Have one score definition for cosinesimilarity

Currently we have different score calculation for cosine similarity,
for ex: script score, approximate search, exact search has diffent formula
to convert distance to cosine similarity that is aligned with OpenSearch
score. To keep it consistent, we will be using one defintion which is used
by Lucene as standard definition for cosine similarity for all search types.

Signed-off-by: Vijayan Balasubramanian <[email protected]>

CHANGELOG.md

@@ -28,6 +28,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 * Allow validation for non knn index only after 2.17.0 (#2315)[https://github.com/opensearch-project/k-NN/pull/2315]
 * Release query vector memory after execution (#2346)[https://github.com/opensearch-project/k-NN/pull/2346]
 * Fix shard level rescoring disabled setting flag (#2352)[https://github.com/opensearch-project/k-NN/pull/2352]
+* Update cosine score translation for nmslib (#2357)[https://github.com/opensearch-project/k-NN/pull/2357]
 ### Infrastructure
 * Updated C++ version in JNI from c++11 to c++17 [#2259](https://github.com/opensearch-project/k-NN/pull/2259)
 * Upgrade bytebuddy and objenesis version to match OpenSearch core and, update github ci runner for macos [#2279](https://github.com/opensearch-project/k-NN/pull/2279)

src/main/java/org/opensearch/knn/index/SpaceType.java

@@ -62,7 +62,9 @@ public float scoreToDistanceTranslation(float score) {
     COSINESIMIL("cosinesimil") {
         @Override
         public float scoreTranslation(float rawScore) {
-            return 1 / (1 + rawScore);
+            // To be consistent, we will be using same formula used by lucene as mentioned below
+            // https://github.com/apache/lucene/blob/0494c824e0ac8049b757582f60d085932a890800/lucene/core/src/java/org/apache/lucene/index/VectorSimilarityFunction.java#L73
+            return Math.max((2.0F - rawScore) / 2.0F, 0.0F);
         }
 
         @Override

src/main/java/org/opensearch/knn/index/engine/nmslib/Nmslib.java

@@ -30,8 +30,8 @@ public class Nmslib extends NativeLibrary {
 
     final static Map<String, KNNMethod> METHODS = ImmutableMap.of(METHOD_HNSW, new NmslibHNSWMethod());
 
-    public final static Nmslib INSTANCE = new Nmslib(METHODS, Collections.emptyMap(), CURRENT_VERSION, EXTENSION);
     private final MethodResolver methodResolver;
+    public final static Nmslib INSTANCE = new Nmslib(METHODS, Collections.emptyMap(), CURRENT_VERSION, EXTENSION);
 
     /**
      * Constructor for Nmslib

src/main/java/org/opensearch/knn/plugin/script/KNNScoringSpace.java

@@ -144,7 +144,12 @@ public CosineSimilarity(Object query, MappedFieldType fieldType) {
         protected BiFunction<float[], float[], Float> getScoringMethod(final float[] processedQuery) {
             SpaceType.COSINESIMIL.validateVector(processedQuery);
             float qVectorSquaredMagnitude = getVectorMagnitudeSquared(processedQuery);
-            return (float[] q, float[] v) -> 1 + KNNScoringUtil.cosinesimilOptimized(q, v, qVectorSquaredMagnitude);
+            // To be consistent, we will be using same formula used by lucene as mentioned below
+            // https://github.com/apache/lucene/blob/0494c824e0ac8049b757582f60d085932a890800/lucene/core/src/java/org/apache/lucene/index/VectorSimilarityFunction.java#L73
+            return (float[] q, float[] v) -> Math.max(
+                (1.0F + KNNScoringUtil.cosinesimilOptimized(q, v, qVectorSquaredMagnitude)) / 2.0F,
+                0.0F
+            );
         }
     }
 

src/test/java/org/opensearch/knn/index/NmslibIT.java

@@ -195,6 +195,64 @@ public void testEndToEnd() throws Exception {
         fail("Graphs are not getting evicted");
     }
 
+    public void testEndToEnd_withApproxAndExactSearch_inSameIndex_ForCosineSpaceType() throws Exception {
+        String indexName = "test-index-1";
+        String fieldName = "test-field-1";
+        SpaceType spaceType = SpaceType.COSINESIMIL;
+        Integer dimension = testData.indexData.vectors[0].length;
+
+        // Create an index
+        XContentBuilder builder = XContentFactory.jsonBuilder()
+            .startObject()
+            .startObject("properties")
+            .startObject(fieldName)
+            .field("type", "knn_vector")
+            .field("dimension", dimension)
+            .field(KNNConstants.METHOD_PARAMETER_SPACE_TYPE, spaceType.getValue())
+            .startObject(KNNConstants.KNN_METHOD)
+            .field(KNNConstants.NAME, KNNConstants.METHOD_HNSW)
+            .field(KNNConstants.KNN_ENGINE, KNNEngine.NMSLIB.getName())
+            .endObject()
+            .endObject()
+            .endObject()
+            .endObject();
+
+        Map<String, Object> mappingMap = xContentBuilderToMap(builder);
+        String mapping = builder.toString();
+
+        createKnnIndex(indexName, buildKNNIndexSettings(0), mapping);
+
+        // Index one document
+        addKnnDoc(indexName, randomAlphaOfLength(5), fieldName, Floats.asList(testData.indexData.vectors[0]).toArray());
+
+        // Assert we have the right number of documents in the index
+        refreshAllIndices();
+        assertEquals(1, getDocCount(indexName));
+        // update threshold setting to skip building graph
+        updateIndexSettings(indexName, Settings.builder().put(KNNSettings.INDEX_KNN_ADVANCED_APPROXIMATE_THRESHOLD, -1));
+        // add duplicate document with different id
+        addKnnDoc(indexName, randomAlphaOfLength(5), fieldName, Floats.asList(testData.indexData.vectors[0]).toArray());
+        assertEquals(2, getDocCount(indexName));
+        final int k = 2;
+        // search index
+        Response response = searchKNNIndex(
+            indexName,
+            KNNQueryBuilder.builder().fieldName(fieldName).vector(testData.queries[0]).k(k).build(),
+            k
+        );
+        String responseBody = EntityUtils.toString(response.getEntity());
+        List<KNNResult> knnResults = parseSearchResponse(responseBody, fieldName);
+        assertEquals(k, knnResults.size());
+
+        List<Float> actualScores = parseSearchResponseScore(responseBody, fieldName);
+
+        // both document should have identical score
+        assertEquals(actualScores.get(0), actualScores.get(1), 0.001);
+
+        // Delete index
+        deleteKNNIndex(indexName);
+    }
+
     @SneakyThrows
     private void validateSearch(
         final String indexName,