[9.1][Vector Search] BugFix wrong vector docvalue_fields

docs/changelog/137862.yaml

@@ -0,0 +1,5 @@
+pr: 137862
+summary: "[Vector Search] Fix  wrong vector docvalue_fields"
+area: Vector Search
+type: bug
+issues: []

rest-api-spec/src/yamlRestTest/resources/rest-api-spec/test/search.vectors/200_dense_vector_docvalue_fields.yml

@@ -161,3 +161,77 @@ setup:
   - close_to: { hits.hits.2.fields.vector2.0.4: { value: -100.0, error: 0.001 } }
 
   - match: {hits.hits.2.fields.vector3.0: [-1, 100, -13, 15, -128]}
+
+---
+"Dense vector cosine docvalue_fields with sparse documents after force merge":
+  - requires:
+      cluster_features: [ "mapper.fix_dense_vector_wrong_fields" ]
+      reason: Support for dense vector doc value fields capability required
+  - do:
+      indices.create:
+        index: test_cosine
+        body:
+          mappings:
+            properties:
+              foo:
+                type: keyword
+              vec:
+                type: dense_vector
+                similarity: cosine
+
+  # Create sparse vector scenario - some docs have vectors, some don't
+  - do:
+      index:
+        index: test_cosine
+        id: "1"
+        body:
+          foo: "bar"
+
+  - do:
+      index:
+        index: test_cosine
+        id: "2"
+        body:
+          foo: "bar"
+
+  - do:
+      index:
+        index: test_cosine
+        id: "3"
+        body:
+          vec: [1, 2]
+
+  - do:
+      indices.refresh:
+        index: test_cosine
+
+  - do:
+      index:
+        index: test_cosine
+        id: "4"
+        body:
+          foo: "bar"
+
+  # Force merge to create the ord/docId mapping issue scenario
+  - do:
+      indices.forcemerge:
+        index: test_cosine
+        max_num_segments: 1
+
+  # Search with docvalue_fields to get the vector values
+  - do:
+      search:
+        index: test_cosine
+        body:
+          docvalue_fields: ["vec"]
+          query:
+            term:
+              _id: "3"
+
+  # Verify that _id=3's vector value is correctly returned as [1, 2]
+  # This test verifies that the ordToDoc fix works correctly
+  - match: { hits.total.value: 1 }
+  - match: { hits.hits.0._id: "3" }
+  - length: { hits.hits.0.fields.vec.0: 2 }
+  - close_to: { hits.hits.0.fields.vec.0.0: { value: 1.0, error: 0.001 } }
+  - close_to: { hits.hits.0.fields.vec.0.1: { value: 2.0, error: 0.001 } }

server/src/main/java/org/elasticsearch/index/mapper/MapperFeatures.java

@@ -49,6 +49,7 @@ public class MapperFeatures implements FeatureSpecification {
     static final NodeFeature SEARCH_LOAD_PER_SHARD = new NodeFeature("mapper.search_load_per_shard");
     static final NodeFeature PATTERNED_TEXT = new NodeFeature("mapper.patterned_text");
     public static final NodeFeature MULTI_FIELD_UNICODE_OPTIMISATION_FIX = new NodeFeature("mapper.multi_field.unicode_optimisation_fix");
+    public static final NodeFeature FIX_DENSE_VECTOR_WRONG_FIELDS = new NodeFeature("mapper.fix_dense_vector_wrong_fields");
 
     @Override
     public Set<NodeFeature> getTestFeatures() {
@@ -83,7 +84,8 @@ public Set<NodeFeature> getTestFeatures() {
             SPARSE_VECTOR_INDEX_OPTIONS_FEATURE,
             PATTERNED_TEXT,
             MULTI_FIELD_UNICODE_OPTIMISATION_FIX,
-            MATCH_ONLY_TEXT_BLOCK_LOADER_FIX
+            MATCH_ONLY_TEXT_BLOCK_LOADER_FIX,
+            FIX_DENSE_VECTOR_WRONG_FIELDS
         );
     }
 }

server/src/main/java/org/elasticsearch/index/mapper/vectors/DenormalizedCosineFloatVectorValues.java

@@ -59,6 +59,11 @@ public VectorScorer scorer(float[] floats) throws IOException {
         return in.scorer(floats);
     }
 
+    @Override
+    public int ordToDoc(int ord) {
+        return in.ordToDoc(ord);
+    }
+
     public float magnitude() {
         return magnitude;
     }

server/src/test/java/org/elasticsearch/index/mapper/vectors/DenormalizedCosineFloatVectorValuesTests.java

@@ -9,8 +9,10 @@
 
 package org.elasticsearch.index.mapper.vectors;
 
+import org.apache.lucene.index.FloatVectorValues;
 import org.apache.lucene.index.KnnVectorValues;
 import org.apache.lucene.index.NumericDocValues;
+import org.apache.lucene.search.VectorScorer;
 import org.elasticsearch.test.ESTestCase;
 
 import java.io.IOException;
@@ -114,4 +116,197 @@ public long cost() {
         };
     }
 
+    public void testOrdToDocWithSparseVectors() throws IOException {
+        // This test simulates a real-world scenario where some documents don't have vector fields
+        // After force merge, the ord (ordinal) and docId mapping becomes crucial
+
+        // Simulate a scenario where we have 6 documents, but only documents 1, 3, 4 have vectors
+        // Doc 0: no vector
+        // Doc 1: vector [0.6, 0.8, 0.0, 0.0], magnitude 5.0 -> original [3.0, 4.0, 0.0, 0.0]
+        // Doc 2: no vector
+        // Doc 3: vector [1.0, 0.0, 0.0, 0.0], magnitude 2.0 -> original [2.0, 0.0, 0.0, 0.0]
+        // Doc 4: vector [0.0, 0.0, 0.6, 0.8], magnitude 10.0 -> original [0.0, 0.0, 6.0, 8.0]
+        // Doc 5: no vector
+
+        // After merge, the vector ordinals will be 0, 1, 2 but they correspond to docIds 1, 3, 4
+        int totalDocs = 6;
+        int[] docIdsWithVectors = { 1, 3, 4 }; // Document IDs that have vectors
+        int numVectors = docIdsWithVectors.length;
+
+        float[][] normalizedVectors = new float[numVectors][];
+        float[] magnitudes = new float[numVectors];
+
+        normalizedVectors[0] = new float[] { 0.6f, 0.8f, 0.0f, 0.0f }; // Doc 1
+        magnitudes[0] = 5.0f;
+
+        normalizedVectors[1] = new float[] { 1.0f, 0.0f, 0.0f, 0.0f }; // Doc 3
+        magnitudes[1] = 2.0f;
+
+        normalizedVectors[2] = new float[] { 0.0f, 0.0f, 0.6f, 0.8f }; // Doc 4
+        magnitudes[2] = 10.0f;
+
+        // Expected original vectors after denormalization
+        float[][] expectedVectors = new float[numVectors][];
+        expectedVectors[0] = new float[] { 3.0f, 4.0f, 0.0f, 0.0f }; // Doc 1
+        expectedVectors[1] = new float[] { 2.0f, 0.0f, 0.0f, 0.0f }; // Doc 3
+        expectedVectors[2] = new float[] { 0.0f, 0.0f, 6.0f, 8.0f }; // Doc 4
+
+        // Create a custom FloatVectorValues that simulates post-merge sparse vector scenario
+        FloatVectorValues sparseVectorValues = new FloatVectorValues() {
+            @Override
+            public int dimension() {
+                return 4;
+            }
+
+            @Override
+            public int size() {
+                return numVectors;
+            }
+
+            @Override
+            public DocIndexIterator iterator() {
+                return new DocIndexIterator() {
+                    private int index = -1;
+
+                    @Override
+                    public int docID() {
+                        return index;
+                    }
+
+                    @Override
+                    public int index() {
+                        return index;
+                    }
+
+                    @Override
+                    public int nextDoc() {
+                        return advance(index + 1);
+                    }
+
+                    @Override
+                    public int advance(int target) {
+                        if (target >= numVectors) return NO_MORE_DOCS;
+                        return index = target;
+                    }
+
+                    @Override
+                    public long cost() {
+                        return numVectors;
+                    }
+                };
+            }
+
+            @Override
+            public FloatVectorValues copy() {
+                throw new UnsupportedOperationException();
+            }
+
+            @Override
+            public VectorScorer scorer(float[] floats) {
+                throw new UnsupportedOperationException();
+            }
+
+            // This is the key method - it maps ordinals to actual document IDs
+            @Override
+            public int ordToDoc(int ord) {
+                // ord 0 -> docId 1, ord 1 -> docId 3, ord 2 -> docId 4
+                return docIdsWithVectors[ord];
+            }
+
+            @Override
+            public float[] vectorValue(int ord) {
+                return normalizedVectors[ord];
+            }
+        };
+
+        // Create magnitudes that correspond to the actual document IDs
+        NumericDocValues sparseMagnitudes = new NumericDocValues() {
+            private int docId = -1;
+
+            @Override
+            public long longValue() {
+                // Find which vector index corresponds to this docId
+                for (int i = 0; i < docIdsWithVectors.length; i++) {
+                    if (docIdsWithVectors[i] == docId) {
+                        return Float.floatToRawIntBits(magnitudes[i]);
+                    }
+                }
+                return Float.floatToRawIntBits(1.0f); // Default magnitude
+            }
+
+            @Override
+            public boolean advanceExact(int target) {
+                docId = target;
+                // Check if this docId has a vector
+                for (int vectorDocId : docIdsWithVectors) {
+                    if (vectorDocId == target) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
+            @Override
+            public int docID() {
+                return docId;
+            }
+
+            @Override
+            public int nextDoc() {
+                return advance(docId + 1);
+            }
+
+            @Override
+            public int advance(int target) {
+                for (int vectorDocId : docIdsWithVectors) {
+                    if (vectorDocId >= target) {
+                        docId = vectorDocId;
+                        return docId;
+                    }
+                }
+                return NO_MORE_DOCS;
+            }
+
+            @Override
+            public long cost() {
+                return totalDocs;
+            }
+        };
+
+        // Test the fixed version (with ordToDoc)
+        DenormalizedCosineFloatVectorValues vectorValues = new DenormalizedCosineFloatVectorValues(sparseVectorValues, sparseMagnitudes);
+
+        // Test that ordToDoc method properly maps ordinals to document IDs
+        KnnVectorValues.DocIndexIterator iterator = vectorValues.iterator();
+
+        for (int ord = 0; ord < numVectors; ord++) {
+            iterator.advance(ord);
+
+            // Verify that ordToDoc works correctly
+            int expectedDocId = docIdsWithVectors[ord];
+            int actualDocId = vectorValues.ordToDoc(ord);
+            assertEquals("ordToDoc should correctly map ord " + ord + " to docId " + expectedDocId, expectedDocId, actualDocId);
+
+            // Get the denormalized vector - this relies on ordToDoc working correctly
+            float[] actualVector = vectorValues.vectorValue(iterator.index());
+            float actualMagnitude = vectorValues.magnitude();
+
+            // Verify the denormalized vector is correct
+            assertArrayEquals(
+                "Vector at ord " + ord + " (docId " + expectedDocId + ") should be correctly denormalized",
+                expectedVectors[ord],
+                actualVector,
+                1e-6f
+            );
+
+            // Verify the magnitude is correct
+            assertEquals(
+                "Magnitude at ord " + ord + " (docId " + expectedDocId + ") should be correct",
+                magnitudes[ord],
+                actualMagnitude,
+                1e-6f
+            );
+        }
+    }
+
 }