fix(local): preserve near-zero cosine vectors

qdrant_client/local/distances.py

@@ -127,16 +127,19 @@ def cosine_similarity(query: types.NumpyArray, vectors: types.NumpyArray) -> typ
     Returns:
         distances
     """
+    vectors = np.array(vectors, copy=True)
+    query = np.array(query, copy=True)
+
     vectors_norm = np.linalg.norm(vectors, axis=-1)[:, np.newaxis]
-    vectors /= np.where(vectors_norm != 0.0, vectors_norm, EPSILON)
+    vectors = vectors / np.where(vectors_norm > EPSILON, vectors_norm, 1.0)
 
     if len(query.shape) == 1:
         query_norm = np.linalg.norm(query)
-        query /= np.where(query_norm != 0.0, query_norm, EPSILON)
+        query = query / np.where(query_norm > EPSILON, query_norm, 1.0)
         return np.dot(vectors, query)
 
     query_norm = np.linalg.norm(query, axis=-1)[:, np.newaxis]
-    query /= np.where(query_norm != 0.0, query_norm, EPSILON)
+    query = query / np.where(query_norm > EPSILON, query_norm, 1.0)
     return np.dot(query, vectors.T)
 
 

qdrant_client/local/local_collection.py

@@ -2442,7 +2442,7 @@ def _update_point(self, point: models.PointStruct) -> None:
 
                 if params.distance == models.Distance.COSINE:
                     vector_norm = np.linalg.norm(vector, axis=-1)[:, np.newaxis]
-                    vector /= np.where(vector_norm != 0.0, vector_norm, EPSILON)
+                    vector /= np.where(vector_norm > EPSILON, vector_norm, 1.0)
                 self.multivectors[vector_name][idx] = np.array(vector)
                 self.deleted_per_vector[vector_name][idx] = 0
             else:
@@ -2541,7 +2541,7 @@ def _add_point(self, point: models.PointStruct) -> None:
                 params = self.get_vector_params(vector_name)
                 if params.distance == models.Distance.COSINE:
                     vector_norm = np.linalg.norm(vector_np, axis=-1)[:, np.newaxis]
-                    vector_np /= np.where(vector_norm != 0.0, vector_norm, EPSILON)
+                    vector_np /= np.where(vector_norm > EPSILON, vector_norm, 1.0)
                 named_vectors[idx] = vector_np
                 self.deleted_per_vector[vector_name] = np.append(
                     self.deleted_per_vector[vector_name], 0
@@ -2686,7 +2686,7 @@ def _update_named_vectors(
             else:
                 if params.distance == models.Distance.COSINE:
                     vector_norm = np.linalg.norm(vector_np, axis=-1)[:, np.newaxis]
-                    vector_np /= np.where(vector_norm != 0.0, vector_norm, EPSILON)
+                    vector_np /= np.where(vector_norm > EPSILON, vector_norm, 1.0)
                 self.multivectors[vector_name][idx] = vector_np
 
     def update_vectors(

qdrant_client/local/tests/test_distances.py

@@ -55,3 +55,27 @@ def test_distances() -> None:
     multivector_query = np.array([[1, 2, 3], [3, 4, 5]])
     docs = [np.array([[1, 2, 3], [0, 1, 2]])]
     assert calculate_multi_distance(multivector_query, docs, models.Distance.DOT)[0] == 40.0
+
+
+def test_cosine_similarity_keeps_near_zero_vectors_unchanged() -> None:
+    tiny = np.array([5e-11] * 4, dtype=np.float32)
+    vectors = tiny.copy()
+    query = np.array([1.0] * 4, dtype=np.float32)
+
+    result = calculate_distance(query, vectors[None, :], models.Distance.COSINE)
+
+    assert np.allclose(vectors, tiny)
+    assert np.allclose(result, [1e-10], atol=1e-12)
+
+
+def test_cosine_similarity_does_not_mutate_inputs() -> None:
+    query = np.array([1.0, 1.0], dtype=np.float32)
+    vectors = np.array([[1.0, 1.0], [0.0, 0.0]], dtype=np.float32)
+    query_before = query.copy()
+    vectors_before = vectors.copy()
+
+    result = calculate_distance(query, vectors, models.Distance.COSINE)
+
+    assert np.allclose(query, query_before)
+    assert np.allclose(vectors, vectors_before)
+    assert np.allclose(result, [1.0, 0.0])