 'tag, 'txn>(
     documents: &mut [RawDocument<'a, 'tag>],
     postings_lists: &mut PostingsListsArena<'tag, 'txn>,
     query_enhancer: &QueryEnhancer,
+    automatons: &[QueryWordAutomaton],
 ) {
     for document in documents {
         if !document.processed_matches.is_empty() { continue }
@@ -181,7 +186,7 @@ fn prepare_raw_matches<'a, 'tag, 'txn>(
             }
         }
 
-        let processed = multiword_rewrite_matches(&mut processed, query_enhancer);
+        let processed = multiword_rewrite_matches(&mut processed, query_enhancer, automatons);
         document.processed_matches = processed.into_vec();
     }
 }
@@ -196,8 +201,9 @@ impl Criterion for Proximity {
         documents: &mut [RawDocument<'a, 'tag>],
         postings_lists: &mut PostingsListsArena<'tag, 'txn>,
         query_enhancer: &QueryEnhancer,
+        automatons: &[QueryWordAutomaton],
     ) {
-        prepare_raw_matches(documents, postings_lists, query_enhancer);
+        prepare_raw_matches(documents, postings_lists, query_enhancer, automatons);
     }
 
     fn evaluate<'a, 'tag, 'txn>(
@@ -264,8 +270,9 @@ impl Criterion for Attribute {
         documents: &mut [RawDocument<'a, 'tag>],
         postings_lists: &mut PostingsListsArena<'tag, 'txn>,
         query_enhancer: &QueryEnhancer,
+        automatons: &[QueryWordAutomaton],
     ) {
-        prepare_raw_matches(documents, postings_lists, query_enhancer);
+        prepare_raw_matches(documents, postings_lists, query_enhancer, automatons);
     }
 
     fn evaluate<'a, 'tag, 'txn>(
@@ -276,16 +283,16 @@ impl Criterion for Attribute {
     ) -> Ordering
     {
         #[inline]
-        fn sum_attribute(matches: &[SimpleMatch]) -> usize {
+        fn best_attribute(matches: &[SimpleMatch]) -> u16 {
-            let mut sum_attribute = 0;
+            let mut best_attribute = u16::max_value();
             for group in matches.linear_group_by_key(|bm| bm.query_index) {
-                sum_attribute += group[0].attribute as usize;
+                best_attribute = cmp::min(best_attribute, group[0].attribute);
             }
-            sum_attribute
+            best_attribute
         }
 
-        let lhs = sum_attribute(&lhs.processed_matches);
+        let lhs = best_attribute(&lhs.processed_matches);
-        let rhs = sum_attribute(&rhs.processed_matches);
+        let rhs = best_attribute(&rhs.processed_matches);
 
         lhs.cmp(&rhs)
     }
@@ -301,8 +308,9 @@ impl Criterion for WordsPosition {
         documents: &mut [RawDocument<'a, 'tag>],
         postings_lists: &mut PostingsListsArena<'tag, 'txn>,
         query_enhancer: &QueryEnhancer,
+        automatons: &[QueryWordAutomaton],
     ) {
-        prepare_raw_matches(documents, postings_lists, query_enhancer);
+        prepare_raw_matches(documents, postings_lists, query_enhancer, automatons);
     }
 
     fn evaluate<'a, 'tag, 'txn>(
@@ -338,6 +346,7 @@ impl Criterion for Exact {
         documents: &mut [RawDocument],
         postings_lists: &mut PostingsListsArena,
         query_enhancer: &QueryEnhancer,
+        automatons: &[QueryWordAutomaton],
     ) {
         for document in documents {
             document.raw_matches.sort_unstable_by_key(|bm| (bm.query_index, Reverse(bm.is_exact)));
@@ -379,6 +388,7 @@ impl Criterion for StableDocId {
         documents: &mut [RawDocument],
         postings_lists: &mut PostingsListsArena,
         query_enhancer: &QueryEnhancer,
+        automatons: &[QueryWordAutomaton],
     ) {
         // ...
     }
@@ -398,17 +408,58 @@ impl Criterion for StableDocId {
 }
 
 pub fn multiword_rewrite_matches(
-    matches: &mut [SimpleMatch],
+    simple_matches: &mut [SimpleMatch],
     query_enhancer: &QueryEnhancer,
+    automatons: &[QueryWordAutomaton],
 ) -> SetBuf<SimpleMatch>
 {
-    let mut padded_matches = Vec::with_capacity(matches.len());
+    let mut matches = Vec::with_capacity(simple_matches.len());
 
     // let before_sort = Instant::now();
     // we sort the matches by word index to make them rewritable
-    matches.sort_unstable_by_key(|m| (m.attribute, m.word_index));
+    simple_matches.sort_unstable_by_key(|m| (m.attribute, m.query_index, m.word_index));
     // debug!("sorting dirty matches took {:.02?}", before_sort.elapsed());
 
+    for same_attribute in simple_matches.linear_group_by_key(|m| m.attribute) {
+        let iter = same_attribute.linear_group_by_key(|m| m.query_index);
+        let mut iter = iter.peekable();
+
+        while let Some(same_query_index) = iter.next() {
+            let query_index = same_query_index[0].query_index;
+
+            // TODO we need to support phrase query of longer length
+            if let Some((i, len)) = automatons[query_index as usize].phrase_query {
+                if i != 0 { continue }
+
+                // is the next query_index group the required one
+                if iter.peek().map_or(false, |g| g[0].query_index == query_index + 1) {
+                    if let Some(next) = iter.next() {
+                        for ma in same_query_index {
+                            for mb in next {
+                                if ma.word_index == mb.word_index + 1 {
+                                    matches.push(*ma);
+                                    matches.push(*mb);
+                                }
+                            }
+                        }
+                    }
+                }
+            } else {
+                matches.extend_from_slice(same_query_index);
+            }
+        }
+    }
+
+    // let is_phrase_query = automatons[match_.query_index as usize].phrase_query_len.is_some();
+    // let next_query_index = match_.query_index + 1;
+    // if is_phrase_query && iter.remainder().iter().find(|m| m.query_index == next_query_index).is_none() {
+    //     continue
+    // }
+
+    matches.sort_unstable_by_key(|m| (m.attribute, m.word_index));
+
+    let mut padded_matches = Vec::with_capacity(matches.len());
+
     // let before_padding = Instant::now();
     // for each attribute of each document
     for same_document_attribute in matches.linear_group_by_key(|m| m.attribute) {