[Backport 2.x] Adding Reciprocal Rank Fusion (RRF) in hybrid query

CHANGELOG.md

@@ -16,6 +16,7 @@ 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/2.18...2.x)
 ### Features
 - Pagination in Hybrid query ([#1048](https://github.com/opensearch-project/neural-search/pull/1048))
+- Implement Reciprocal Rank Fusion score normalization/combination technique in hybrid query ([#874](https://github.com/opensearch-project/neural-search/pull/874))
 ### Enhancements
 - Explainability in hybrid query ([#970](https://github.com/opensearch-project/neural-search/pull/970))
 - Implement pruning for neural sparse ingestion pipeline and two phase search processor ([#988](https://github.com/opensearch-project/neural-search/pull/988))

qa/restart-upgrade/src/test/java/org/opensearch/neuralsearch/bwc/KnnRadialSearchIT.java

@@ -69,7 +69,6 @@ private void validateIndexQuery(final String modelId) {
             .modelId(modelId)
             .maxDistance(100000f)
             .build();
-
         Map<String, Object> responseWithMaxDistanceQuery = search(getIndexNameForTest(), neuralQueryBuilderWithMaxDistanceQuery, 1);
         assertNotNull(responseWithMaxDistanceQuery);
     }

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

@@ -30,22 +30,24 @@
 import org.opensearch.neuralsearch.ml.MLCommonsClientAccessor;
 import org.opensearch.neuralsearch.processor.NeuralQueryEnricherProcessor;
 import org.opensearch.neuralsearch.processor.NeuralSparseTwoPhaseProcessor;
-import org.opensearch.neuralsearch.processor.NormalizationProcessor;
 import org.opensearch.neuralsearch.processor.NormalizationProcessorWorkflow;
 import org.opensearch.neuralsearch.processor.ExplanationResponseProcessor;
 import org.opensearch.neuralsearch.processor.SparseEncodingProcessor;
 import org.opensearch.neuralsearch.processor.TextEmbeddingProcessor;
 import org.opensearch.neuralsearch.processor.TextChunkingProcessor;
 import org.opensearch.neuralsearch.processor.TextImageEmbeddingProcessor;
+import org.opensearch.neuralsearch.processor.RRFProcessor;
+import org.opensearch.neuralsearch.processor.NormalizationProcessor;
 import org.opensearch.neuralsearch.processor.combination.ScoreCombinationFactory;
 import org.opensearch.neuralsearch.processor.combination.ScoreCombiner;
 import org.opensearch.neuralsearch.processor.factory.ExplanationResponseProcessorFactory;
 import org.opensearch.neuralsearch.processor.factory.TextChunkingProcessorFactory;
-import org.opensearch.neuralsearch.processor.factory.NormalizationProcessorFactory;
 import org.opensearch.neuralsearch.processor.factory.RerankProcessorFactory;
 import org.opensearch.neuralsearch.processor.factory.SparseEncodingProcessorFactory;
 import org.opensearch.neuralsearch.processor.factory.TextEmbeddingProcessorFactory;
 import org.opensearch.neuralsearch.processor.factory.TextImageEmbeddingProcessorFactory;
+import org.opensearch.neuralsearch.processor.factory.RRFProcessorFactory;
+import org.opensearch.neuralsearch.processor.factory.NormalizationProcessorFactory;
 import org.opensearch.neuralsearch.processor.normalization.ScoreNormalizationFactory;
 import org.opensearch.neuralsearch.processor.normalization.ScoreNormalizer;
 import org.opensearch.neuralsearch.processor.rerank.RerankProcessor;
@@ -157,7 +159,9 @@ public Map<String, org.opensearch.search.pipeline.Processor.Factory<SearchPhaseR
     ) {
         return Map.of(
             NormalizationProcessor.TYPE,
-            new NormalizationProcessorFactory(normalizationProcessorWorkflow, scoreNormalizationFactory, scoreCombinationFactory)
+            new NormalizationProcessorFactory(normalizationProcessorWorkflow, scoreNormalizationFactory, scoreCombinationFactory),
+            RRFProcessor.TYPE,
+            new RRFProcessorFactory(normalizationProcessorWorkflow, scoreNormalizationFactory, scoreCombinationFactory)
         );
     }
 

src/main/java/org/opensearch/neuralsearch/processor/AbstractScoreHybridizationProcessor.java

@@ -0,0 +1,65 @@
+/*
+ * Copyright OpenSearch Contributors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+package org.opensearch.neuralsearch.processor;
+
+import org.opensearch.action.search.SearchPhaseContext;
+import org.opensearch.action.search.SearchPhaseResults;
+import org.opensearch.search.SearchPhaseResult;
+import org.opensearch.search.internal.SearchContext;
+import org.opensearch.search.pipeline.PipelineProcessingContext;
+import org.opensearch.search.pipeline.SearchPhaseResultsProcessor;
+
+import java.util.Optional;
+
+/**
+ * Base class for all score hybridization processors. This class is responsible for executing the score hybridization process.
+ * It is a pipeline processor that is executed after the query phase and before the fetch phase.
+ */
+public abstract class AbstractScoreHybridizationProcessor implements SearchPhaseResultsProcessor {
+    /**
+     * Method abstracts functional aspect of score normalization and score combination. Exact methods for each processing stage
+     * are set as part of class constructor. This method is called when there is no pipeline context
+     * @param searchPhaseResult {@link SearchPhaseResults} DTO that has query search results. Results will be mutated as part of this method execution
+     * @param searchPhaseContext {@link SearchContext}
+     */
+    @Override
+    public <Result extends SearchPhaseResult> void process(
+        final SearchPhaseResults<Result> searchPhaseResult,
+        final SearchPhaseContext searchPhaseContext
+    ) {
+        hybridizeScores(searchPhaseResult, searchPhaseContext, Optional.empty());
+    }
+
+    /**
+     * Method abstracts functional aspect of score normalization and score combination. Exact methods for each processing stage
+     * are set as part of class constructor. This method is called when there is pipeline context
+     * @param searchPhaseResult {@link SearchPhaseResults} DTO that has query search results. Results will be mutated as part of this method execution
+     * @param searchPhaseContext {@link SearchContext}
+     * @param requestContext {@link PipelineProcessingContext} processing context of search pipeline
+     * @param <Result>
+     */
+    @Override
+    public <Result extends SearchPhaseResult> void process(
+        final SearchPhaseResults<Result> searchPhaseResult,
+        final SearchPhaseContext searchPhaseContext,
+        final PipelineProcessingContext requestContext
+    ) {
+        hybridizeScores(searchPhaseResult, searchPhaseContext, Optional.ofNullable(requestContext));
+    }
+
+    /**
+     * Method abstracts functional aspect of score normalization and score combination. Exact methods for each processing stage
+     * are set as part of class constructor
+     * @param searchPhaseResult
+     * @param searchPhaseContext
+     * @param requestContextOptional
+     * @param <Result>
+     */
+    abstract <Result extends SearchPhaseResult> void hybridizeScores(
+        SearchPhaseResults<Result> searchPhaseResult,
+        SearchPhaseContext searchPhaseContext,
+        Optional<PipelineProcessingContext> requestContextOptional
+    );
+}

src/main/java/org/opensearch/neuralsearch/processor/ExplanationResponseProcessor.java

@@ -111,8 +111,9 @@ public SearchResponse processResponse(
                         );
                     }
                     // Create and set final explanation combining all components
+                    Float finalScore = Float.isNaN(searchHit.getScore()) ? 0.0f : searchHit.getScore();
                     Explanation finalExplanation = Explanation.match(
-                        searchHit.getScore(),
+                        finalScore,
                         // combination level explanation is always a single detail
                         combinationExplanation.getScoreDetails().get(0).getValue(),
                         normalizedExplanation

src/main/java/org/opensearch/neuralsearch/processor/InferenceProcessor.java

@@ -319,7 +319,7 @@ Map<String, Object> buildMapWithTargetKeys(IngestDocument ingestDocument) {
                 buildNestedMap(originalKey, targetKey, sourceAndMetadataMap, treeRes);
                 mapWithProcessorKeys.put(originalKey, treeRes.get(originalKey));
             } else {
-                mapWithProcessorKeys.put(String.valueOf(targetKey), normalizeSourceValue(sourceAndMetadataMap.get(originalKey)));
+                mapWithProcessorKeys.put(String.valueOf(targetKey), sourceAndMetadataMap.get(originalKey));
             }
         }
         return mapWithProcessorKeys;
@@ -357,9 +357,8 @@ void buildNestedMap(String parentKey, Object processorKey, Map<String, Object> s
             }
             treeRes.merge(parentKey, next, REMAPPING_FUNCTION);
         } else {
-            Object parentValue = sourceAndMetadataMap.get(parentKey);
             String key = String.valueOf(processorKey);
-            treeRes.put(key, normalizeSourceValue(parentValue));
+            treeRes.put(key, sourceAndMetadataMap.get(parentKey));
         }
     }
 
@@ -404,7 +403,7 @@ private void validateEmbeddingFieldsValue(IngestDocument ingestDocument) {
             indexName,
             clusterService,
             environment,
-            true
+            false
         );
     }
 

src/main/java/org/opensearch/neuralsearch/processor/NormalizationExecuteDTO.java

@@ -0,0 +1,35 @@
+/*
+ * Copyright OpenSearch Contributors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+package org.opensearch.neuralsearch.processor;
+
+import lombok.AllArgsConstructor;
+import lombok.Builder;
+import lombok.Getter;
+import lombok.NonNull;
+import org.opensearch.neuralsearch.processor.combination.ScoreCombinationTechnique;
+import org.opensearch.neuralsearch.processor.normalization.ScoreNormalizationTechnique;
+import org.opensearch.search.fetch.FetchSearchResult;
+import org.opensearch.search.query.QuerySearchResult;
+
+import java.util.List;
+import java.util.Optional;
+
+/**
+ * DTO object to hold data in NormalizationProcessorWorkflow class
+ * in NormalizationProcessorWorkflow.
+ */
+@AllArgsConstructor
+@Builder
+@Getter
+public class NormalizationExecuteDTO {
+    @NonNull
+    private List<QuerySearchResult> querySearchResults;
+    @NonNull
+    private Optional<FetchSearchResult> fetchSearchResultOptional;
+    @NonNull
+    private ScoreNormalizationTechnique normalizationTechnique;
+    @NonNull
+    private ScoreCombinationTechnique combinationTechnique;
+}

src/main/java/org/opensearch/neuralsearch/processor/NormalizationProcessor.java

@@ -19,9 +19,7 @@
 import org.opensearch.neuralsearch.processor.normalization.ScoreNormalizationTechnique;
 import org.opensearch.search.SearchPhaseResult;
 import org.opensearch.search.fetch.FetchSearchResult;
-import org.opensearch.search.internal.SearchContext;
 import org.opensearch.search.pipeline.PipelineProcessingContext;
-import org.opensearch.search.pipeline.SearchPhaseResultsProcessor;
 import org.opensearch.search.query.QuerySearchResult;
 
 import lombok.AllArgsConstructor;
@@ -33,7 +31,7 @@
  */
 @Log4j2
 @AllArgsConstructor
-public class NormalizationProcessor implements SearchPhaseResultsProcessor {
+public class NormalizationProcessor extends AbstractScoreHybridizationProcessor {
     public static final String TYPE = "normalization-processor";
 
     private final String tag;
@@ -42,38 +40,8 @@ public class NormalizationProcessor implements SearchPhaseResultsProcessor {
     private final ScoreCombinationTechnique combinationTechnique;
     private final NormalizationProcessorWorkflow normalizationWorkflow;
 
-    /**
-     * Method abstracts functional aspect of score normalization and score combination. Exact methods for each processing stage
-     * are set as part of class constructor. This method is called when there is no pipeline context
-     * @param searchPhaseResult {@link SearchPhaseResults} DTO that has query search results. Results will be mutated as part of this method execution
-     * @param searchPhaseContext {@link SearchContext}
-     */
     @Override
-    public <Result extends SearchPhaseResult> void process(
-        final SearchPhaseResults<Result> searchPhaseResult,
-        final SearchPhaseContext searchPhaseContext
-    ) {
-        prepareAndExecuteNormalizationWorkflow(searchPhaseResult, searchPhaseContext, Optional.empty());
-    }
-
-    /**
-     * Method abstracts functional aspect of score normalization and score combination. Exact methods for each processing stage
-     * are set as part of class constructor
-     * @param searchPhaseResult {@link SearchPhaseResults} DTO that has query search results. Results will be mutated as part of this method execution
-     * @param searchPhaseContext {@link SearchContext}
-     * @param requestContext {@link PipelineProcessingContext} processing context of search pipeline
-     * @param <Result>
-     */
-    @Override
-    public <Result extends SearchPhaseResult> void process(
-        final SearchPhaseResults<Result> searchPhaseResult,
-        final SearchPhaseContext searchPhaseContext,
-        final PipelineProcessingContext requestContext
-    ) {
-        prepareAndExecuteNormalizationWorkflow(searchPhaseResult, searchPhaseContext, Optional.ofNullable(requestContext));
-    }
-
-    private <Result extends SearchPhaseResult> void prepareAndExecuteNormalizationWorkflow(
+    <Result extends SearchPhaseResult> void hybridizeScores(
         SearchPhaseResults<Result> searchPhaseResult,
         SearchPhaseContext searchPhaseContext,
         Optional<PipelineProcessingContext> requestContextOptional

src/main/java/org/opensearch/neuralsearch/processor/NormalizationProcessorWorkflow.java

@@ -22,14 +22,12 @@
 import org.opensearch.action.search.SearchPhaseContext;
 import org.opensearch.common.lucene.search.TopDocsAndMaxScore;
 import org.opensearch.neuralsearch.processor.combination.CombineScoresDto;
-import org.opensearch.neuralsearch.processor.combination.ScoreCombinationTechnique;
 import org.opensearch.neuralsearch.processor.combination.ScoreCombiner;
 import org.opensearch.neuralsearch.processor.explain.CombinedExplanationDetails;
 import org.opensearch.neuralsearch.processor.explain.DocIdAtSearchShard;
 import org.opensearch.neuralsearch.processor.explain.ExplanationDetails;
 import org.opensearch.neuralsearch.processor.explain.ExplainableTechnique;
 import org.opensearch.neuralsearch.processor.explain.ExplanationPayload;
-import org.opensearch.neuralsearch.processor.normalization.ScoreNormalizationTechnique;
 import org.opensearch.neuralsearch.processor.normalization.ScoreNormalizer;
 import org.opensearch.search.SearchHit;
 import org.opensearch.search.SearchHits;
@@ -57,44 +55,30 @@ public class NormalizationProcessorWorkflow {
 
     /**
      * Start execution of this workflow
-     * @param querySearchResults input data with QuerySearchResult from multiple shards
-     * @param normalizationTechnique technique for score normalization
-     * @param combinationTechnique technique for score combination
+     * @param request contains querySearchResults input data with QuerySearchResult
+     * from multiple shards, fetchSearchResultOptional, normalizationTechnique technique for score normalization
+     *  combinationTechnique technique for score combination, and nullable rankConstant only used in RRF technique
      */
-    public void execute(
-        final List<QuerySearchResult> querySearchResults,
-        final Optional<FetchSearchResult> fetchSearchResultOptional,
-        final ScoreNormalizationTechnique normalizationTechnique,
-        final ScoreCombinationTechnique combinationTechnique,
-        final SearchPhaseContext searchPhaseContext
-    ) {
-        NormalizationProcessorWorkflowExecuteRequest request = NormalizationProcessorWorkflowExecuteRequest.builder()
-            .querySearchResults(querySearchResults)
-            .fetchSearchResultOptional(fetchSearchResultOptional)
-            .normalizationTechnique(normalizationTechnique)
-            .combinationTechnique(combinationTechnique)
-            .explain(false)
-            .searchPhaseContext(searchPhaseContext)
-            .build();
-        execute(request);
-    }
-
     public void execute(final NormalizationProcessorWorkflowExecuteRequest request) {
         List<QuerySearchResult> querySearchResults = request.getQuerySearchResults();
         Optional<FetchSearchResult> fetchSearchResultOptional = request.getFetchSearchResultOptional();
-
-        // save original state
-        List<Integer> unprocessedDocIds = unprocessedDocIds(querySearchResults);
+        List<Integer> unprocessedDocIds = unprocessedDocIds(request.getQuerySearchResults());
 
         // pre-process data
         log.debug("Pre-process query results");
         List<CompoundTopDocs> queryTopDocs = getQueryTopDocs(querySearchResults);
 
         explain(request, queryTopDocs);
 
+        // Data transfer object for score normalization used to pass nullable rankConstant which is only used in RRF
+        NormalizeScoresDTO normalizeScoresDTO = NormalizeScoresDTO.builder()
+            .queryTopDocs(queryTopDocs)
+            .normalizationTechnique(request.getNormalizationTechnique())
+            .build();
+
         // normalize
         log.debug("Do score normalization");
-        scoreNormalizer.normalizeScores(queryTopDocs, request.getNormalizationTechnique());
+        scoreNormalizer.normalizeScores(normalizeScoresDTO);
 
         CombineScoresDto combineScoresDTO = CombineScoresDto.builder()
             .queryTopDocs(queryTopDocs)

src/main/java/org/opensearch/neuralsearch/processor/NormalizeScoresDTO.java

@@ -0,0 +1,26 @@
+/*
+ * Copyright OpenSearch Contributors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+package org.opensearch.neuralsearch.processor;
+
+import lombok.AllArgsConstructor;
+import lombok.Builder;
+import lombok.Getter;
+import lombok.NonNull;
+import org.opensearch.neuralsearch.processor.normalization.ScoreNormalizationTechnique;
+
+import java.util.List;
+
+/**
+ * DTO object to hold data required for score normalization.
+ */
+@AllArgsConstructor
+@Builder
+@Getter
+public class NormalizeScoresDTO {
+    @NonNull
+    private List<CompoundTopDocs> queryTopDocs;
+    @NonNull
+    private ScoreNormalizationTechnique normalizationTechnique;
+}