[Enhancement] Implement pruning for neural sparse search (#988)

* add impl Signed-off-by: zhichao-aws <[email protected]> * add UT Signed-off-by: zhichao-aws <[email protected]> * rename pruneType; UT Signed-off-by: zhichao-aws <[email protected]> * changelog Signed-off-by: zhichao-aws <[email protected]> * ut Signed-off-by: zhichao-aws <[email protected]> * add it Signed-off-by: zhichao-aws <[email protected]> * change on 2-phase Signed-off-by: zhichao-aws <[email protected]> * UT Signed-off-by: zhichao-aws <[email protected]> * it Signed-off-by: zhichao-aws <[email protected]> * rename Signed-off-by: zhichao-aws <[email protected]> * enhance: more detailed error message Signed-off-by: zhichao-aws <[email protected]> * refactor to prune and split Signed-off-by: zhichao-aws <[email protected]> * changelog Signed-off-by: zhichao-aws <[email protected]> * fix UT cov Signed-off-by: zhichao-aws <[email protected]> * address review comments Signed-off-by: zhichao-aws <[email protected]> * enlarge score diff range Signed-off-by: zhichao-aws <[email protected]> * address comments: check lowScores non null instead of flag Signed-off-by: zhichao-aws <[email protected]> --------- Signed-off-by: zhichao-aws <[email protected]>
opensearch-project · Dec 18, 2024 · e8fe284 · e8fe284
1 parent fba4961
commit e8fe284
Show file tree

Hide file tree

Showing 18 changed files with 1,197 additions and 140 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -19,6 +19,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 ### Enhancements
 - Explainability in hybrid query ([#970](https://github.com/opensearch-project/neural-search/pull/970))
 - Support new knn query parameter expand_nested ([#1013](https://github.com/opensearch-project/neural-search/pull/1013))
+- Implement pruning for neural sparse ingestion pipeline and two phase search processor ([#988](https://github.com/opensearch-project/neural-search/pull/988))
 ### Bug Fixes
 -  Address inconsistent scoring in hybrid query results ([#998](https://github.com/opensearch-project/neural-search/pull/998))
 ### Infrastructure

diff --git a/src/main/java/org/opensearch/neuralsearch/processor/NeuralSparseTwoPhaseProcessor.java b/src/main/java/org/opensearch/neuralsearch/processor/NeuralSparseTwoPhaseProcessor.java
@@ -9,23 +9,22 @@
 import lombok.Getter;
 import lombok.Setter;
 import org.opensearch.action.search.SearchRequest;
-import org.opensearch.common.collect.Tuple;
 import org.opensearch.index.query.BoolQueryBuilder;
 import org.opensearch.index.query.QueryBuilder;
 import org.opensearch.ingest.ConfigurationUtils;
 import org.opensearch.neuralsearch.query.NeuralSparseQueryBuilder;
+import org.opensearch.neuralsearch.util.prune.PruneType;
+import org.opensearch.neuralsearch.util.prune.PruneUtils;
 import org.opensearch.search.builder.SearchSourceBuilder;
 import org.opensearch.search.pipeline.AbstractProcessor;
 import org.opensearch.search.pipeline.Processor;
 import org.opensearch.search.pipeline.SearchRequestProcessor;
 import org.opensearch.search.rescore.QueryRescorerBuilder;
 import org.opensearch.search.rescore.RescorerBuilder;
 
-import java.util.Collections;
 import java.util.Locale;
 import java.util.Map;
 import java.util.Objects;
-import java.util.stream.Collectors;
 
 /**
  * A SearchRequestProcessor to generate two-phase NeuralSparseQueryBuilder,
@@ -37,41 +36,37 @@ public class NeuralSparseTwoPhaseProcessor extends AbstractProcessor implements
 
     public static final String TYPE = "neural_sparse_two_phase_processor";
     private boolean enabled;
-    private float ratio;
+    private float pruneRatio;
+    private PruneType pruneType;
     private float windowExpansion;
     private int maxWindowSize;
     private static final String PARAMETER_KEY = "two_phase_parameter";
-    private static final String RATIO_KEY = "prune_ratio";
     private static final String ENABLE_KEY = "enabled";
     private static final String EXPANSION_KEY = "expansion_rate";
     private static final String MAX_WINDOW_SIZE_KEY = "max_window_size";
     private static final boolean DEFAULT_ENABLED = true;
     private static final float DEFAULT_RATIO = 0.4f;
+    private static final PruneType DEFAULT_PRUNE_TYPE = PruneType.MAX_RATIO;
     private static final float DEFAULT_WINDOW_EXPANSION = 5.0f;
     private static final int DEFAULT_MAX_WINDOW_SIZE = 10000;
     private static final int DEFAULT_BASE_QUERY_SIZE = 10;
     private static final int MAX_WINDOWS_SIZE_LOWER_BOUND = 50;
     private static final float WINDOW_EXPANSION_LOWER_BOUND = 1.0f;
-    private static final float RATIO_LOWER_BOUND = 0f;
-    private static final float RATIO_UPPER_BOUND = 1f;
 
     protected NeuralSparseTwoPhaseProcessor(
         String tag,
         String description,
         boolean ignoreFailure,
         boolean enabled,
-        float ratio,
+        float pruneRatio,
+        PruneType pruneType,
         float windowExpansion,
         int maxWindowSize
     ) {
         super(tag, description, ignoreFailure);
         this.enabled = enabled;
-        if (ratio < RATIO_LOWER_BOUND || ratio > RATIO_UPPER_BOUND) {
-            throw new IllegalArgumentException(
-                String.format(Locale.ROOT, "The two_phase_parameter.prune_ratio must be within [0, 1]. Received: %f", ratio)
-            );
-        }
-        this.ratio = ratio;
+        this.pruneRatio = pruneRatio;
+        this.pruneType = pruneType;
         if (windowExpansion < WINDOW_EXPANSION_LOWER_BOUND) {
             throw new IllegalArgumentException(
                 String.format(Locale.ROOT, "The two_phase_parameter.expansion_rate must >= 1.0. Received: %f", windowExpansion)
@@ -93,7 +88,7 @@ protected NeuralSparseTwoPhaseProcessor(
      */
     @Override
     public SearchRequest processRequest(final SearchRequest request) {
-        if (!enabled || ratio == 0f) {
+        if (!enabled || pruneRatio == 0f) {
             return request;
         }
         QueryBuilder queryBuilder = request.source().query();
@@ -117,43 +112,6 @@ public String getType() {
         return TYPE;
     }
 
-    /**
-     * Based on ratio, split a Map into two map by the value.
-     *
-     * @param queryTokens    the queryTokens map, key is the token String, value is the score.
-     * @param thresholdRatio The ratio that control how tokens map be split.
-     * @return A tuple has two element, { token map whose value above threshold, token map whose value below threshold }
-     */
-    public static Tuple<Map<String, Float>, Map<String, Float>> splitQueryTokensByRatioedMaxScoreAsThreshold(
-        final Map<String, Float> queryTokens,
-        final float thresholdRatio
-    ) {
-        if (Objects.isNull(queryTokens)) {
-            throw new IllegalArgumentException("Query tokens cannot be null or empty.");
-        }
-        float max = 0f;
-        for (Float value : queryTokens.values()) {
-            max = Math.max(value, max);
-        }
-        float threshold = max * thresholdRatio;
-
-        Map<Boolean, Map<String, Float>> queryTokensByScore = queryTokens.entrySet()
-            .stream()
-            .collect(
-                Collectors.partitioningBy(entry -> entry.getValue() >= threshold, Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue))
-            );
-
-        Map<String, Float> highScoreTokens = queryTokensByScore.get(Boolean.TRUE);
-        Map<String, Float> lowScoreTokens = queryTokensByScore.get(Boolean.FALSE);
-        if (Objects.isNull(highScoreTokens)) {
-            highScoreTokens = Collections.emptyMap();
-        }
-        if (Objects.isNull(lowScoreTokens)) {
-            lowScoreTokens = Collections.emptyMap();
-        }
-        return Tuple.tuple(highScoreTokens, lowScoreTokens);
-    }
-
     private QueryBuilder getNestedQueryBuilderFromNeuralSparseQueryBuilderMap(
         final Multimap<NeuralSparseQueryBuilder, Float> queryBuilderFloatMap
     ) {
@@ -201,7 +159,10 @@ private Multimap<NeuralSparseQueryBuilder, Float> collectNeuralSparseQueryBuilde
              *     - Docs besides TopDocs: Score = HighScoreToken's score
              *     - Final TopDocs: Score = HighScoreToken's score + LowScoreToken's score
              */
-            NeuralSparseQueryBuilder modifiedQueryBuilder = neuralSparseQueryBuilder.getCopyNeuralSparseQueryBuilderForTwoPhase(ratio);
+            NeuralSparseQueryBuilder modifiedQueryBuilder = neuralSparseQueryBuilder.getCopyNeuralSparseQueryBuilderForTwoPhase(
+                pruneRatio,
+                pruneType
+            );
             result.put(modifiedQueryBuilder, updatedBoost);
         }
         // We only support BoostQuery, BooleanQuery and NeuralSparseQuery now. For other compound query type which are not support now, will
@@ -248,16 +209,40 @@ public NeuralSparseTwoPhaseProcessor create(
             boolean enabled = ConfigurationUtils.readBooleanProperty(TYPE, tag, config, ENABLE_KEY, DEFAULT_ENABLED);
             Map<String, Object> twoPhaseConfigMap = ConfigurationUtils.readOptionalMap(TYPE, tag, config, PARAMETER_KEY);
 
-            float ratio = DEFAULT_RATIO;
+            float pruneRatio = DEFAULT_RATIO;
             float windowExpansion = DEFAULT_WINDOW_EXPANSION;
             int maxWindowSize = DEFAULT_MAX_WINDOW_SIZE;
+            PruneType pruneType = DEFAULT_PRUNE_TYPE;
             if (Objects.nonNull(twoPhaseConfigMap)) {
-                ratio = ((Number) twoPhaseConfigMap.getOrDefault(RATIO_KEY, ratio)).floatValue();
+                pruneRatio = ((Number) twoPhaseConfigMap.getOrDefault(PruneUtils.PRUNE_RATIO_FIELD, pruneRatio)).floatValue();
                 windowExpansion = ((Number) twoPhaseConfigMap.getOrDefault(EXPANSION_KEY, windowExpansion)).floatValue();
                 maxWindowSize = ((Number) twoPhaseConfigMap.getOrDefault(MAX_WINDOW_SIZE_KEY, maxWindowSize)).intValue();
+                pruneType = PruneType.fromString(
+                    twoPhaseConfigMap.getOrDefault(PruneUtils.PRUNE_TYPE_FIELD, pruneType.getValue()).toString()
+                );
+            }
+            if (!PruneUtils.isValidPruneRatio(pruneType, pruneRatio)) {
+                throw new IllegalArgumentException(
+                    String.format(
+                        Locale.ROOT,
+                        "Illegal prune_ratio %f for prune_type: %s. %s",
+                        pruneRatio,
+                        pruneType.getValue(),
+                        PruneUtils.getValidPruneRatioDescription(pruneType)
+                    )
+                );
             }
 
-            return new NeuralSparseTwoPhaseProcessor(tag, description, ignoreFailure, enabled, ratio, windowExpansion, maxWindowSize);
+            return new NeuralSparseTwoPhaseProcessor(
+                tag,
+                description,
+                ignoreFailure,
+                enabled,
+                pruneRatio,
+                pruneType,
+                windowExpansion,
+                maxWindowSize
+            );
         }
     }
 

diff --git a/src/main/java/org/opensearch/neuralsearch/processor/SparseEncodingProcessor.java b/src/main/java/org/opensearch/neuralsearch/processor/SparseEncodingProcessor.java
@@ -9,14 +9,17 @@
 import java.util.function.BiConsumer;
 import java.util.function.Consumer;
 
+import lombok.Getter;
 import org.opensearch.cluster.service.ClusterService;
 import org.opensearch.core.action.ActionListener;
 import org.opensearch.env.Environment;
 import org.opensearch.ingest.IngestDocument;
 import org.opensearch.neuralsearch.ml.MLCommonsClientAccessor;
+import org.opensearch.neuralsearch.util.prune.PruneType;
 import org.opensearch.neuralsearch.util.TokenWeightUtil;
 
 import lombok.extern.log4j.Log4j2;
+import org.opensearch.neuralsearch.util.prune.PruneUtils;
 
 /**
  * This processor is used for user input data text sparse encoding processing, model_id can be used to indicate which model user use,
@@ -27,18 +30,26 @@ public final class SparseEncodingProcessor extends InferenceProcessor {
 
     public static final String TYPE = "sparse_encoding";
     public static final String LIST_TYPE_NESTED_MAP_KEY = "sparse_encoding";
+    @Getter
+    private final PruneType pruneType;
+    @Getter
+    private final float pruneRatio;
 
     public SparseEncodingProcessor(
         String tag,
         String description,
         int batchSize,
         String modelId,
         Map<String, Object> fieldMap,
+        PruneType pruneType,
+        float pruneRatio,
         MLCommonsClientAccessor clientAccessor,
         Environment environment,
         ClusterService clusterService
     ) {
         super(tag, description, batchSize, TYPE, LIST_TYPE_NESTED_MAP_KEY, modelId, fieldMap, clientAccessor, environment, clusterService);
+        this.pruneType = pruneType;
+        this.pruneRatio = pruneRatio;
     }
 
     @Override
@@ -49,17 +60,23 @@ public void doExecute(
         BiConsumer<IngestDocument, Exception> handler
     ) {
         mlCommonsClientAccessor.inferenceSentencesWithMapResult(this.modelId, inferenceList, ActionListener.wrap(resultMaps -> {
-            setVectorFieldsToDocument(ingestDocument, ProcessMap, TokenWeightUtil.fetchListOfTokenWeightMap(resultMaps));
+            List<Map<String, Float>> sparseVectors = TokenWeightUtil.fetchListOfTokenWeightMap(resultMaps)
+                .stream()
+                .map(vector -> PruneUtils.pruneSparseVector(pruneType, pruneRatio, vector))
+                .toList();
+            setVectorFieldsToDocument(ingestDocument, ProcessMap, sparseVectors);
             handler.accept(ingestDocument, null);
         }, e -> { handler.accept(null, e); }));
     }
 
     @Override
     public void doBatchExecute(List<String> inferenceList, Consumer<List<?>> handler, Consumer<Exception> onException) {
-        mlCommonsClientAccessor.inferenceSentencesWithMapResult(
-            this.modelId,
-            inferenceList,
-            ActionListener.wrap(resultMaps -> handler.accept(TokenWeightUtil.fetchListOfTokenWeightMap(resultMaps)), onException)
-        );
+        mlCommonsClientAccessor.inferenceSentencesWithMapResult(this.modelId, inferenceList, ActionListener.wrap(resultMaps -> {
+            List<Map<String, Float>> sparseVectors = TokenWeightUtil.fetchListOfTokenWeightMap(resultMaps)
+                .stream()
+                .map(vector -> PruneUtils.pruneSparseVector(pruneType, pruneRatio, vector))
+                .toList();
+            handler.accept(sparseVectors);
+        }, onException));
     }
 }
diff --git a/...in/java/org/opensearch/neuralsearch/processor/factory/SparseEncodingProcessorFactory.java b/...in/java/org/opensearch/neuralsearch/processor/factory/SparseEncodingProcessorFactory.java
@@ -6,10 +6,13 @@
 
 import static org.opensearch.ingest.ConfigurationUtils.readMap;
 import static org.opensearch.ingest.ConfigurationUtils.readStringProperty;
-import static org.opensearch.neuralsearch.processor.TextEmbeddingProcessor.TYPE;
+import static org.opensearch.ingest.ConfigurationUtils.readOptionalStringProperty;
+import static org.opensearch.ingest.ConfigurationUtils.readDoubleProperty;
 import static org.opensearch.neuralsearch.processor.TextEmbeddingProcessor.MODEL_ID_FIELD;
 import static org.opensearch.neuralsearch.processor.TextEmbeddingProcessor.FIELD_MAP_FIELD;
+import static org.opensearch.neuralsearch.processor.SparseEncodingProcessor.TYPE;
 
+import java.util.Locale;
 import java.util.Map;
 
 import org.opensearch.cluster.service.ClusterService;
@@ -19,6 +22,8 @@
 import org.opensearch.neuralsearch.processor.SparseEncodingProcessor;
 
 import lombok.extern.log4j.Log4j2;
+import org.opensearch.neuralsearch.util.prune.PruneUtils;
+import org.opensearch.neuralsearch.util.prune.PruneType;
 
 /**
  * Factory for sparse encoding ingest processor for ingestion pipeline. Instantiates processor based on user provided input.
@@ -40,7 +45,40 @@ public SparseEncodingProcessorFactory(MLCommonsClientAccessor clientAccessor, En
     protected AbstractBatchingProcessor newProcessor(String tag, String description, int batchSize, Map<String, Object> config) {
         String modelId = readStringProperty(TYPE, tag, config, MODEL_ID_FIELD);
         Map<String, Object> fieldMap = readMap(TYPE, tag, config, FIELD_MAP_FIELD);
+        // if the field is miss, will return PruneType.None
+        PruneType pruneType = PruneType.fromString(readOptionalStringProperty(TYPE, tag, config, PruneUtils.PRUNE_TYPE_FIELD));
+        float pruneRatio = 0;
+        if (pruneType != PruneType.NONE) {
+            // if we have prune type, then prune ratio field must have value
+            // readDoubleProperty will throw exception if value is not present
+            pruneRatio = readDoubleProperty(TYPE, tag, config, PruneUtils.PRUNE_RATIO_FIELD).floatValue();
+            if (!PruneUtils.isValidPruneRatio(pruneType, pruneRatio)) {
+                throw new IllegalArgumentException(
+                    String.format(
+                        Locale.ROOT,
+                        "Illegal prune_ratio %f for prune_type: %s. %s",
+                        pruneRatio,
+                        pruneType.getValue(),
+                        PruneUtils.getValidPruneRatioDescription(pruneType)
+                    )
+                );
+            }
+        } else if (config.containsKey(PruneUtils.PRUNE_RATIO_FIELD)) {
+            // if we don't have prune type, then prune ratio field must not have value
+            throw new IllegalArgumentException("prune_ratio field is not supported when prune_type is not provided");
+        }
 
-        return new SparseEncodingProcessor(tag, description, batchSize, modelId, fieldMap, clientAccessor, environment, clusterService);
+        return new SparseEncodingProcessor(
+            tag,
+            description,
+            batchSize,
+            modelId,
+            fieldMap,
+            pruneType,
+            pruneRatio,
+            clientAccessor,
+            environment,
+            clusterService
+        );
     }
 }
diff --git a/src/main/java/org/opensearch/neuralsearch/query/NeuralSparseQueryBuilder.java b/src/main/java/org/opensearch/neuralsearch/query/NeuralSparseQueryBuilder.java
@@ -47,8 +47,8 @@
 import lombok.NoArgsConstructor;
 import lombok.Setter;
 import lombok.experimental.Accessors;
-
-import static org.opensearch.neuralsearch.processor.NeuralSparseTwoPhaseProcessor.splitQueryTokensByRatioedMaxScoreAsThreshold;
+import org.opensearch.neuralsearch.util.prune.PruneType;
+import org.opensearch.neuralsearch.util.prune.PruneUtils;
 
 /**
  * SparseEncodingQueryBuilder is responsible for handling "neural_sparse" query types. It uses an ML NEURAL_SPARSE model
@@ -90,6 +90,7 @@ public class NeuralSparseQueryBuilder extends AbstractQueryBuilder<NeuralSparseQ
     // 2. If it's the sub query only build for two-phase, the value will be set to -1 * ratio of processor.
     // Then in the DoToQuery, we can use this to determine which type are this queryBuilder.
     private float twoPhasePruneRatio = 0F;
+    private PruneType twoPhasePruneType = PruneType.NONE;
 
     private static final Version MINIMAL_SUPPORTED_VERSION_DEFAULT_MODEL_ID = Version.V_2_13_0;
 
@@ -129,22 +130,23 @@ public NeuralSparseQueryBuilder(StreamInput in) throws IOException {
 
     /**
      * Copy this QueryBuilder for two phase rescorer, set the copy one's twoPhasePruneRatio to -1.
-     * @param ratio the parameter of the NeuralSparseTwoPhaseProcessor, control how to split the queryTokens to two phase.
+     * @param pruneRatio the parameter of the NeuralSparseTwoPhaseProcessor, control how to split the queryTokens to two phase.
      * @return A copy NeuralSparseQueryBuilder for twoPhase, it will be added to the rescorer.
      */
-    public NeuralSparseQueryBuilder getCopyNeuralSparseQueryBuilderForTwoPhase(float ratio) {
-        this.twoPhasePruneRatio(ratio);
+    public NeuralSparseQueryBuilder getCopyNeuralSparseQueryBuilderForTwoPhase(float pruneRatio, PruneType pruneType) {
+        this.twoPhasePruneRatio(pruneRatio);
+        this.twoPhasePruneType(pruneType);
         NeuralSparseQueryBuilder copy = new NeuralSparseQueryBuilder().fieldName(this.fieldName)
             .queryName(this.queryName)
             .queryText(this.queryText)
             .modelId(this.modelId)
             .maxTokenScore(this.maxTokenScore)
-            .twoPhasePruneRatio(-1f * ratio);
+            .twoPhasePruneRatio(-1f * pruneRatio);
         if (Objects.nonNull(this.queryTokensSupplier)) {
             Map<String, Float> tokens = queryTokensSupplier.get();
             // Splitting tokens based on a threshold value: tokens greater than the threshold are stored in v1,
             // while those less than or equal to the threshold are stored in v2.
-            Tuple<Map<String, Float>, Map<String, Float>> splitTokens = splitQueryTokensByRatioedMaxScoreAsThreshold(tokens, ratio);
+            Tuple<Map<String, Float>, Map<String, Float>> splitTokens = PruneUtils.splitSparseVector(pruneType, pruneRatio, tokens);
             this.queryTokensSupplier(() -> splitTokens.v1());
             copy.queryTokensSupplier(() -> splitTokens.v2());
         } else {
@@ -346,9 +348,10 @@ private BiConsumer<Client, ActionListener<?>> getModelInferenceAsync(SetOnce<Map
             ActionListener.wrap(mapResultList -> {
                 Map<String, Float> queryTokens = TokenWeightUtil.fetchListOfTokenWeightMap(mapResultList).get(0);
                 if (Objects.nonNull(twoPhaseSharedQueryToken)) {
-                    Tuple<Map<String, Float>, Map<String, Float>> splitQueryTokens = splitQueryTokensByRatioedMaxScoreAsThreshold(
-                        queryTokens,
-                        twoPhasePruneRatio
+                    Tuple<Map<String, Float>, Map<String, Float>> splitQueryTokens = PruneUtils.splitSparseVector(
+                        twoPhasePruneType,
+                        twoPhasePruneRatio,
+                        queryTokens
                     );
                     setOnce.set(splitQueryTokens.v1());
                     twoPhaseSharedQueryToken = splitQueryTokens.v2();