Speed up (filtered) KNN queries for flat vector fields (elastic#130251)

For dense vector fields using the `flat` index, we already know a brute-force search will be used—so there’s no need to go through the codec’s approximate KNN logic. This change skips that step and builds the brute-force query directly, making things faster and simpler.

I tested this on a setup with **10 million random vectors**, each with **1596 dimensions** and **17,500 partitions**, using the `random_vector` track.
The results:

### Performance Comparison

| Metric            | Before    | After      | Change    |
| ----------------- | --------- | ---------- | --------- |
| **Throughput**    | 221 ops/s | 2762 ops/s | 🟢 +1149% |
| **Latency (p50)** | 29.2 ms   | 1.6 ms     | 🔻 -94.4% |
| **Latency (p99)** | 81.6 ms   | 3.5 ms     | 🔻 -95.7% |

Filtered KNN queries on flat vectors are now over 10x faster on my laptop!

Author: jimczi

docs/changelog/130251.yaml

@@ -0,0 +1,5 @@
+pr: 130251
+summary: Speed up (filtered) KNN queries for flat vector fields
+area: Vector Search
+type: enhancement
+issues: []

qa/vector/src/main/java/org/elasticsearch/test/knn/KnnSearcher.java

@@ -310,7 +310,7 @@ TopDocs doVectorQuery(float[] vector, IndexSearcher searcher) throws IOException
         }
         if (overSamplingFactor > 1f) {
             // oversample the topK results to get more candidates for the final result
-            knnQuery = new RescoreKnnVectorQuery(VECTOR_FIELD, vector, similarityFunction, this.topK, knnQuery);
+            knnQuery = RescoreKnnVectorQuery.fromInnerQuery(VECTOR_FIELD, vector, similarityFunction, this.topK, topK, knnQuery);
         }
         QueryProfiler profiler = new QueryProfiler();
         TopDocs docs = searcher.search(knnQuery, this.topK);

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

@@ -30,6 +30,8 @@
 import org.apache.lucene.index.SegmentWriteState;
 import org.apache.lucene.index.VectorEncoding;
 import org.apache.lucene.index.VectorSimilarityFunction;
+import org.apache.lucene.search.BooleanClause;
+import org.apache.lucene.search.BooleanQuery;
 import org.apache.lucene.search.FieldExistsQuery;
 import org.apache.lucene.search.Query;
 import org.apache.lucene.search.join.BitSetProducer;
@@ -77,6 +79,7 @@
 import org.elasticsearch.search.lookup.Source;
 import org.elasticsearch.search.vectors.DenseVectorQuery;
 import org.elasticsearch.search.vectors.DiversifyingChildrenIVFKnnFloatVectorQuery;
+import org.elasticsearch.search.vectors.DiversifyingParentBlockQuery;
 import org.elasticsearch.search.vectors.ESDiversifyingChildrenByteKnnVectorQuery;
 import org.elasticsearch.search.vectors.ESDiversifyingChildrenFloatKnnVectorQuery;
 import org.elasticsearch.search.vectors.ESKnnByteVectorQuery;
@@ -1391,6 +1394,18 @@ public final boolean equals(Object other) {
         public final int hashCode() {
             return Objects.hash(type, doHashCode());
         }
+
+        /**
+         * Indicates whether the underlying vector search is performed using a flat (exhaustive) approach.
+         * <p>
+         * When {@code true}, it means the search does not use any approximate nearest neighbor (ANN)
+         * acceleration structures such as HNSW or IVF. Instead, it performs a brute-force comparison
+         * against all candidate vectors. This information can be used by higher-level components
+         * to decide whether additional acceleration or optimization is necessary.
+         *
+         * @return {@code true} if the vector search is flat (exhaustive), {@code false} if it uses ANN structures
+         */
+        abstract boolean isFlat();
     }
 
     abstract static class QuantizedIndexOptions extends DenseVectorIndexOptions {
@@ -1762,6 +1777,11 @@ int doHashCode() {
             return Objects.hash(confidenceInterval, rescoreVector);
         }
 
+        @Override
+        boolean isFlat() {
+            return true;
+        }
+
         @Override
         public boolean updatableTo(DenseVectorIndexOptions update) {
             return update.type.equals(this.type)
@@ -1810,6 +1830,11 @@ public boolean doEquals(DenseVectorIndexOptions o) {
         public int doHashCode() {
             return Objects.hash(type);
         }
+
+        @Override
+        boolean isFlat() {
+            return true;
+        }
     }
 
     public static class Int4HnswIndexOptions extends QuantizedIndexOptions {
@@ -1860,6 +1885,11 @@ public int doHashCode() {
             return Objects.hash(m, efConstruction, confidenceInterval, rescoreVector);
         }
 
+        @Override
+        boolean isFlat() {
+            return false;
+        }
+
         @Override
         public String toString() {
             return "{type="
@@ -1931,6 +1961,11 @@ public int doHashCode() {
             return Objects.hash(confidenceInterval, rescoreVector);
         }
 
+        @Override
+        boolean isFlat() {
+            return true;
+        }
+
         @Override
         public String toString() {
             return "{type=" + type + ", confidence_interval=" + confidenceInterval + ", rescore_vector=" + rescoreVector + "}";
@@ -1999,6 +2034,11 @@ public int doHashCode() {
             return Objects.hash(m, efConstruction, confidenceInterval, rescoreVector);
         }
 
+        @Override
+        boolean isFlat() {
+            return false;
+        }
+
         @Override
         public String toString() {
             return "{type="
@@ -2088,6 +2128,11 @@ public int doHashCode() {
             return Objects.hash(m, efConstruction);
         }
 
+        @Override
+        boolean isFlat() {
+            return false;
+        }
+
         @Override
         public String toString() {
             return "{type=" + type + ", m=" + m + ", ef_construction=" + efConstruction + "}";
@@ -2126,6 +2171,11 @@ int doHashCode() {
             return Objects.hash(m, efConstruction, rescoreVector);
         }
 
+        @Override
+        boolean isFlat() {
+            return false;
+        }
+
         @Override
         public XContentBuilder toXContent(XContentBuilder builder, Params params) throws IOException {
             builder.startObject();
@@ -2179,6 +2229,11 @@ int doHashCode() {
             return CLASS_NAME_HASH;
         }
 
+        @Override
+        boolean isFlat() {
+            return true;
+        }
+
         @Override
         public XContentBuilder toXContent(XContentBuilder builder, Params params) throws IOException {
             builder.startObject();
@@ -2237,6 +2292,11 @@ int doHashCode() {
             return Objects.hash(clusterSize, defaultNProbe, rescoreVector);
         }
 
+        @Override
+        boolean isFlat() {
+            return false;
+        }
+
         @Override
         public XContentBuilder toXContent(XContentBuilder builder, Params params) throws IOException {
             builder.startObject();
@@ -2485,9 +2545,21 @@ private Query createKnnBitQuery(
             KnnSearchStrategy searchStrategy
         ) {
             elementType.checkDimensions(dims, queryVector.length);
-            Query knnQuery = parentFilter != null
-                ? new ESDiversifyingChildrenByteKnnVectorQuery(name(), queryVector, filter, k, numCands, parentFilter, searchStrategy)
-                : new ESKnnByteVectorQuery(name(), queryVector, k, numCands, filter, searchStrategy);
+            Query knnQuery;
+            if (indexOptions != null && indexOptions.isFlat()) {
+                var exactKnnQuery = parentFilter != null
+                    ? new DiversifyingParentBlockQuery(parentFilter, createExactKnnBitQuery(queryVector))
+                    : createExactKnnBitQuery(queryVector);
+                knnQuery = filter == null
+                    ? exactKnnQuery
+                    : new BooleanQuery.Builder().add(exactKnnQuery, BooleanClause.Occur.SHOULD)
+                        .add(filter, BooleanClause.Occur.FILTER)
+                        .build();
+            } else {
+                knnQuery = parentFilter != null
+                    ? new ESDiversifyingChildrenByteKnnVectorQuery(name(), queryVector, filter, k, numCands, parentFilter, searchStrategy)
+                    : new ESKnnByteVectorQuery(name(), queryVector, k, numCands, filter, searchStrategy);
+            }
             if (similarityThreshold != null) {
                 knnQuery = new VectorSimilarityQuery(
                     knnQuery,
@@ -2513,9 +2585,22 @@ private Query createKnnByteQuery(
                 float squaredMagnitude = VectorUtil.dotProduct(queryVector, queryVector);
                 elementType.checkVectorMagnitude(similarity, ElementType.errorByteElementsAppender(queryVector), squaredMagnitude);
             }
-            Query knnQuery = parentFilter != null
-                ? new ESDiversifyingChildrenByteKnnVectorQuery(name(), queryVector, filter, k, numCands, parentFilter, searchStrategy)
-                : new ESKnnByteVectorQuery(name(), queryVector, k, numCands, filter, searchStrategy);
+
+            Query knnQuery;
+            if (indexOptions != null && indexOptions.isFlat()) {
+                var exactKnnQuery = parentFilter != null
+                    ? new DiversifyingParentBlockQuery(parentFilter, createExactKnnByteQuery(queryVector))
+                    : createExactKnnByteQuery(queryVector);
+                knnQuery = filter == null
+                    ? exactKnnQuery
+                    : new BooleanQuery.Builder().add(exactKnnQuery, BooleanClause.Occur.SHOULD)
+                        .add(filter, BooleanClause.Occur.FILTER)
+                        .build();
+            } else {
+                knnQuery = parentFilter != null
+                    ? new ESDiversifyingChildrenByteKnnVectorQuery(name(), queryVector, filter, k, numCands, parentFilter, searchStrategy)
+                    : new ESKnnByteVectorQuery(name(), queryVector, k, numCands, filter, searchStrategy);
+            }
             if (similarityThreshold != null) {
                 knnQuery = new VectorSimilarityQuery(
                     knnQuery,
@@ -2568,7 +2653,16 @@ && isNotUnitVector(squaredMagnitude)) {
                 numCands = Math.max(adjustedK, numCands);
             }
             Query knnQuery;
-            if (indexOptions instanceof BBQIVFIndexOptions bbqIndexOptions) {
+            if (indexOptions != null && indexOptions.isFlat()) {
+                var exactKnnQuery = parentFilter != null
+                    ? new DiversifyingParentBlockQuery(parentFilter, createExactKnnFloatQuery(queryVector))
+                    : createExactKnnFloatQuery(queryVector);
+                knnQuery = filter == null
+                    ? exactKnnQuery
+                    : new BooleanQuery.Builder().add(exactKnnQuery, BooleanClause.Occur.SHOULD)
+                        .add(filter, BooleanClause.Occur.FILTER)
+                        .build();
+            } else if (indexOptions instanceof BBQIVFIndexOptions bbqIndexOptions) {
                 knnQuery = parentFilter != null
                     ? new DiversifyingChildrenIVFKnnFloatVectorQuery(
                         name(),
@@ -2594,11 +2688,12 @@ && isNotUnitVector(squaredMagnitude)) {
                     : new ESKnnFloatVectorQuery(name(), queryVector, adjustedK, numCands, filter, knnSearchStrategy);
             }
             if (rescore) {
-                knnQuery = new RescoreKnnVectorQuery(
+                knnQuery = RescoreKnnVectorQuery.fromInnerQuery(
                     name(),
                     queryVector,
                     similarity.vectorSimilarityFunction(indexVersionCreated, ElementType.FLOAT),
                     k,
+                    adjustedK,
                     knnQuery
                 );
             }
@@ -2624,7 +2719,7 @@ ElementType getElementType() {
             return elementType;
         }
 
-        public IndexOptions getIndexOptions() {
+        public DenseVectorIndexOptions getIndexOptions() {
             return indexOptions;
         }
 

server/src/main/java/org/elasticsearch/search/vectors/DenseVectorQuery.java

@@ -207,4 +207,5 @@ public int docID() {
             return iterator.docID();
         }
     }
+
 }