|
| 1 | +/* |
| 2 | + * Copyright OpenSearch Contributors |
| 3 | + * SPDX-License-Identifier: Apache-2.0 |
| 4 | + */ |
| 5 | +package org.opensearch.neuralsearch.processor; |
| 6 | + |
| 7 | +import com.google.common.collect.ArrayListMultimap; |
| 8 | +import com.google.common.collect.Multimap; |
| 9 | +import lombok.Getter; |
| 10 | +import lombok.Setter; |
| 11 | +import org.opensearch.action.search.SearchRequest; |
| 12 | +import org.opensearch.common.collect.Tuple; |
| 13 | +import org.opensearch.index.query.BoolQueryBuilder; |
| 14 | +import org.opensearch.index.query.QueryBuilder; |
| 15 | +import org.opensearch.ingest.ConfigurationUtils; |
| 16 | +import org.opensearch.neuralsearch.query.NeuralSparseQueryBuilder; |
| 17 | +import org.opensearch.search.builder.SearchSourceBuilder; |
| 18 | +import org.opensearch.search.pipeline.AbstractProcessor; |
| 19 | +import org.opensearch.search.pipeline.Processor; |
| 20 | +import org.opensearch.search.pipeline.SearchRequestProcessor; |
| 21 | +import org.opensearch.search.rescore.QueryRescorerBuilder; |
| 22 | +import org.opensearch.search.rescore.RescorerBuilder; |
| 23 | + |
| 24 | +import java.util.Collections; |
| 25 | +import java.util.Locale; |
| 26 | +import java.util.Map; |
| 27 | +import java.util.Objects; |
| 28 | +import java.util.stream.Collectors; |
| 29 | + |
| 30 | +/** |
| 31 | + * A SearchRequestProcessor to generate two-phase NeuralSparseQueryBuilder, |
| 32 | + * and add it to the Rescore of a searchRequest. |
| 33 | + */ |
| 34 | +@Setter |
| 35 | +@Getter |
| 36 | +public class NeuralSparseTwoPhaseProcessor extends AbstractProcessor implements SearchRequestProcessor { |
| 37 | + |
| 38 | + public static final String TYPE = "neural_sparse_two_phase_processor"; |
| 39 | + private boolean enabled; |
| 40 | + private float ratio; |
| 41 | + private float windowExpansion; |
| 42 | + private int maxWindowSize; |
| 43 | + private static final String PARAMETER_KEY = "two_phase_parameter"; |
| 44 | + private static final String RATIO_KEY = "prune_ratio"; |
| 45 | + private static final String ENABLE_KEY = "enabled"; |
| 46 | + private static final String EXPANSION_KEY = "expansion_rate"; |
| 47 | + private static final String MAX_WINDOW_SIZE_KEY = "max_window_size"; |
| 48 | + private static final boolean DEFAULT_ENABLED = true; |
| 49 | + private static final float DEFAULT_RATIO = 0.4f; |
| 50 | + private static final float DEFAULT_WINDOW_EXPANSION = 5.0f; |
| 51 | + private static final int DEFAULT_MAX_WINDOW_SIZE = 10000; |
| 52 | + private static final int DEFAULT_BASE_QUERY_SIZE = 10; |
| 53 | + private static final int MAX_WINDOWS_SIZE_LOWER_BOUND = 50; |
| 54 | + private static final float WINDOW_EXPANSION_LOWER_BOUND = 1.0f; |
| 55 | + private static final float RATIO_LOWER_BOUND = 0f; |
| 56 | + private static final float RATIO_UPPER_BOUND = 1f; |
| 57 | + |
| 58 | + protected NeuralSparseTwoPhaseProcessor( |
| 59 | + String tag, |
| 60 | + String description, |
| 61 | + boolean ignoreFailure, |
| 62 | + boolean enabled, |
| 63 | + float ratio, |
| 64 | + float windowExpansion, |
| 65 | + int maxWindowSize |
| 66 | + ) { |
| 67 | + super(tag, description, ignoreFailure); |
| 68 | + this.enabled = enabled; |
| 69 | + if (ratio < RATIO_LOWER_BOUND || ratio > RATIO_UPPER_BOUND) { |
| 70 | + throw new IllegalArgumentException( |
| 71 | + String.format(Locale.ROOT, "The two_phase_parameter.prune_ratio must be within [0, 1]. Received: %f", ratio) |
| 72 | + ); |
| 73 | + } |
| 74 | + this.ratio = ratio; |
| 75 | + if (windowExpansion < WINDOW_EXPANSION_LOWER_BOUND) { |
| 76 | + throw new IllegalArgumentException( |
| 77 | + String.format(Locale.ROOT, "The two_phase_parameter.expansion_rate must >= 1.0. Received: %f", windowExpansion) |
| 78 | + ); |
| 79 | + } |
| 80 | + this.windowExpansion = windowExpansion; |
| 81 | + if (maxWindowSize < MAX_WINDOWS_SIZE_LOWER_BOUND) { |
| 82 | + throw new IllegalArgumentException( |
| 83 | + String.format(Locale.ROOT, "The two_phase_parameter.max_window_size must >= 50. Received: %n" + maxWindowSize) |
| 84 | + ); |
| 85 | + } |
| 86 | + this.maxWindowSize = maxWindowSize; |
| 87 | + } |
| 88 | + |
| 89 | + /** |
| 90 | + * Process the search request of neural_sparse_two_phase_processor |
| 91 | + * @param request the search request (which may have been modified by an earlier processor) |
| 92 | + * @return request the search request that add the two-phase rescore query of neural sparse query. |
| 93 | + */ |
| 94 | + @Override |
| 95 | + public SearchRequest processRequest(final SearchRequest request) { |
| 96 | + if (!enabled || ratio == 0f) { |
| 97 | + return request; |
| 98 | + } |
| 99 | + QueryBuilder queryBuilder = request.source().query(); |
| 100 | + // Collect the nested NeuralSparseQueryBuilder in the whole query. |
| 101 | + Multimap<NeuralSparseQueryBuilder, Float> queryBuilderMap; |
| 102 | + queryBuilderMap = collectNeuralSparseQueryBuilder(queryBuilder, 1.0f); |
| 103 | + if (queryBuilderMap.isEmpty()) { |
| 104 | + return request; |
| 105 | + } |
| 106 | + // Make a nestedQueryBuilder which includes all the two-phase QueryBuilder. |
| 107 | + QueryBuilder nestedTwoPhaseQueryBuilder = getNestedQueryBuilderFromNeuralSparseQueryBuilderMap(queryBuilderMap); |
| 108 | + nestedTwoPhaseQueryBuilder.boost(getOriginQueryWeightAfterRescore(request.source())); |
| 109 | + // Add it to the rescorer. |
| 110 | + RescorerBuilder<QueryRescorerBuilder> twoPhaseRescorer = buildRescoreQueryBuilderForTwoPhase(nestedTwoPhaseQueryBuilder, request); |
| 111 | + request.source().addRescorer(twoPhaseRescorer); |
| 112 | + return request; |
| 113 | + } |
| 114 | + |
| 115 | + @Override |
| 116 | + public String getType() { |
| 117 | + return TYPE; |
| 118 | + } |
| 119 | + |
| 120 | + /** |
| 121 | + * Based on ratio, split a Map into two map by the value. |
| 122 | + * |
| 123 | + * @param queryTokens the queryTokens map, key is the token String, value is the score. |
| 124 | + * @param thresholdRatio The ratio that control how tokens map be split. |
| 125 | + * @return A tuple has two element, { token map whose value above threshold, token map whose value below threshold } |
| 126 | + */ |
| 127 | + public static Tuple<Map<String, Float>, Map<String, Float>> splitQueryTokensByRatioedMaxScoreAsThreshold( |
| 128 | + final Map<String, Float> queryTokens, |
| 129 | + final float thresholdRatio |
| 130 | + ) { |
| 131 | + if (Objects.isNull(queryTokens)) { |
| 132 | + throw new IllegalArgumentException("Query tokens cannot be null or empty."); |
| 133 | + } |
| 134 | + float max = 0f; |
| 135 | + for (Float value : queryTokens.values()) { |
| 136 | + max = Math.max(value, max); |
| 137 | + } |
| 138 | + float threshold = max * thresholdRatio; |
| 139 | + |
| 140 | + Map<Boolean, Map<String, Float>> queryTokensByScore = queryTokens.entrySet() |
| 141 | + .stream() |
| 142 | + .collect( |
| 143 | + Collectors.partitioningBy(entry -> entry.getValue() >= threshold, Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue)) |
| 144 | + ); |
| 145 | + |
| 146 | + Map<String, Float> highScoreTokens = queryTokensByScore.get(Boolean.TRUE); |
| 147 | + Map<String, Float> lowScoreTokens = queryTokensByScore.get(Boolean.FALSE); |
| 148 | + if (Objects.isNull(highScoreTokens)) { |
| 149 | + highScoreTokens = Collections.emptyMap(); |
| 150 | + } |
| 151 | + if (Objects.isNull(lowScoreTokens)) { |
| 152 | + lowScoreTokens = Collections.emptyMap(); |
| 153 | + } |
| 154 | + return Tuple.tuple(highScoreTokens, lowScoreTokens); |
| 155 | + } |
| 156 | + |
| 157 | + private QueryBuilder getNestedQueryBuilderFromNeuralSparseQueryBuilderMap( |
| 158 | + final Multimap<NeuralSparseQueryBuilder, Float> queryBuilderFloatMap |
| 159 | + ) { |
| 160 | + BoolQueryBuilder boolQueryBuilder = new BoolQueryBuilder(); |
| 161 | + queryBuilderFloatMap.asMap().forEach((neuralSparseQueryBuilder, boosts) -> { |
| 162 | + float reduceBoost = boosts.stream().reduce(0.0f, Float::sum); |
| 163 | + boolQueryBuilder.should(neuralSparseQueryBuilder.boost(reduceBoost)); |
| 164 | + }); |
| 165 | + return boolQueryBuilder; |
| 166 | + } |
| 167 | + |
| 168 | + private float getOriginQueryWeightAfterRescore(final SearchSourceBuilder searchSourceBuilder) { |
| 169 | + if (Objects.isNull(searchSourceBuilder.rescores())) { |
| 170 | + return 1.0f; |
| 171 | + } |
| 172 | + return searchSourceBuilder.rescores() |
| 173 | + .stream() |
| 174 | + .map(rescorerBuilder -> ((QueryRescorerBuilder) rescorerBuilder).getQueryWeight()) |
| 175 | + .reduce(1.0f, (a, b) -> a * b); |
| 176 | + } |
| 177 | + |
| 178 | + private Multimap<NeuralSparseQueryBuilder, Float> collectNeuralSparseQueryBuilder(final QueryBuilder queryBuilder, float baseBoost) { |
| 179 | + Multimap<NeuralSparseQueryBuilder, Float> result = ArrayListMultimap.create(); |
| 180 | + |
| 181 | + if (queryBuilder instanceof BoolQueryBuilder) { |
| 182 | + BoolQueryBuilder boolQueryBuilder = (BoolQueryBuilder) queryBuilder; |
| 183 | + float updatedBoost = baseBoost * boolQueryBuilder.boost(); |
| 184 | + for (QueryBuilder subQuery : boolQueryBuilder.should()) { |
| 185 | + Multimap<NeuralSparseQueryBuilder, Float> subResult = collectNeuralSparseQueryBuilder(subQuery, updatedBoost); |
| 186 | + result.putAll(subResult); |
| 187 | + } |
| 188 | + } else if (queryBuilder instanceof NeuralSparseQueryBuilder) { |
| 189 | + NeuralSparseQueryBuilder neuralSparseQueryBuilder = (NeuralSparseQueryBuilder) queryBuilder; |
| 190 | + float updatedBoost = baseBoost * neuralSparseQueryBuilder.boost(); |
| 191 | + /* |
| 192 | + * We obtain a copied modifiedQueryBuilder from the valid origin NeuralSparseQueryBuilder. After this, |
| 193 | + * when the original NeuralSparseQueryBuilder starts to rewrite, it will only retain the tokens that |
| 194 | + * have higher scores (controlled by the maxScore * ratio). The tokens with lower scores will be |
| 195 | + * passed to the modifiedQueryBuilder's queryTokenSupplier. |
| 196 | + * |
| 197 | + * By doing this, we reduce the score computation time for the original NeuralSparseQueryBuilder, |
| 198 | + * and use the modifiedQueryBuilder to make a score increment on TopDocs. |
| 199 | + * |
| 200 | + * When 2-phase is enabled: |
| 201 | + * - Docs besides TopDocs: Score = HighScoreToken's score |
| 202 | + * - Final TopDocs: Score = HighScoreToken's score + LowScoreToken's score |
| 203 | + */ |
| 204 | + NeuralSparseQueryBuilder modifiedQueryBuilder = neuralSparseQueryBuilder.getCopyNeuralSparseQueryBuilderForTwoPhase(ratio); |
| 205 | + result.put(modifiedQueryBuilder, updatedBoost); |
| 206 | + } |
| 207 | + // We only support BoostQuery, BooleanQuery and NeuralSparseQuery now. For other compound query type which are not support now, will |
| 208 | + // do nothing and just quit. |
| 209 | + return result; |
| 210 | + } |
| 211 | + |
| 212 | + private RescorerBuilder<QueryRescorerBuilder> buildRescoreQueryBuilderForTwoPhase( |
| 213 | + final QueryBuilder nestedTwoPhaseQueryBuilder, |
| 214 | + final SearchRequest searchRequest |
| 215 | + ) { |
| 216 | + RescorerBuilder<QueryRescorerBuilder> twoPhaseRescorer = new QueryRescorerBuilder(nestedTwoPhaseQueryBuilder); |
| 217 | + int requestSize = searchRequest.source().size(); |
| 218 | + int windowSize = (int) ((requestSize == -1 ? DEFAULT_BASE_QUERY_SIZE : requestSize) * windowExpansion); |
| 219 | + if (windowSize > maxWindowSize || windowSize < 0) { |
| 220 | + throw new IllegalArgumentException( |
| 221 | + String.format( |
| 222 | + Locale.ROOT, |
| 223 | + "The two-phase window size of neural_sparse_two_phase_processor should be [0,%d], but get the value of %d", |
| 224 | + maxWindowSize, |
| 225 | + windowSize |
| 226 | + ) |
| 227 | + ); |
| 228 | + } |
| 229 | + twoPhaseRescorer.windowSize(windowSize); |
| 230 | + return twoPhaseRescorer; |
| 231 | + } |
| 232 | + |
| 233 | + /** |
| 234 | + * Factory to create NeuralSparseTwoPhaseProcessor, provide default parameter, |
| 235 | + * |
| 236 | + */ |
| 237 | + public static class Factory implements Processor.Factory<SearchRequestProcessor> { |
| 238 | + @Override |
| 239 | + public NeuralSparseTwoPhaseProcessor create( |
| 240 | + Map<String, Processor.Factory<SearchRequestProcessor>> processorFactories, |
| 241 | + String tag, |
| 242 | + String description, |
| 243 | + boolean ignoreFailure, |
| 244 | + Map<String, Object> config, |
| 245 | + PipelineContext pipelineContext |
| 246 | + ) throws IllegalArgumentException { |
| 247 | + |
| 248 | + boolean enabled = ConfigurationUtils.readBooleanProperty(TYPE, tag, config, ENABLE_KEY, DEFAULT_ENABLED); |
| 249 | + Map<String, Object> twoPhaseConfigMap = ConfigurationUtils.readOptionalMap(TYPE, tag, config, PARAMETER_KEY); |
| 250 | + |
| 251 | + float ratio = DEFAULT_RATIO; |
| 252 | + float windowExpansion = DEFAULT_WINDOW_EXPANSION; |
| 253 | + int maxWindowSize = DEFAULT_MAX_WINDOW_SIZE; |
| 254 | + if (Objects.nonNull(twoPhaseConfigMap)) { |
| 255 | + ratio = ((Number) twoPhaseConfigMap.getOrDefault(RATIO_KEY, ratio)).floatValue(); |
| 256 | + windowExpansion = ((Number) twoPhaseConfigMap.getOrDefault(EXPANSION_KEY, windowExpansion)).floatValue(); |
| 257 | + maxWindowSize = ((Number) twoPhaseConfigMap.getOrDefault(MAX_WINDOW_SIZE_KEY, maxWindowSize)).intValue(); |
| 258 | + } |
| 259 | + |
| 260 | + return new NeuralSparseTwoPhaseProcessor(tag, description, ignoreFailure, enabled, ratio, windowExpansion, maxWindowSize); |
| 261 | + } |
| 262 | + } |
| 263 | + |
| 264 | +} |
![opensearch-trigger-bot[bot]](https://avatars.githubusercontent.com/u/80134844?v=4&size=40){kind=avatar}
0 commit comments