feat(search): HNSW New Arguments

packages/search/lib/commands/CREATE.spec.ts

@@ -245,6 +245,27 @@ describe('FT.CREATE', () => {
           );
         });
 
+        it('HNSW algorithm with SQ8 COMPRESSION and TRAINING_THRESHOLD', () => {
+          assert.deepEqual(
+            parseArgs(CREATE, 'index', {
+              field: {
+                type: SCHEMA_FIELD_TYPE.VECTOR,
+                ALGORITHM: SCHEMA_VECTOR_FIELD_ALGORITHM.HNSW,
+                TYPE: 'FLOAT32',
+                DIM: 64,
+                DISTANCE_METRIC: 'L2',
+                COMPRESSION: 'SQ8',
+                TRAINING_THRESHOLD: 4096
+              }
+            }),
+            [
+              'FT.CREATE', 'index', 'SCHEMA', 'field', 'VECTOR', 'HNSW', '10', 'TYPE',
+              'FLOAT32', 'DIM', '64', 'DISTANCE_METRIC', 'L2', 'COMPRESSION', 'SQ8',
+              'TRAINING_THRESHOLD', '4096'
+            ]
+          );
+        });
+
         it('VAMANA algorithm', () => {
           assert.deepEqual(
             parseArgs(CREATE, 'index', {
@@ -708,6 +729,60 @@ describe('FT.CREATE', () => {
     );
   }, GLOBAL.SERVERS.OPEN);
 
+  testUtils.testWithClientIfVersionWithinRange([[8, 12], 'LATEST'], 'client.ft.create vector hnsw sq8 compression', async client => {
+    assert.equal(
+      await client.ft.create('index_hnsw_sq8', {
+        field: {
+          type: SCHEMA_FIELD_TYPE.VECTOR,
+          ALGORITHM: SCHEMA_VECTOR_FIELD_ALGORITHM.HNSW,
+          TYPE: 'FLOAT32',
+          DIM: 64,
+          DISTANCE_METRIC: 'L2',
+          COMPRESSION: 'SQ8',
+          TRAINING_THRESHOLD: 4096
+        }
+      }),
+      'OK'
+    );
+
+    const [attribute] = (await client.ft.info('index_hnsw_sq8')).attributes;
+    assert.equal(attribute.compression, 'SQ8');
+    assert.equal(Number(attribute.training_threshold), 4096);
+
+    assert.equal(
+      await client.ft.create('index_hnsw_sq8_zero_threshold', {
+        field: {
+          type: SCHEMA_FIELD_TYPE.VECTOR,
+          ALGORITHM: SCHEMA_VECTOR_FIELD_ALGORITHM.HNSW,
+          TYPE: 'FLOAT16',
+          DIM: 64,
+          DISTANCE_METRIC: 'L2',
+          COMPRESSION: 'SQ8',
+          TRAINING_THRESHOLD: 0
+        }
+      }),
+      'OK'
+    );
+
+    const [zeroAttribute] = (await client.ft.info('index_hnsw_sq8_zero_threshold')).attributes;
+    assert.equal(Number(zeroAttribute.training_threshold), 0);
+
+    // TRAINING_THRESHOLD is rejected without COMPRESSION
+    await assert.rejects(
+      client.ft.create('index_hnsw_threshold_without_compression', {
+        field: {
+          type: SCHEMA_FIELD_TYPE.VECTOR,
+          ALGORITHM: SCHEMA_VECTOR_FIELD_ALGORITHM.HNSW,
+          TYPE: 'FLOAT32',
+          DIM: 64,
+          DISTANCE_METRIC: 'L2',
+          TRAINING_THRESHOLD: 4096
+        }
+      }),
+      /TRAINING_THRESHOLD is irrelevant/
+    );
+  }, GLOBAL.SERVERS.OPEN);
+
   testUtils.testWithClientIfVersionWithinRange([[8, 2], 'LATEST'], 'client.ft.create vector svs-vamana', async client => {
     assert.equal(
       await client.ft.create("index_svs_vamana_min_config", {

packages/search/lib/commands/CREATE.ts

@@ -86,6 +86,17 @@ interface SchemaHNSWVectorField extends SchemaVectorField {
    * available since 8.10
    */
   RERANK?: boolean;
+  /**
+   * Scalar quantization of the stored vectors. Requires TYPE FLOAT32 or FLOAT16.
+   * available since 8.12
+   */
+  COMPRESSION?: 'SQ8';
+  /**
+   * Number of vectors collected before computing the quantization parameters (0 - 102400).
+   * Applicable only with COMPRESSION; 0 disables mean normalization.
+   * available since 8.12
+   */
+  TRAINING_THRESHOLD?: number;
 }
 
 export const VAMANA_COMPRESSION_ALGORITHM = {
@@ -302,6 +313,14 @@ export function parseSchema(parser: CommandParser, schema: RediSearchSchema) {
                 args.push('RERANK', fieldOptions.RERANK ? 'TRUE' : 'FALSE');
               }
 
+              if (fieldOptions.COMPRESSION) {
+                args.push('COMPRESSION', fieldOptions.COMPRESSION);
+              }
+
+              if (fieldOptions.TRAINING_THRESHOLD !== undefined) {
+                args.push('TRAINING_THRESHOLD', fieldOptions.TRAINING_THRESHOLD.toString());
+              }
+
               break;
 
             case SCHEMA_VECTOR_FIELD_ALGORITHM['VAMANA']: