[FEATURE] Add rescoreContext for neuralsparse query's two-phase speed up.

CHANGELOG.md

@@ -20,6 +20,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 ## [Unreleased 2.x](https://github.com/opensearch-project/neural-search/compare/2.13...2.x)
 ### Features
 - Support k-NN radial search parameters in neural search([#697](https://github.com/opensearch-project/neural-search/pull/697))
+- Enhance neural_sparse query's latency performance with two-phase rescore query([#695](https://github.com/opensearch-project/neural-search/pull/695)).
 ### Enhancements
 - BWC tests for text chunking processor ([#661](https://github.com/opensearch-project/neural-search/pull/661))
 - Allowing execution of hybrid query on index alias with filters ([#670](https://github.com/opensearch-project/neural-search/pull/670))

src/main/java/org/opensearch/neuralsearch/plugin/NeuralSearch.java

@@ -5,6 +5,10 @@
 package org.opensearch.neuralsearch.plugin;
 
 import static org.opensearch.neuralsearch.settings.NeuralSearchSettings.NEURAL_SEARCH_HYBRID_SEARCH_DISABLED;
+import static org.opensearch.neuralsearch.settings.NeuralSearchSettings.NEURAL_SPARSE_TWO_PHASE_DEFAULT_ENABLED;
+import static org.opensearch.neuralsearch.settings.NeuralSearchSettings.NEURAL_SPARSE_TWO_PHASE_DEFAULT_PRUNING_RATIO;
+import static org.opensearch.neuralsearch.settings.NeuralSearchSettings.NEURAL_SPARSE_TWO_PHASE_DEFAULT_WINDOW_SIZE_EXPANSION;
+import static org.opensearch.neuralsearch.settings.NeuralSearchSettings.NEURAL_SPARSE_TWO_PHASE_MAX_WINDOW_SIZE;
 import static org.opensearch.neuralsearch.settings.NeuralSearchSettings.RERANKER_MAX_DOC_FIELDS;
 
 import java.util.Arrays;
@@ -18,13 +22,15 @@
 import org.opensearch.cluster.metadata.IndexNameExpressionResolver;
 import org.opensearch.cluster.service.ClusterService;
 import org.opensearch.common.settings.Setting;
+import org.opensearch.common.settings.Settings;
 import org.opensearch.common.util.FeatureFlags;
 import org.opensearch.core.common.io.stream.NamedWriteableRegistry;
 import org.opensearch.core.xcontent.NamedXContentRegistry;
 import org.opensearch.env.Environment;
 import org.opensearch.env.NodeEnvironment;
 import org.opensearch.ingest.Processor;
 import org.opensearch.ml.client.MachineLearningNodeClient;
+import org.opensearch.ml.repackage.com.google.common.collect.ImmutableList;
 import org.opensearch.neuralsearch.ml.MLCommonsClientAccessor;
 import org.opensearch.neuralsearch.processor.NeuralQueryEnricherProcessor;
 import org.opensearch.neuralsearch.processor.NormalizationProcessor;
@@ -47,6 +53,7 @@
 import org.opensearch.neuralsearch.query.HybridQueryBuilder;
 import org.opensearch.neuralsearch.query.NeuralQueryBuilder;
 import org.opensearch.neuralsearch.query.NeuralSparseQueryBuilder;
+import org.opensearch.neuralsearch.query.NeuralSparseTwoPhaseParameters;
 import org.opensearch.neuralsearch.query.ext.RerankSearchExtBuilder;
 import org.opensearch.neuralsearch.search.query.HybridQueryPhaseSearcher;
 import org.opensearch.neuralsearch.util.NeuralSearchClusterUtil;
@@ -91,9 +98,11 @@ public Collection<Object> createComponents(
         final IndexNameExpressionResolver indexNameExpressionResolver,
         final Supplier<RepositoriesService> repositoriesServiceSupplier
     ) {
+        Settings settings = environment.settings();
         NeuralSearchClusterUtil.instance().initialize(clusterService);
         NeuralQueryBuilder.initialize(clientAccessor);
         NeuralSparseQueryBuilder.initialize(clientAccessor);
+        NeuralSparseTwoPhaseParameters.initialize(clusterService, settings);
         normalizationProcessorWorkflow = new NormalizationProcessorWorkflow(new ScoreNormalizer(), new ScoreCombiner());
         return List.of(clientAccessor);
     }
@@ -155,7 +164,14 @@ public Map<String, org.opensearch.search.pipeline.Processor.Factory<SearchPhaseR
 
     @Override
     public List<Setting<?>> getSettings() {
-        return List.of(NEURAL_SEARCH_HYBRID_SEARCH_DISABLED, RERANKER_MAX_DOC_FIELDS);
+        return ImmutableList.of(
+            NEURAL_SEARCH_HYBRID_SEARCH_DISABLED,
+            RERANKER_MAX_DOC_FIELDS,
+            NEURAL_SPARSE_TWO_PHASE_DEFAULT_ENABLED,
+            NEURAL_SPARSE_TWO_PHASE_DEFAULT_WINDOW_SIZE_EXPANSION,
+            NEURAL_SPARSE_TWO_PHASE_DEFAULT_PRUNING_RATIO,
+            NEURAL_SPARSE_TWO_PHASE_MAX_WINDOW_SIZE
+        );
     }
 
     @Override
@@ -185,4 +201,5 @@ public List<SearchPlugin.SearchExtSpec<?>> getSearchExts() {
             )
         );
     }
+
 }

src/main/java/org/opensearch/neuralsearch/query/NeuralSparseQuery.java

@@ -0,0 +1,114 @@
+/*
+ * Copyright OpenSearch Contributors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+package org.opensearch.neuralsearch.query;
+
+import java.io.IOException;
+import java.util.Locale;
+import java.util.Objects;
+
+import lombok.AllArgsConstructor;
+import lombok.Getter;
+import lombok.NonNull;
+import org.apache.lucene.search.IndexSearcher;
+import org.apache.lucene.search.Query;
+import org.apache.lucene.search.QueryVisitor;
+import org.apache.lucene.search.ScoreMode;
+import org.apache.lucene.search.Weight;
+
+/**
+ * Implementation of Query interface for type NeuralSparseQuery when TwoPhaseNeuralSparse Enabled.
+ * Initialized, it currentQuery include all tokenQuery. After call setCurrentQueryToHighScoreTokenQuery,
+ * it will perform highScoreTokenQuery.
+ */
+@AllArgsConstructor
+@Getter
+@NonNull
+public final class NeuralSparseQuery extends Query {
+
+    private Query currentQuery;
+    private final Query highScoreTokenQuery;
+    private final Query lowScoreTokenQuery;
+    private final Float rescoreWindowSizeExpansion;
+
+    /**
+     *
+     * @param field The field of this query.
+     * @return String of NeuralSparseQuery object.
+     */
+    @Override
+    public String toString(String field) {
+        return String.format(
+            Locale.ROOT,
+            "NeuralSparseQuery(%s,%s,%s)",
+            currentQuery.toString(field),
+            highScoreTokenQuery.toString(field),
+            lowScoreTokenQuery.toString(field)
+        );
+    }
+
+    public Query rewrite(IndexSearcher indexSearcher) throws IOException {
+        Query rewrittenCurrentQuery = currentQuery.rewrite(indexSearcher);
+        Query rewrittenFirstStepQuery = highScoreTokenQuery.rewrite(indexSearcher);
+        Query rewrittenSecondPhaseQuery = lowScoreTokenQuery.rewrite(indexSearcher);
+        if (rewrittenCurrentQuery == currentQuery
+            && rewrittenFirstStepQuery == highScoreTokenQuery
+            && rewrittenSecondPhaseQuery == lowScoreTokenQuery) {
+            return this;
+        }
+        return new NeuralSparseQuery(rewrittenCurrentQuery, rewrittenFirstStepQuery, rewrittenSecondPhaseQuery, rescoreWindowSizeExpansion);
+    }
+
+    /**
+     * This interface is let the lucene to visit this query.
+     * Briefly, the query to be performed always be a subset of current query.
+     * So in this function use currentQuery.visit().
+     * @param queryVisitor a QueryVisitor to be called by each query in the tree
+     */
+    @Override
+    public void visit(QueryVisitor queryVisitor) {
+        currentQuery.visit(queryVisitor);
+    }
+
+    @Override
+    public boolean equals(Object other) {
+        return sameClassAs(other) && equalsTo(getClass().cast(other));
+    }
+
+    private boolean equalsTo(NeuralSparseQuery other) {
+        return Objects.equals(currentQuery, other.currentQuery)
+            && Objects.equals(highScoreTokenQuery, other.highScoreTokenQuery)
+            && Objects.equals(lowScoreTokenQuery, other.lowScoreTokenQuery)
+            && Objects.equals(rescoreWindowSizeExpansion, other.rescoreWindowSizeExpansion);
+    }
+
+    @Override
+    public int hashCode() {
+        int h = classHash();
+        h = 31 * h + Objects.hashCode(highScoreTokenQuery) + Objects.hashCode(lowScoreTokenQuery) + Objects.hashCode(currentQuery) + Objects
+            .hashCode(rescoreWindowSizeExpansion);
+        return h;
+    }
+
+    /**
+     * This function is always performed after setCurrentQueryToHighScoreTokenQuery. And determine which query's weight to score docs.
+     * @param searcher The searcher that execute the neural_sparse query.
+     * @param scoreMode How the produced scorers will be consumed.
+     * @param boost The boost that is propagated by the parent queries.
+     * @return The weight of currentQuery.
+     * @throws IOException If creteWeight failed.
+     */
+    @Override
+    public Weight createWeight(IndexSearcher searcher, ScoreMode scoreMode, float boost) throws IOException {
+        return currentQuery.createWeight(searcher, scoreMode, boost);
+    }
+
+    /**
+     * Before call this function, the currentQuery of this object is a BooleanQuery include all tokens FeatureFiledQuery.
+     * After call this function, the currentQuery of this object change to a BooleanQuery include high score tokens FeatureFiledQuery.
+     */
+    public void setCurrentQueryToHighScoreTokenQuery() {
+        this.currentQuery = highScoreTokenQuery;
+    }
+}