fix(openai): avoid PydanticSerializationUnexpectedValue for structured output (#35543)

libs/partners/openai/langchain_openai/chat_models/base.py

@@ -1548,7 +1548,13 @@ class BaseChatOpenAI(BaseChatModel):
         generations = []
 
         response_dict = (
-            response if isinstance(response, dict) else response.model_dump()
+            response
+            if isinstance(response, dict)
+            # `parsed` may hold arbitrary Pydantic models from structured output.
+            # Exclude it from this dump and copy it from the typed response below.
+            else response.model_dump(
+                exclude={"choices": {"__all__": {"message": {"parsed"}}}}
+            )
         )
         # Sometimes the AI Model calling will get error, we should raise it (this is
         # typically followed by a null value for `choices`, which we raise for

libs/partners/openai/tests/unit_tests/chat_models/test_base.py

@@ -1388,6 +1388,49 @@ def test_structured_outputs_parser() -> None:
     assert result == parsed_response
 
 
+def test_create_chat_result_avoids_parsed_model_dump_warning() -> None:
+    class ModelOutput(BaseModel):
+        output: str
+
+    class MockParsedMessage(openai.BaseModel):
+        role: Literal["assistant"] = "assistant"
+        content: str = '{"output": "Paris"}'
+        parsed: None = None
+        refusal: str | None = None
+
+    class MockChoice(openai.BaseModel):
+        index: int = 0
+        finish_reason: Literal["stop"] = "stop"
+        message: MockParsedMessage
+
+    class MockChatCompletion(openai.BaseModel):
+        id: str = "chatcmpl-1"
+        object: str = "chat.completion"
+        created: int = 0
+        model: str = "gpt-4o-mini"
+        choices: list[MockChoice]
+        usage: dict[str, int] | None = None
+
+    parsed_response = ModelOutput(output="Paris")
+    response = MockChatCompletion.model_construct(
+        choices=[
+            MockChoice.model_construct(
+                message=MockParsedMessage.model_construct(parsed=parsed_response)
+            )
+        ],
+        usage={"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+    )
+
+    llm = ChatOpenAI(model="gpt-4o-mini")
+    with warnings.catch_warnings(record=True) as caught_warnings:
+        warnings.simplefilter("always")
+        result = llm._create_chat_result(response)
+
+    warning_messages = [str(warning.message) for warning in caught_warnings]
+    assert not any("field_name='parsed'" in message for message in warning_messages)
+    assert result.generations[0].message.additional_kwargs["parsed"] == parsed_response
+
+
 def test_structured_outputs_parser_valid_falsy_response() -> None:
     class LunchBox(BaseModel):
         sandwiches: list[str]