Merge 00a38aa565 into 40be7fa1a8

litellm/integrations/opentelemetry.py

@@ -2823,7 +2823,7 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
         primitive_value: Final = self._cast_as_primitive_value_type(value)
         span.set_attribute(key, primitive_value)
 
-    def _transform_messages_to_otel_semantic_conventions(self, messages: list[dict] | str) -> list[dict]:
+    def _transform_messages_to_otel_semantic_conventions(self, messages: list[Any] | str) -> list[dict]:
         """
         Transforms LiteLLM/OpenAI style messages into OTEL GenAI 1.38 compliant format.
         OTEL expects a 'parts' array instead of a single 'content' string.
@@ -2839,6 +2839,11 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
 
         transformed: Final = []
         for msg in messages:
+            # An embedding's input is logged as its messages, so a token-array input
+            # (`[[9906, 1917]]` or `[9906, 1917]`) yields lists or ints with no role or content.
+            # Anything else is message-like: a dict, or a `Message` object from chat output.
+            if isinstance(msg, (list, int)):
+                continue
             role = msg.get("role", "user")
             content = msg.get("content", "")
             parts = []

tests/unit/integrations/test_opentelemetry.py

@@ -3325,6 +3325,59 @@ class TestOpenTelemetrySemanticConventions138(unittest.TestCase):
         self.assertEqual(parsed[0]["parts"][0]["type"], "text")
         self.assertEqual(parsed[0]["parts"][0]["content"], "Hello world")
 
+    @parameterized.expand(
+        [
+            ("batch_of_token_arrays", [[9906, 1917], [15339]]),
+            ("single_token_array", [9906, 1917]),
+        ]
+    )
+    def test_token_array_embedding_input_keeps_span_attributes(self, _name, token_input):
+        """
+        An embedding whose input is token arrays is logged with those arrays as its
+        messages. set_attributes must still set every attribute after the input messages
+        rather than stopping at the first entry that is not a message dict.
+        """
+        otel = OpenTelemetry()
+        mock_span = MagicMock()
+
+        kwargs = {
+            "model": "text-embedding-3-small",
+            "messages": token_input,
+            "optional_params": {},
+            "litellm_params": {"custom_llm_provider": "openai"},
+            "standard_logging_object": {
+                "id": "test-id",
+                "call_type": "aembedding",
+                "metadata": {},
+            },
+        }
+        response_obj = {
+            "model": "text-embedding-3-small",
+            "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2]}],
+            "usage": {"prompt_tokens": 3, "total_tokens": 3},
+        }
+
+        with patch("litellm.integrations.opentelemetry.verbose_logger") as logger:
+            otel.set_attributes(span=mock_span, kwargs=kwargs, response_obj=response_obj)
+        logger.exception.assert_not_called()
+
+        attrs = {c[0][0]: c[0][1] for c in mock_span.set_attribute.call_args_list}
+        self.assertEqual(json.loads(attrs["gen_ai.input.messages"]), [])
+        self.assertEqual(attrs["gen_ai.operation.name"], "aembedding")
+
+    def test_choice_message_objects_are_still_transformed(self):
+        """
+        Chat output reaches the message conversion as `Message` objects, not dicts, so
+        skipping token-array entries must not skip them.
+        """
+        from litellm.types.utils import Choices, Message
+
+        choices = [Choices(finish_reason="stop", index=0, message=Message(role="assistant", content="Hi"))]
+        self.assertEqual(
+            OpenTelemetry()._transform_choices_to_otel_semantic_conventions(choices),
+            [{"role": "assistant", "parts": [{"type": "text", "content": "Hi"}], "finish_reason": "stop"}],
+        )
+
     def test_output_messages_uses_parts_structure(self):
         """
         Test that gen_ai.output.messages uses the OTEL 1.38 parts array structure.