Support score type threshold in radial search (opensearch-project#1589)

* Support score type threshold in radial search Signed-off-by: Junqiu Lei <[email protected]>
junqiu-lei · Apr 17, 2024 · 244450b · 244450b
1 parent 1887e12
commit 244450b
Show file tree

Hide file tree

Showing 14 changed files with 675 additions and 87 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -16,6 +16,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 ### Features
 * Support distance type radius search for Lucene engine [#1498](https://github.com/opensearch-project/k-NN/pull/1498)
 * Support distance type radius search for Faiss engine [#1546](https://github.com/opensearch-project/k-NN/pull/1546)
+* Support score type threshold in radial search [#1589](https://github.com/opensearch-project/k-NN/pull/1589)
 ### Enhancements
 * Make the HitQueue size more appropriate for exact search [#1549](https://github.com/opensearch-project/k-NN/pull/1549)
 * Support script score when doc value is disabled [#1573](https://github.com/opensearch-project/k-NN/pull/1573)

diff --git a/src/main/java/org/opensearch/knn/index/SpaceType.java b/src/main/java/org/opensearch/knn/index/SpaceType.java
@@ -36,6 +36,14 @@ public float scoreTranslation(float rawScore) {
         public VectorSimilarityFunction getVectorSimilarityFunction() {
             return VectorSimilarityFunction.EUCLIDEAN;
         }
+
+        @Override
+        public float scoreToDistanceTranslation(float score) {
+            if (score == 0) {
+                throw new IllegalArgumentException(String.format(Locale.ROOT, "score cannot be 0 when space type is [%s]", getValue()));
+            }
+            return 1 / score - 1;
+        }
     },
     COSINESIMIL("cosinesimil") {
         @Override
@@ -170,4 +178,14 @@ public static SpaceType getSpace(String spaceTypeName) {
         }
         throw new IllegalArgumentException("Unable to find space: " + spaceTypeName);
     }
+
+    /**
+     * Translate a score to a distance for this space type
+     *
+     * @param score score to translate
+     * @return translated distance
+     */
+    public float scoreToDistanceTranslation(float score) {
+        throw new UnsupportedOperationException(String.format("Space [%s] does not have a score to distance translation", getValue()));
+    }
 }
diff --git a/src/main/java/org/opensearch/knn/index/query/KNNQueryBuilder.java b/src/main/java/org/opensearch/knn/index/query/KNNQueryBuilder.java
@@ -52,6 +52,7 @@ public class KNNQueryBuilder extends AbstractQueryBuilder<KNNQueryBuilder> {
     public static final ParseField FILTER_FIELD = new ParseField("filter");
     public static final ParseField IGNORE_UNMAPPED_FIELD = new ParseField("ignore_unmapped");
     public static final ParseField DISTANCE_FIELD = new ParseField("distance");
+    public static final ParseField SCORE_FIELD = new ParseField("score");
     public static final int K_MAX = 10000;
     /**
      * The name for the knn query
@@ -64,6 +65,7 @@ public class KNNQueryBuilder extends AbstractQueryBuilder<KNNQueryBuilder> {
     private final float[] vector;
     private int k = 0;
     private Float distance = null;
+    private Float score = null;
     private QueryBuilder filter;
     private boolean ignoreUnmapped = false;
 
@@ -92,13 +94,14 @@ public KNNQueryBuilder(String fieldName, float[] vector) {
      *
      * @param k K nearest neighbours for the given vector
      */
-    public KNNQueryBuilder k(int k) {
+    public KNNQueryBuilder k(Integer k) {
+        if (k == null) {
+            throw new IllegalArgumentException("[" + NAME + "] requires k to be set");
+        }
+        validateSingleQueryType(k, distance, score);
         if (k <= 0 || k > K_MAX) {
             throw new IllegalArgumentException("[" + NAME + "] requires 0 < k <= " + K_MAX);
         }
-        if (distance != null) {
-            throw new IllegalArgumentException("[" + NAME + "] requires either k or distance must be set");
-        }
         this.k = k;
         return this;
     }
@@ -112,13 +115,28 @@ public KNNQueryBuilder distance(Float distance) {
         if (distance == null) {
             throw new IllegalArgumentException("[" + NAME + "] requires distance to be set");
         }
-        if (k != 0) {
-            throw new IllegalArgumentException("[" + NAME + "] requires either k or distance must be set");
-        }
+        validateSingleQueryType(k, distance, score);
         this.distance = distance;
         return this;
     }
 
+    /**
+     * Builder method for score
+     *
+     * @param score the score threshold for the nearest neighbours
+     */
+    public KNNQueryBuilder score(Float score) {
+        if (score == null) {
+            throw new IllegalArgumentException("[" + NAME + "] requires score to be set");
+        }
+        validateSingleQueryType(k, distance, score);
+        if (score <= 0) {
+            throw new IllegalArgumentException("[" + NAME + "] requires score greater than 0");
+        }
+        this.score = score;
+        return this;
+    }
+
     /**
      * Builder method for filter
      *
@@ -163,6 +181,7 @@ public KNNQueryBuilder(String fieldName, float[] vector, int k, QueryBuilder fil
         this.filter = filter;
         this.ignoreUnmapped = false;
         this.distance = null;
+        this.score = null;
     }
 
     public static void initialize(ModelDao modelDao) {
@@ -200,6 +219,9 @@ public KNNQueryBuilder(StreamInput in) throws IOException {
             if (isClusterOnOrAfterMinRequiredVersion(KNNConstants.RADIAL_SEARCH_KEY)) {
                 distance = in.readOptionalFloat();
             }
+            if (isClusterOnOrAfterMinRequiredVersion(KNNConstants.RADIAL_SEARCH_KEY)) {
+                score = in.readOptionalFloat();
+            }
         } catch (IOException ex) {
             throw new RuntimeException("[KNN] Unable to create KNNQueryBuilder", ex);
         }
@@ -211,6 +233,7 @@ public static KNNQueryBuilder fromXContent(XContentParser parser) throws IOExcep
         float boost = AbstractQueryBuilder.DEFAULT_BOOST;
         Integer k = null;
         Float distance = null;
+        Float score = null;
         QueryBuilder filter = null;
         String queryName = null;
         String currentFieldName = null;
@@ -241,6 +264,8 @@ public static KNNQueryBuilder fromXContent(XContentParser parser) throws IOExcep
                             queryName = parser.text();
                         } else if (DISTANCE_FIELD.match(currentFieldName, parser.getDeprecationHandler())) {
                             distance = (Float) NumberFieldMapper.NumberType.FLOAT.parse(parser.objectBytes(), false);
+                        } else if (SCORE_FIELD.match(currentFieldName, parser.getDeprecationHandler())) {
+                            score = (Float) NumberFieldMapper.NumberType.FLOAT.parse(parser.objectBytes(), false);
                         } else {
                             throw new ParsingException(
                                 parser.getTokenLocation(),
@@ -270,9 +295,7 @@ public static KNNQueryBuilder fromXContent(XContentParser parser) throws IOExcep
             }
         }
 
-        if ((k != null && distance != null) || (k == null && distance == null)) {
-            throw new IllegalArgumentException("[" + NAME + "] requires either k or distance must be set");
-        }
+        validateSingleQueryType(k, distance, score);
 
         KNNQueryBuilder knnQueryBuilder = new KNNQueryBuilder(fieldName, ObjectsToFloats(vector)).filter(filter)
             .ignoreUnmapped(ignoreUnmapped)
@@ -281,8 +304,10 @@ public static KNNQueryBuilder fromXContent(XContentParser parser) throws IOExcep
 
         if (k != null) {
             knnQueryBuilder.k(k);
-        } else {
+        } else if (distance != null) {
             knnQueryBuilder.distance(distance);
+        } else if (score != null) {
+            knnQueryBuilder.score(score);
         }
 
         return knnQueryBuilder;
@@ -300,6 +325,9 @@ protected void doWriteTo(StreamOutput out) throws IOException {
         if (isClusterOnOrAfterMinRequiredVersion(KNNConstants.RADIAL_SEARCH_KEY)) {
             out.writeOptionalFloat(distance);
         }
+        if (isClusterOnOrAfterMinRequiredVersion(KNNConstants.RADIAL_SEARCH_KEY)) {
+            out.writeOptionalFloat(score);
+        }
     }
 
     /**
@@ -324,6 +352,10 @@ public float getDistance() {
         return this.distance;
     }
 
+    public float getScore() {
+        return this.score;
+    }
+
     public QueryBuilder getFilter() {
         return this.filter;
     }
@@ -358,6 +390,9 @@ public void doXContent(XContentBuilder builder, Params params) throws IOExceptio
         if (ignoreUnmapped) {
             builder.field(IGNORE_UNMAPPED_FIELD.getPreferredName(), ignoreUnmapped);
         }
+        if (score != null) {
+            builder.field(SCORE_FIELD.getPreferredName(), score);
+        }
         printBoostAndQueryName(builder);
         builder.endObject();
         builder.endObject();
@@ -397,8 +432,8 @@ protected Query doToQuery(QueryShardContext context) {
             spaceType = knnMethodContext.getSpaceType();
         }
 
-        // Currently, k-NN supports distance type radius search.
-        // We need transform distance radius to right type of engine required radius.
+        // Currently, k-NN supports distance and score types radial search
+        // We need transform distance/score to right type of engine required radius.
         Float radius = null;
         if (this.distance != null) {
             if (this.distance < 0 && SpaceType.INNER_PRODUCT.equals(spaceType) == false) {
@@ -407,6 +442,13 @@ protected Query doToQuery(QueryShardContext context) {
             radius = knnEngine.distanceToRadialThreshold(this.distance, spaceType);
         }
 
+        if (this.score != null) {
+            if (this.score > 1 && SpaceType.INNER_PRODUCT.equals(spaceType) == false) {
+                throw new IllegalArgumentException("[" + NAME + "] requires score to be in the range (0, 1] for space type: " + spaceType);
+            }
+            radius = knnEngine.scoreToRadialThreshold(this.score, spaceType);
+        }
+
         if (fieldDimension != vector.length) {
             throw new IllegalArgumentException(
                 String.format("Query vector has invalid dimension: %d. Dimension should be: %d", vector.length, fieldDimension)
@@ -464,7 +506,7 @@ protected Query doToQuery(QueryShardContext context) {
                 .build();
             return RNNQueryFactory.create(createQueryRequest);
         }
-        throw new IllegalArgumentException("[" + NAME + "] requires either k or distance must be set");
+        throw new IllegalArgumentException("[" + NAME + "] requires either k or distance or score to be set");
     }
 
     private ModelMetadata getModelMetadataForField(KNNVectorFieldMapper.KNNVectorFieldType knnVectorField) {
@@ -499,4 +541,24 @@ protected int doHashCode() {
     public String getWriteableName() {
         return NAME;
     }
+
+    private static void validateSingleQueryType(Integer k, Float distance, Float score) {
+        int countSetFields = 0;
+
+        if (k != null && k != 0) {
+            countSetFields++;
+        }
+        if (distance != null) {
+            countSetFields++;
+        }
+        if (score != null) {
+            countSetFields++;
+        }
+
+        if (countSetFields != 1) {
+            throw new IllegalArgumentException(
+                "[" + NAME + "] requires only one query type to be set, it can be either k, distance, or score"
+            );
+        }
+    }
 }
diff --git a/src/main/java/org/opensearch/knn/index/util/Faiss.java b/src/main/java/org/opensearch/knn/index/util/Faiss.java
@@ -56,6 +56,7 @@
  * Implements NativeLibrary for the faiss native library
  */
 class Faiss extends NativeLibrary {
+    Map<SpaceType, Function<Float, Float>> scoreTransform;
 
     // TODO: Current version is not really current version. Instead, it encodes information in the file name
     // about the compatibility version the file is created with. In the future, we should refactor this so that it
@@ -68,6 +69,12 @@ class Faiss extends NativeLibrary {
         rawScore -> SpaceType.INNER_PRODUCT.scoreTranslation(-1 * rawScore)
     );
 
+    // Map that overrides radial search score threshold to faiss required distance, check more details in knn documentation:
+    // https://opensearch.org/docs/latest/search-plugins/knn/approximate-knn/#spaces
+    private final static Map<SpaceType, Function<Float, Float>> SCORE_TO_DISTANCE_TRANSFORMATIONS = ImmutableMap.<
+        SpaceType,
+        Function<Float, Float>>builder().put(SpaceType.INNER_PRODUCT, score -> score > 1 ? 1 - score : 1 / score - 1).build();
+
     // Define encoders supported by faiss
     private final static MethodComponentContext ENCODER_DEFAULT = new MethodComponentContext(
         KNNConstants.ENCODER_FLAT,
@@ -301,7 +308,13 @@ class Faiss extends NativeLibrary {
         ).addSpaces(SpaceType.L2, SpaceType.INNER_PRODUCT).build()
     );
 
-    final static Faiss INSTANCE = new Faiss(METHODS, SCORE_TRANSLATIONS, CURRENT_VERSION, KNNConstants.FAISS_EXTENSION);
+    final static Faiss INSTANCE = new Faiss(
+        METHODS,
+        SCORE_TRANSLATIONS,
+        CURRENT_VERSION,
+        KNNConstants.FAISS_EXTENSION,
+        SCORE_TO_DISTANCE_TRANSFORMATIONS
+    );
 
     /**
      * Constructor for Faiss
@@ -315,9 +328,11 @@ private Faiss(
         Map<String, KNNMethod> methods,
         Map<SpaceType, Function<Float, Float>> scoreTranslation,
         String currentVersion,
-        String extension
+        String extension,
+        Map<SpaceType, Function<Float, Float>> scoreTransform
     ) {
         super(methods, scoreTranslation, currentVersion, extension);
+        this.scoreTransform = scoreTransform;
     }
 
     @Override
@@ -326,6 +341,15 @@ public Float distanceToRadialThreshold(Float distance, SpaceType spaceType) {
         return distance;
     }
 
+    @Override
+    public Float scoreToRadialThreshold(Float score, SpaceType spaceType) {
+        // Faiss engine uses distance as is and need transformation
+        if (this.scoreTransform.containsKey(spaceType)) {
+            return this.scoreTransform.get(spaceType).apply(score);
+        }
+        return spaceType.scoreToDistanceTranslation(score);
+    }
+
     /**
      * MethodAsMap builder is used to create the map that will be passed to the jni to create the faiss index.
      * Faiss's index factory takes an "index description" that it uses to build the index. In this description,

diff --git a/src/main/java/org/opensearch/knn/index/util/KNNEngine.java b/src/main/java/org/opensearch/knn/index/util/KNNEngine.java
@@ -158,6 +158,11 @@ public Float distanceToRadialThreshold(Float distance, SpaceType spaceType) {
         return knnLibrary.distanceToRadialThreshold(distance, spaceType);
     }
 
+    @Override
+    public Float scoreToRadialThreshold(Float score, SpaceType spaceType) {
+        return knnLibrary.scoreToRadialThreshold(score, spaceType);
+    }
+
     @Override
     public ValidationException validateMethod(KNNMethodContext knnMethodContext) {
         return knnLibrary.validateMethod(knnMethodContext);

diff --git a/src/main/java/org/opensearch/knn/index/util/KNNLibrary.java b/src/main/java/org/opensearch/knn/index/util/KNNLibrary.java
@@ -78,6 +78,16 @@ public interface KNNLibrary {
      */
     Float distanceToRadialThreshold(Float distance, SpaceType spaceType);
 
+    /**
+     * Translate the score threshold input from end user to the engine's threshold.
+     *
+     * @param score score threshold input from end user
+     * @param spaceType spaceType used to compute the threshold
+     *
+     * @return transformed score for the library
+     */
+    Float scoreToRadialThreshold(Float score, SpaceType spaceType);
+
     /**
      * Validate the knnMethodContext for the given library. A ValidationException should be thrown if the method is
      * deemed invalid.

diff --git a/src/main/java/org/opensearch/knn/index/util/Lucene.java b/src/main/java/org/opensearch/knn/index/util/Lucene.java
@@ -48,11 +48,13 @@ public class Lucene extends JVMLibrary {
         ).addSpaces(SpaceType.L2, SpaceType.COSINESIMIL, SpaceType.INNER_PRODUCT).build()
     );
 
+    // Map that overrides the default distance translations for Lucene, check more details in knn documentation:
+    // https://opensearch.org/docs/latest/search-plugins/knn/approximate-knn/#spaces
     private final static Map<SpaceType, Function<Float, Float>> DISTANCE_TRANSLATIONS = ImmutableMap.<
         SpaceType,
         Function<Float, Float>>builder()
         .put(SpaceType.COSINESIMIL, distance -> (2 - distance) / 2)
-        .put(SpaceType.L2, distance -> 1 / (1 + distance))
+        .put(SpaceType.INNER_PRODUCT, distance -> distance <= 0 ? 1 / (1 - distance) : distance + 1)
         .build();
 
     final static Lucene INSTANCE = new Lucene(METHODS, Version.LATEST.toString(), DISTANCE_TRANSLATIONS);
@@ -90,7 +92,16 @@ public float score(float rawScore, SpaceType spaceType) {
     @Override
     public Float distanceToRadialThreshold(Float distance, SpaceType spaceType) {
         // Lucene requires score threshold to be parameterized when calling the radius search.
-        return this.distanceTransform.get(spaceType).apply(distance);
+        if (this.distanceTransform.containsKey(spaceType)) {
+            return this.distanceTransform.get(spaceType).apply(distance);
+        }
+        return spaceType.scoreTranslation(distance);
+    }
+
+    @Override
+    public Float scoreToRadialThreshold(Float score, SpaceType spaceType) {
+        // Lucene engine uses distance as is and does not need transformation
+        return score;
     }
 
     @Override

diff --git a/src/main/java/org/opensearch/knn/index/util/Nmslib.java b/src/main/java/org/opensearch/knn/index/util/Nmslib.java
@@ -73,4 +73,8 @@ private Nmslib(
     public Float distanceToRadialThreshold(Float distance, SpaceType spaceType) {
         return distance;
     }
+
+    public Float scoreToRadialThreshold(Float score, SpaceType spaceType) {
+        return score;
+    }
 }