Implement words as a graph-based ranking rule and fix some bugs

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 {

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>() {
+                if let Some(graph) = state.downcast_ref::<RankingRuleGraph<WordsGraph>>() {
-                    self.events.push(SearchEvents::WordsGraph { query_graph: query_graph.clone() });
+                    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)?;
@@ -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/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();
+        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];
+                *costs_to_end.get_mut(self.query_graph.end_node) = vec![0];
-
+                return true;
-        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() {
             let mut self_costs = Vec::<u64>::new();
 
             let cur_node_edges = &self.edges_of_node.get(cur_node);
@@ -232,13 +226,8 @@ 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() {
+            true
-                if !enqueued.contains(prev_node) {
+        });
-                    node_stack.push_back(prev_node);
-                    enqueued.insert(prev_node);
-                }
-            }
-        }
         costs_to_end
     }
 
@@ -247,17 +236,9 @@ 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());
+        self.traverse_breadth_first_backward(node_with_removed_outgoing_conditions, |cur_node| {
-        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() {
             let mut costs_to_remove = FxHashSet::default();
-            for c in costs.get(cur_node) {
+            costs_to_remove.extend(costs.get(cur_node).iter().copied());
-                costs_to_remove.insert(*c);
-            }
 
             let cur_node_edges = &self.edges_of_node.get(cur_node);
             for edge_idx in cur_node_edges.iter() {
@@ -265,23 +246,79 @@ 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 false;
                     }
                 }
             }
             if costs_to_remove.is_empty() {
-                continue 'main_loop;
+                return false;
             }
             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();
+            true
+        });
+    }
 
-            for prev_node in self.query_graph.nodes.get(cur_node).predecessors.iter() {
+    /// Traverse the graph backwards from the given node such that every time
-                if !enqueued.contains(prev_node) {
+    /// a node is visited, we are guaranteed that all its successors either:
-                    node_stack.push_back(prev_node);
+    /// 1. have already been visited; OR
-                    enqueued.insert(prev_node);
+    /// 2. were not reachable from the given node
+    pub fn traverse_breadth_first_backward(
+        &self,
+        from: Interned<QueryNode>,
+        mut visit: impl FnMut(Interned<QueryNode>) -> bool,
+    ) {
+        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);
+            if visit(cur_node) {
+                for prev_node in self.query_graph.nodes.get(cur_node).predecessors.iter() {
+                    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/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;
-    }
-}