Now supports cosine distance metric [API]

Added type C4VectorMetric and field C4VectorIndexOptions::metric.
Default value should be kC4VectorMetricEuclidean.

C/include/c4IndexTypes.h

@@ -27,9 +27,16 @@ typedef C4_ENUM(uint32_t, C4IndexType){
         kC4FullTextIndex,    ///< Full-text index
         kC4ArrayIndex,       ///< Index of array values, for use with UNNEST
         kC4PredictiveIndex,  ///< Index of prediction() results (Enterprise Edition only)
-        kC4VectorIndex,      ///< Index of ML vector similarity.
+        kC4VectorIndex,      ///< Index of ML vector similarity (Enterprise Edition only)
 };
 
+/** Distance metric to use in vector indexes. */
+typedef C4_ENUM(uint32_t, C4VectorMetric){
+        kC4VectorMetricDefault,    ///< Use default metric, Euclidean
+        kC4VectorMetricEuclidean,  ///< Euclidean distance (squared)
+        kC4VectorMetricCosine,     ///< Cosine distance (1.0 - cosine similarity)
+};                                 // Values must match IndexSpec::VectorOptions::Metric
+
 /** Types of encoding (compression) to use in vector indexes. */
 typedef C4_ENUM(uint32_t, C4VectorEncoding){
         kC4VectorEncodingDefault,  ///< Use default encoding, which is currently SQ
@@ -40,6 +47,7 @@ typedef C4_ENUM(uint32_t, C4VectorEncoding){
 /** Options for vector indexes. */
 typedef struct C4VectorIndexOptions {
     unsigned         numCentroids;  ///< Number of buckets to partition the vectors between
+    C4VectorMetric   metric;        ///< Distance metric
     C4VectorEncoding encoding;      ///< Vector compression type
 } C4VectorIndexOptions;
 

LiteCore/Database/CollectionImpl.hh

@@ -400,7 +400,7 @@ namespace litecore {
                         ftsOpt.ignoreDiacritics = indexOptions->ignoreDiacritics;
                         ftsOpt.disableStemming  = indexOptions->disableStemming;
                         ftsOpt.stopWords        = indexOptions->stopWords;
-                        options = ftsOpt;
+                        options                 = ftsOpt;
                         break;
                     }
                 case kC4VectorIndex:

LiteCore/Query/IndexSpec.hh

@@ -48,13 +48,20 @@ namespace litecore {
 
         /// Options for a vector index.
         struct VectorOptions {
+            enum Metric {
+                DefaultMetric,  ///< Use default metric, Euclidean
+                Euclidean,      ///< Euclidean distance (squared)
+                Cosine,         ///< Cosine distance (1.0 - cosine similarity)
+            };                  // Note: values must match C4VectorMetric in c4IndexTypes.h
+
             enum Encoding {
                 DefaultEncoding,  ///< Use default encoding, which is currently SQ
                 NoEncoding,       ///< No encoding; 4 bytes per dimension, no data loss
                 SQEncoding,       ///< Scalar Quantizer; 1 byte per dimension (recommended)
             };                    // Note: values must match C4VectorEncoding in c4IndexTypes.h
 
             unsigned numCentroids{2048};         ///< Number of centroids/buckets to divide the index into
+            Metric   metric{DefaultMetric};      ///< Distance metric
             Encoding encoding{DefaultEncoding};  ///< Vector encoding/compression
         };
 

LiteCore/Query/QueryParser.hh

@@ -120,7 +120,7 @@ namespace litecore {
         string predictiveIdentifier(const Value*) const;
         string predictiveTableName(const Value*) const;
 
-        /// Translates the JSON-parsed Value to blob-format vector for use by sqlite-vss.
+        /// Translates the JSON-parsed Value to blob-format vector for use by vectorsearch.
         string vectorExpressionSQL(const Value*);
         string vectorIndexTableName(const Value* matchExpr);
 

LiteCore/Query/SQLiteKeyStore+VectorIndex.cc

@@ -29,12 +29,13 @@ using namespace fleece::impl;
 
 namespace litecore {
 
-    // Vector search index for ML / predictive query, using the sqlite-vss extension.
-    // sqlite-vss documentation: https://github.com/asg017/sqlite-vss/blob/main/docs.md
+    // Vector search index for ML / predictive query, using the vectorsearch extension.
+    // https://github.com/couchbaselabs/mobile-vector-search/blob/main/README_Extension.md
 
     static constexpr const char* kVectorEncodingNames[] = {nullptr, "none", "PQ", "SQ"};
+    static constexpr const char* kVectorMetricNames[]   = {nullptr, "euclidean2", "cosine"};
 
-    // Creates a vector-similarity index using the sqlite-vss extension.
+    // Creates a vector-similarity index.
     bool SQLiteKeyStore::createVectorIndex(const IndexSpec& spec) {
         auto vectorTableName = db().auxiliaryTableName(tableName(), KeyStore::kVectorSeparator, spec.name);
 
@@ -53,6 +54,9 @@ namespace litecore {
             IndexSpec::VectorOptions options;
             if ( IndexSpec::VectorOptions const* o = spec.vectorOptions() ) { options = *o; }
             createStmt << "centroids=" << options.numCentroids << ",minToTrain=" << options.numCentroids * 25;
+            if ( options.metric != IndexSpec::VectorOptions::DefaultMetric ) {
+                createStmt << ",metric=" << kVectorMetricNames[options.metric];
+            }
             if ( options.encoding != IndexSpec::VectorOptions::DefaultEncoding ) {
                 createStmt << ",encoding=" << kVectorEncodingNames[options.encoding];
             }