Optimize decode_vector with NumPy arrays for faster vector decoding

Author: Alex Waszkiewicz

python/cocoindex/engine_value.py

@@ -365,57 +365,38 @@ def decode(value: Any) -> Any | None:
         return lambda value: value
 
     if isinstance(src_type, BasicValueType) and src_type.kind == "Vector":
-        field_path_str = "".join(field_path)
-        if not isinstance(dst_type_variant, AnalyzedListType):
-            raise ValueError(
-                f"Type mismatch for `{''.join(field_path)}`: "
-                f"declared `{dst_type_info.core_type}`, a list type expected"
-            )
-        expected_dim = (
-            dst_type_variant.vector_info.dim
-            if dst_type_variant and dst_type_variant.vector_info
-            else None
-        )
 
-        vec_elem_decoder = None
-        scalar_dtype = None
-        if dst_type_variant and dst_type_info.base_type is np.ndarray:
-            if is_numpy_number_type(dst_type_variant.elem_type):
-                scalar_dtype = dst_type_variant.elem_type
-        else:
-            # mypy: vector info exists for Vector kind
-            assert src_type.vector is not None  # type: ignore[unreachable]
-            vec_elem_decoder = make_engine_value_decoder(
-                field_path + ["[*]"],
-                src_type.vector.element_type,
-                analyze_type_info(
-                    dst_type_variant.elem_type if dst_type_variant else Any
-                ),
-            )
+    # force numeric vectors to be converted directly to NumPy arrays
+    scalar_dtype = (
+        dst_type_variant.elem_type
+        if dst_type_variant and is_numpy_number_type(dst_type_variant.elem_type)
+        else None
+    )
+    expected_dim = (
+        dst_type_variant.vector_info.dim
+        if dst_type_variant and dst_type_variant.vector_info
+        else None
+    )
 
-        def decode_vector(value: Any) -> Any | None:
-            if value is None:
-                if dst_type_info.nullable:
-                    return None
-                raise ValueError(
-                    f"Received null for non-nullable vector `{field_path_str}`"
-                )
-            if not isinstance(value, (np.ndarray, list)):
-                raise TypeError(
-                    f"Expected NDArray or list for vector `{field_path_str}`, got {type(value)}"
-                )
-            if expected_dim is not None and len(value) != expected_dim:
-                raise ValueError(
-                    f"Vector dimension mismatch for `{field_path_str}`: "
-                    f"expected {expected_dim}, got {len(value)}"
-                )
+    def decode_vector(value: Any) -> Any | None:
+        if value is None:
+            return None if dst_type_info.nullable else np.zeros(expected_dim, dtype=scalar_dtype)
+
+        if not isinstance(value, (list, np.ndarray)):
+            raise TypeError(f"Expected NDArray or list for vector `{''.join(field_path)}`, got {type(value)}")
+
+
+        arr = np.array(value, dtype=scalar_dtype)
+
+        if expected_dim is not None and arr.shape[0] != expected_dim:
+            raise ValueError(
+                f"Vector dimension mismatch for `{''.join(field_path)}`: "
+                f"expected {expected_dim}, got {arr.shape[0]}"
+            )
+        return arr
 
-            if vec_elem_decoder is not None:  # for Non-NDArray vector
-                return [vec_elem_decoder(v) for v in value]
-            else:  # for NDArray vector
-                return np.array(value, dtype=scalar_dtype)
+    return decode_vector
 
-        return decode_vector
 
     if isinstance(dst_type_variant, AnalyzedBasicType):
         if not _is_type_kind_convertible_to(src_type_kind, dst_type_variant.kind):