[8.x] kNN vector rescoring for quantized vectors (#116663) (#118418)

* kNN vector rescoring for quantized vectors (#116663) (cherry picked from commit 5996772) # Conflicts: # server/src/main/java/org/elasticsearch/search/vectors/KnnSearchBuilder.java # x-pack/plugin/rank-rrf/src/main/java/org/elasticsearch/xpack/rank/rrf/RRFRankBuilder.java * FloatVectorValues have a different interface in this Lucene version
elastic · Dec 11, 2024 · d40813d · d40813d
1 parent 20f7919
commit d40813d
Show file tree

Hide file tree

Showing 66 changed files with 2,363 additions and 296 deletions.
diff --git a/docs/changelog/116663.yaml b/docs/changelog/116663.yaml
@@ -0,0 +1,5 @@
+pr: 116663
+summary: KNN vector rescoring for quantized vectors
+area: Vector Search
+type: feature
+issues: []
diff --git a/...or/src/internalClusterTest/java/org/elasticsearch/percolator/PercolatorQuerySearchIT.java b/...or/src/internalClusterTest/java/org/elasticsearch/percolator/PercolatorQuerySearchIT.java
@@ -1359,7 +1359,7 @@ public void testKnnQueryNotSupportedInPercolator() throws IOException {
             """);
         indicesAdmin().prepareCreate("index1").setMapping(mappings).get();
         ensureGreen();
-        QueryBuilder knnVectorQueryBuilder = new KnnVectorQueryBuilder("my_vector", new float[] { 1, 1, 1, 1, 1 }, 10, 10, null);
+        QueryBuilder knnVectorQueryBuilder = new KnnVectorQueryBuilder("my_vector", new float[] { 1, 1, 1, 1, 1 }, 10, 10, null, null);
 
         IndexRequestBuilder indexRequestBuilder = prepareIndex("index1").setId("knn_query1")
             .setSource(jsonBuilder().startObject().field("my_query", knnVectorQueryBuilder).endObject());

diff --git a/...spec/src/yamlRestTest/resources/rest-api-spec/test/search.retrievers/20_knn_retriever.yml b/...spec/src/yamlRestTest/resources/rest-api-spec/test/search.retrievers/20_knn_retriever.yml
@@ -18,7 +18,7 @@ setup:
                 dims: 5
                 index: true
                 index_options:
-                  type: hnsw
+                  type: int8_hnsw
                 similarity: l2_norm
 
   - do:
@@ -73,3 +73,59 @@ setup:
   - match: {hits.total.value: 1}
   - match: {hits.hits.0._id: "3"}
   - match: {hits.hits.0.fields.name.0: "rabbit.jpg"}
+
+---
+"Vector rescoring has no effect for non-quantized vectors and provides same results as non-rescored knn":
+  - requires:
+      reason: 'Quantized vector rescoring is required'
+      test_runner_features: [capabilities]
+      capabilities:
+        - method: GET
+          path: /_search
+          capabilities: [knn_quantized_vector_rescore]
+  - skip:
+      features: "headers"
+
+  # Rescore
+  - do:
+      headers:
+        Content-Type: application/json
+      search:
+        rest_total_hits_as_int: true
+        index: index1
+        body:
+          knn:
+            field: vector
+            query_vector: [2, 2, 2, 2, 3]
+            k: 3
+            num_candidates: 3
+            rescore_vector:
+              num_candidates_factor: 1.5
+
+  # Get rescoring scores - hit ordering may change depending on how things are distributed
+  - match: { hits.total: 3 }
+  - set: { hits.hits.0._score: rescore_score0 }
+  - set: { hits.hits.1._score: rescore_score1 }
+  - set: { hits.hits.2._score: rescore_score2 }
+
+  # Exact knn via script score
+  - do:
+      headers:
+        Content-Type: application/json
+      search:
+        rest_total_hits_as_int: true
+        index: index1
+        body:
+          query:
+            script_score:
+              query: {match_all: {} }
+              script:
+                source: "1.0 / (1.0 + Math.pow(l2norm(params.query_vector, 'vector'), 2.0))"
+                params:
+                  query_vector: [2, 2, 2, 2, 3]
+
+  # Compare scores as hit IDs may change depending on how things are distributed
+  - match: { hits.total: 3 }
+  - match: { hits.hits.0._score: $rescore_score0 }
+  - match: { hits.hits.1._score: $rescore_score1 }
+  - match: { hits.hits.2._score: $rescore_score2 }
diff --git a/...c/src/yamlRestTest/resources/rest-api-spec/test/search.vectors/210_knn_search_profile.yml b/...c/src/yamlRestTest/resources/rest-api-spec/test/search.vectors/210_knn_search_profile.yml
@@ -0,0 +1,137 @@
+setup:
+  - requires:
+      reason: 'Quantized vector rescoring is required'
+      test_runner_features: [ capabilities ]
+      capabilities:
+        - method: GET
+          path: /_search
+          capabilities: [ knn_quantized_vector_rescore ]
+  - skip:
+      features: "headers"
+
+  - do:
+      indices.create:
+        index: bbq_hnsw
+        body:
+          settings:
+            index:
+              number_of_shards: 1
+          mappings:
+            properties:
+              vector:
+                type: dense_vector
+                dims: 64
+                index: true
+                similarity: max_inner_product
+                index_options:
+                  type: bbq_hnsw
+
+  - do:
+      index:
+        index: bbq_hnsw
+        id: "1"
+        body:
+          vector: [0.077,  0.32 , -0.205,  0.63 ,  0.032,  0.201,  0.167, -0.313,
+                   0.176,  0.531, -0.375,  0.334, -0.046,  0.078, -0.349,  0.272,
+                   0.307, -0.083,  0.504,  0.255, -0.404,  0.289, -0.226, -0.132,
+                   -0.216,  0.49 ,  0.039,  0.507, -0.307,  0.107,  0.09 , -0.265,
+                   -0.285,  0.336, -0.272,  0.369, -0.282,  0.086, -0.132,  0.475,
+                   -0.224,  0.203,  0.439,  0.064,  0.246, -0.396,  0.297,  0.242,
+                   -0.028,  0.321, -0.022, -0.009, -0.001  ,  0.031, -0.533,  0.45,
+                   -0.683,  1.331,  0.194, -0.157, -0.1  , -0.279, -0.098, -0.176]
+  # Flush in order to provoke a merge later
+  - do:
+      indices.flush:
+        index: bbq_hnsw
+
+  - do:
+      index:
+        index: bbq_hnsw
+        id: "2"
+        body:
+          vector: [0.196,  0.514,  0.039,  0.555, -0.042,  0.242,  0.463, -0.348,
+                   -0.08 ,  0.442, -0.067, -0.05 , -0.001,  0.298, -0.377,  0.048,
+                   0.307,  0.159,  0.278,  0.119, -0.057,  0.333, -0.289, -0.438,
+                   -0.014,  0.361, -0.169,  0.292, -0.229,  0.123,  0.031, -0.138,
+                   -0.139,  0.315, -0.216,  0.322, -0.445, -0.059,  0.071,  0.429,
+                   -0.602, -0.142,  0.11 ,  0.192,  0.259, -0.241,  0.181, -0.166,
+                   0.082,  0.107, -0.05 ,  0.155,  0.011,  0.161, -0.486,  0.569,
+                   -0.489,  0.901,  0.208,  0.011, -0.209, -0.153, -0.27 , -0.013]
+  # Flush in order to provoke a merge later
+  - do:
+      indices.flush:
+        index: bbq_hnsw
+
+  - do:
+      index:
+        index: bbq_hnsw
+        id: "3"
+        body:
+          name: rabbit.jpg
+          vector: [0.139,  0.178, -0.117,  0.399,  0.014, -0.139,  0.347, -0.33 ,
+                   0.139,  0.34 , -0.052, -0.052, -0.249,  0.327, -0.288,  0.049,
+                   0.464,  0.338,  0.516,  0.247, -0.104,  0.259, -0.209, -0.246,
+                   -0.11 ,  0.323,  0.091,  0.442, -0.254,  0.195, -0.109, -0.058,
+                   -0.279,  0.402, -0.107,  0.308, -0.273,  0.019,  0.082,  0.399,
+                   -0.658, -0.03 ,  0.276,  0.041,  0.187, -0.331,  0.165,  0.017,
+                   0.171, -0.203, -0.198,  0.115, -0.007,  0.337, -0.444,  0.615,
+                   -0.657,  1.285,  0.2  , -0.062,  0.038,  0.089, -0.068, -0.058]
+  # Flush in order to provoke a merge later
+  - do:
+      indices.flush:
+        index: bbq_hnsw
+
+  - do:
+      indices.forcemerge:
+        index: bbq_hnsw
+        max_num_segments: 1
+---
+"Profile rescored knn search":
+
+  - do:
+      search:
+        index: bbq_hnsw
+        body:
+          profile: true
+          knn:
+            field: vector
+            query_vector: [0.128,  0.067, -0.08 ,  0.395, -0.11 , -0.259,  0.473, -0.393,
+                           0.292,  0.571, -0.491,  0.444, -0.288,  0.198, -0.343,  0.015,
+                           0.232,  0.088,  0.228,  0.151, -0.136,  0.236, -0.273, -0.259,
+                           -0.217,  0.359, -0.207,  0.352, -0.142,  0.192, -0.061, -0.17 ,
+                           -0.343,  0.189, -0.221,  0.32 , -0.301, -0.1  ,  0.005,  0.232,
+                           -0.344,  0.136,  0.252,  0.157, -0.13 , -0.244,  0.193, -0.034,
+                           -0.12 , -0.193, -0.102,  0.252, -0.185, -0.167, -0.575,  0.582,
+                           -0.426,  0.983,  0.212,  0.204,  0.03 , -0.276, -0.425, -0.158]
+            k: 3
+            num_candidates: 3
+            "rescore_vector":
+              "num_candidates_factor": 2.0
+
+  # We expect the knn search ops + rescoring num_cnaidates (for rescoring) per shard
+  - match: { profile.shards.0.dfs.knn.0.vector_operations_count: 6 }
+
+  # Search with similarity to check number of operations are propagated correctly
+  - do:
+      search:
+        index: bbq_hnsw
+        body:
+          profile: true
+          knn:
+            field: vector
+            query_vector: [0.128,  0.067, -0.08 ,  0.395, -0.11 , -0.259,  0.473, -0.393,
+                           0.292,  0.571, -0.491,  0.444, -0.288,  0.198, -0.343,  0.015,
+                           0.232,  0.088,  0.228,  0.151, -0.136,  0.236, -0.273, -0.259,
+                           -0.217,  0.359, -0.207,  0.352, -0.142,  0.192, -0.061, -0.17 ,
+                           -0.343,  0.189, -0.221,  0.32 , -0.301, -0.1  ,  0.005,  0.232,
+                           -0.344,  0.136,  0.252,  0.157, -0.13 , -0.244,  0.193, -0.034,
+                           -0.12 , -0.193, -0.102,  0.252, -0.185, -0.167, -0.575,  0.582,
+                           -0.426,  0.983,  0.212,  0.204,  0.03 , -0.276, -0.425, -0.158]
+            k: 3
+            num_candidates: 3
+            similarity: 100000
+            "rescore_vector":
+              "num_candidates_factor": 2.0
+
+  # We expect the knn search ops + rescoring num_cnaidates (for rescoring) per shard
+  - match: { profile.shards.0.dfs.knn.0.vector_operations_count: 6 }
diff --git a/rest-api-spec/src/yamlRestTest/resources/rest-api-spec/test/search.vectors/40_knn_search.yml b/rest-api-spec/src/yamlRestTest/resources/rest-api-spec/test/search.vectors/40_knn_search.yml
@@ -541,3 +541,58 @@ setup:
             num_candidates: 3
 
   - match: { hits.total.value: 0 }
+---
+"Vector rescoring has no effect for non-quantized vectors and provides same results as non-rescored knn":
+  - requires:
+      reason: 'Quantized vector rescoring is required'
+      test_runner_features: [capabilities]
+      capabilities:
+        - method: GET
+          path: /_search
+          capabilities: [knn_quantized_vector_rescore]
+  - skip:
+      features: "headers"
+
+  # Non-rescored knn
+  - do:
+      headers:
+        Content-Type: application/json
+      search:
+        rest_total_hits_as_int: true
+        index: test
+        body:
+          fields: [ "name" ]
+          knn:
+            field: vector
+            query_vector: [-0.5, 90.0, -10, 14.8, -156.0]
+            k: 3
+            num_candidates: 3
+
+  # Get scores - hit ordering may change depending on how things are distributed
+  - match: { hits.total: 3 }
+  - set: { hits.hits.0._score: knn_score0 }
+  - set: { hits.hits.1._score: knn_score1 }
+  - set: { hits.hits.2._score: knn_score2 }
+
+  # Rescored knn
+  - do:
+      headers:
+        Content-Type: application/json
+      search:
+        rest_total_hits_as_int: true
+        index: test
+        body:
+          fields: [ "name" ]
+          knn:
+            field: vector
+            query_vector: [-0.5, 90.0, -10, 14.8, -156.0]
+            k: 3
+            num_candidates: 3
+            rescore_vector:
+              num_candidates_factor: 1.5
+
+  # Compare scores as hit IDs may change depending on how things are distributed
+  - match: { hits.total: 3 }
+  - match: { hits.hits.0._score: $knn_score0 }
+  - match: { hits.hits.1._score: $knn_score1 }
+  - match: { hits.hits.2._score: $knn_score2 }
diff --git a/...c/src/yamlRestTest/resources/rest-api-spec/test/search.vectors/41_knn_search_bbq_hnsw.yml b/...c/src/yamlRestTest/resources/rest-api-spec/test/search.vectors/41_knn_search_bbq_hnsw.yml
@@ -108,6 +108,75 @@ setup:
   - match: { hits.hits.1._id: "3" }
   - match: { hits.hits.2._id: "2" }
 ---
+"Vector rescoring has same scoring as exact search for kNN section":
+  - requires:
+      reason: 'Quantized vector rescoring is required'
+      test_runner_features: [capabilities]
+      capabilities:
+        - method: GET
+          path: /_search
+          capabilities: [knn_quantized_vector_rescore]
+  - skip:
+      features: "headers"
+
+  # Rescore
+  - do:
+      headers:
+        Content-Type: application/json
+      search:
+        rest_total_hits_as_int: true
+        index: bbq_hnsw
+        body:
+          knn:
+            field: vector
+            query_vector: [0.128,  0.067, -0.08 ,  0.395, -0.11 , -0.259,  0.473, -0.393,
+                           0.292,  0.571, -0.491,  0.444, -0.288,  0.198, -0.343,  0.015,
+                           0.232,  0.088,  0.228,  0.151, -0.136,  0.236, -0.273, -0.259,
+                           -0.217,  0.359, -0.207,  0.352, -0.142,  0.192, -0.061, -0.17 ,
+                           -0.343,  0.189, -0.221,  0.32 , -0.301, -0.1  ,  0.005,  0.232,
+                           -0.344,  0.136,  0.252,  0.157, -0.13 , -0.244,  0.193, -0.034,
+                           -0.12 , -0.193, -0.102,  0.252, -0.185, -0.167, -0.575,  0.582,
+                           -0.426,  0.983,  0.212,  0.204,  0.03 , -0.276, -0.425, -0.158]
+            k: 3
+            num_candidates: 3
+            rescore_vector:
+              num_candidates_factor: 1.5
+
+  # Get rescoring scores - hit ordering may change depending on how things are distributed
+  - match: { hits.total: 3 }
+  - set: { hits.hits.0._score: rescore_score0 }
+  - set: { hits.hits.1._score: rescore_score1 }
+  - set: { hits.hits.2._score: rescore_score2 }
+
+  # Exact knn via script score
+  - do:
+      headers:
+        Content-Type: application/json
+      search:
+        rest_total_hits_as_int: true
+        body:
+          query:
+            script_score:
+              query: {match_all: {} }
+              script:
+                source: "double similarity = dotProduct(params.query_vector, 'vector'); return similarity < 0 ? 1 / (1 + -1 * similarity) : similarity + 1"
+                params:
+                  query_vector: [0.128,  0.067, -0.08 ,  0.395, -0.11 , -0.259,  0.473, -0.393,
+                           0.292,  0.571, -0.491,  0.444, -0.288,  0.198, -0.343,  0.015,
+                           0.232,  0.088,  0.228,  0.151, -0.136,  0.236, -0.273, -0.259,
+                           -0.217,  0.359, -0.207,  0.352, -0.142,  0.192, -0.061, -0.17 ,
+                           -0.343,  0.189, -0.221,  0.32 , -0.301, -0.1  ,  0.005,  0.232,
+                           -0.344,  0.136,  0.252,  0.157, -0.13 , -0.244,  0.193, -0.034,
+                           -0.12 , -0.193, -0.102,  0.252, -0.185, -0.167, -0.575,  0.582,
+                           -0.426,  0.983,  0.212,  0.204,  0.03 , -0.276, -0.425, -0.158]
+
+  # Compare scores as hit IDs may change depending on how things are distributed
+  - match: { hits.total: 3 }
+  - match: { hits.hits.0._score: $rescore_score0 }
+  - match: { hits.hits.1._score: $rescore_score1 }
+  - match: { hits.hits.2._score: $rescore_score2 }
+
+---
 "Test bad quantization parameters":
   - do:
       catch: bad_request