fix: decode base64 embeddings as little-endian floats

src/openai/lib/_parsing/_embeddings.py

@@ -2,6 +2,7 @@
 
 import array
 import base64
+import sys
 from typing import cast
 
 from ..._types import Omit, NotGiven
@@ -28,12 +29,16 @@ def parse_embedding_response(
         data = cast(object, embedding.embedding)
         if not isinstance(data, str):
             continue
+        decoded = base64.b64decode(data)
         if not use_numpy:
             # use array for base64 optimisation
-            embedding.embedding = array.array("f", base64.b64decode(data)).tolist()
+            floats = array.array("f", decoded)
+            if sys.byteorder == "big":
+                floats.byteswap()
+            embedding.embedding = floats.tolist()
         else:
             embedding.embedding = np.frombuffer(  # type: ignore[no-untyped-call]
-                base64.b64decode(data), dtype="float32"
+                decoded, dtype="<f4"
             ).tolist()
 
     return obj

tests/lib/test_embeddings.py

@@ -1,9 +1,10 @@
 from __future__ import annotations
 
-import json
-import array
 import base64
 import binascii
+import json
+import struct
+from types import SimpleNamespace
 from typing import Any, cast
 from typing_extensions import Literal
 
@@ -18,7 +19,7 @@
 from openai.types.create_embedding_response import CreateEmbeddingResponse
 
 VALUES = [0.125, -2.5, 3.75]
-ENCODED = base64.b64encode(array.array("f", VALUES).tobytes()).decode("ascii")
+ENCODED = base64.b64encode(struct.pack("<3f", *VALUES)).decode("ascii")
 EncodingFormat = Literal["float", "base64"] | Omit
 ResponseMode = Literal["normal", "raw", "streaming"]
 
@@ -89,6 +90,36 @@ def unexpected_decoder() -> bool:
     assert embeddings_parser.parse_embedding_response(response, encoding_format=omit) is response
 
 
+def test_stdlib_decoder_swaps_big_endian_values(monkeypatch: pytest.MonkeyPatch) -> None:
+    response = make_response(ENCODED)
+    decoded = base64.b64decode(ENCODED)
+
+    class Floats:
+        swapped = False
+
+        def byteswap(self) -> None:
+            self.swapped = True
+
+        def tolist(self) -> list[float]:
+            assert self.swapped
+            return VALUES
+
+    floats = Floats()
+
+    def make_array(typecode: str, initializer: bytes) -> Floats:
+        assert typecode == "f"
+        assert initializer == decoded
+        return floats
+
+    monkeypatch.setattr(embeddings_parser, "has_numpy", lambda: False)
+    monkeypatch.setattr(embeddings_parser, "sys", SimpleNamespace(byteorder="big"))
+    monkeypatch.setattr(embeddings_parser.array, "array", make_array)
+
+    parsed = embeddings_parser.parse_embedding_response(response, encoding_format=omit)
+
+    assert parsed.data[0].embedding == VALUES
+
+
 @pytest.mark.parametrize("encoding_format", ["float", "base64", None])
 @pytest.mark.parametrize("vectors", [(ENCODED,), ("abc",), ()], ids=["encoded", "invalid", "empty"])
 def test_explicit_format_is_untouched(