Assume normalized vecs

CHANGELOG.md

@@ -5,6 +5,7 @@
 - Use anonymous wasm modules for better performance [#264](https://github.com/hypermodeAI/runtime/pull/264)
 - Support optional parameters [#269](https://github.com/hypermodeAI/runtime/pull/269)
 - Handle null parameters [#270](https://github.com/hypermodeAI/runtime/pull/2270)
+- Assume normalized vectors when calculating cosine similarity [#265](https://github.com/hypermodeAI/runtime/pull/265)
 
 ## 2024-06-26 - Version 0.9.4
 

collections/in_mem/sequential/vector_index_test.go

@@ -47,6 +47,11 @@ func TestMultipleSequentialVectorIndexes(t *testing.T) {
 				for k := range vecs[j] {
 					vecs[j][k] += float32(i) / 10
 				}
+				var err error
+				vecs[j], err = utils.Normalize(vecs[j])
+				if err != nil {
+					t.Errorf("Failed to normalize vector: %v", err)
+				}
 			}
 
 			err := index.InsertVectorsToMemory(ctx, textIds, textIds, keys, vecs)

collections/utils/helper.go

@@ -36,6 +36,17 @@ func IsBetterScoreForSimilarity(a, b float64) bool {
 	return a > b
 }
 
+func Normalize(v []float32) ([]float32, error) {
+	norm := norm(v)
+	if norm == 0 {
+		return nil, errors.New("can not normalize vector with zero norm")
+	}
+	for i := range v {
+		v[i] /= norm
+	}
+	return v, nil
+}
+
 func norm(v []float32) float32 {
 	vectorNorm, _ := DotProduct(v, v)
 	return math32.Sqrt(vectorNorm)
@@ -52,20 +63,14 @@ func DotProduct(a, b []float32) (float32, error) {
 	return dotProduct, nil
 }
 
+// assume normalization for vectors
 func CosineSimilarity(a, b []float32) (float64, error) {
 	dotProd, err := DotProduct(a, b)
 	if err != nil {
 		return 0, err
 	}
-	normA := norm(a)
-	normB := norm(b)
-	if normA == 0 || normB == 0 {
-		err := errors.New("can not compute cosine similarity on zero vector")
-		var empty float64
-		return empty, err
-	}
 
-	return float64(dotProd) / (float64(normA) * float64(normB)), nil
+	return float64(dotProd), nil
 }
 
 func ConcatStrings(strs ...string) string {