fix(llmobs): filter openai Omit/NotGiven sentinels from span metadata

ddtrace/llmobs/_integrations/utils.py

@@ -39,6 +39,17 @@
 
 logger = get_logger(__name__)
 
+# The openai SDK uses `Omit`/`NotGiven` sentinels as defaults for unset request parameters.
+# Callers (e.g. PydanticAI) may forward these sentinels explicitly, so filter them out of span
+# metadata rather than serializing them to noisy repr strings. Identify them by class name instead
+# of importing openai: this shared utils module is provider-agnostic, so it must not depend on a
+# specific vendor SDK (which also avoids a circular import while ddtrace is patching openai).
+_OPENAI_SENTINEL_TYPE_NAMES = ("Omit", "NotGiven")
+
+
+def _is_openai_sentinel(value: Any) -> bool:
+    return type(value).__name__ in _OPENAI_SENTINEL_TYPE_NAMES
+
 
 COMMON_METADATA_KEYS = (
     "stream",
@@ -572,7 +583,7 @@ def get_metadata_from_kwargs(
         keys_to_include += OPENAI_METADATA_CHAT_KEYS if operation == "chat" else OPENAI_METADATA_COMPLETION_KEYS
     elif integration_name == "litellm":
         keys_to_include += LITELLM_METADATA_CHAT_KEYS if operation == "chat" else LITELLM_METADATA_COMPLETION_KEYS
-    metadata = {k: load_data_value(v) for k, v in kwargs.items() if k in keys_to_include}
+    metadata = {k: load_data_value(v) for k, v in kwargs.items() if k in keys_to_include and not _is_openai_sentinel(v)}
     return metadata
 
 
@@ -822,7 +833,13 @@ def openai_get_metadata_from_response(
     metadata = {}
 
     if kwargs:
-        metadata.update({k: v for k, v in kwargs.items() if k in OPENAI_METADATA_RESPONSE_KEYS + COMMON_METADATA_KEYS})
+        metadata.update(
+            {
+                k: v
+                for k, v in kwargs.items()
+                if k in OPENAI_METADATA_RESPONSE_KEYS + COMMON_METADATA_KEYS and not _is_openai_sentinel(v)
+            }
+        )
 
     if not response:
         return metadata

releasenotes/notes/llmobs-filter-openai-omit-metadata-15533386303440c7.yaml

@@ -0,0 +1,4 @@
+---
+fixes:
+  - |
+    LLM Observability: Resolves an issue where the OpenAI integration recorded unset request parameters (OpenAI SDK's ``Omit``/``NotGiven`` sentinel values) as noise in LLM span metadata.

tests/contrib/openai/test_openai_llmobs.py

@@ -118,6 +118,44 @@ def test_chat_completion_unknown_provider(self, mock_completions_post, openai, o
         assert len(spans) == 1
         assert get_llmobs_model_provider(spans[0]) == "unknown"
 
+    @mock.patch("openai._base_client.SyncAPIClient.post")
+    def test_chat_completion_filters_openai_sentinel_metadata(
+        self, mock_completions_post, openai, openai_llmobs, test_spans
+    ):
+        """openai.Omit / openai.NotGiven sentinels for unset params must not pollute span metadata.
+
+        Frameworks like PydanticAI forward every chat-completion parameter explicitly, defaulting
+        any the caller didn't set to ``openai.omit``. Without filtering, these sentinels are
+        serialized into span metadata as noisy repr strings (e.g. "<openai.Omit object at 0x...>"),
+        making the field unqueryable. Regression test for MLOS-693.
+        """
+        mock_completions_post.return_value = mock_openai_chat_completions_response
+        # Sentinel availability varies by openai version: NotGiven was exported starting in ~1.30,
+        # Omit only in openai>=2, and neither in very old clients (e.g. 1.0.0). Only construct and
+        # exercise the sentinels this installed version actually exposes; skip if it has none.
+        not_given_cls = getattr(openai, "NotGiven", None)
+        omit_cls = getattr(openai, "Omit", None)
+        sentinel_kwargs = {}
+        if not_given_cls is not None:
+            sentinel_kwargs["presence_penalty"] = not_given_cls()
+            sentinel_kwargs["seed"] = not_given_cls()
+        if omit_cls is not None:
+            sentinel_kwargs["temperature"] = omit_cls()
+            sentinel_kwargs["frequency_penalty"] = omit_cls()
+        if not sentinel_kwargs:
+            pytest.skip("installed openai exposes no Omit/NotGiven sentinel types")
+        client = openai.OpenAI(base_url="http://localhost:8000")
+        client.chat.completions.create(
+            model="gpt-3.5-turbo",
+            messages=multi_message_input,
+            top_p=0.9,
+            **sentinel_kwargs,
+        )
+        spans = [s for trace in test_spans.pop_traces() for s in trace]
+        assert len(spans) == 1
+        # Only the explicitly-set value should survive; the Omit/NotGiven sentinels are dropped.
+        assert get_llmobs_metadata(spans[0]) == {"top_p": 0.9}
+
     @mock.patch("openai._base_client.SyncAPIClient.post")
     def test_provider_attribution_with_concurrent_openai_and_azure_clients(
         self, mock_completions_post, openai, azure_openai_config, openai_llmobs, test_spans