Skip to content

The agent engineering platform.

fix(openai): Always add raw response object to OpenAI client errors for invoke (#32655)

Jacob Lee1459d4f4ce02
diff --git a/libs/partners/openai/langchain_openai/chat_models/base.py b/libs/partners/openai/langchain_openai/chat_models/base.py
index fefa81ea42..afdcf47445 100644
--- a/libs/partners/openai/langchain_openai/chat_models/base.py
+++ b/libs/partners/openai/langchain_openai/chat_models/base.py
@@ -1142,42 +1142,51 @@ class BaseChatOpenAI(BaseChatModel):
return generate_from_stream(stream_iter)
payload = self._get_request_payload(messages, stop=stop, **kwargs)
generation_info = None
- if "response_format" in payload:
- if self.include_response_headers:
- warnings.warn(
- "Cannot currently include response headers when response_format is "
- "specified."
- )
- payload.pop("stream")
- try:
- response = self.root_client.beta.chat.completions.parse(**payload)
- except openai.BadRequestError as e:
- _handle_openai_bad_request(e)
- elif self._use_responses_api(payload):
- original_schema_obj = kwargs.get("response_format")
- if original_schema_obj and _is_pydantic_class(original_schema_obj):
- response = self.root_client.responses.parse(**payload)
- else:
- if self.include_response_headers:
- raw_response = self.root_client.with_raw_response.responses.create(
- **payload
+ raw_response = None
+ try:
+ if "response_format" in payload:
+ payload.pop("stream")
+ try:
+ raw_response = (
+ self.root_client.chat.completions.with_raw_response.parse(
+ **payload
+ )
)
response = raw_response.parse()
- generation_info = {"headers": dict(raw_response.headers)}
+ except openai.BadRequestError as e:
+ _handle_openai_bad_request(e)
+ elif self._use_responses_api(payload):
+ original_schema_obj = kwargs.get("response_format")
+ if original_schema_obj and _is_pydantic_class(original_schema_obj):
+ raw_response = self.root_client.responses.with_raw_response.parse(
+ **payload
+ )
else:
- response = self.root_client.responses.create(**payload)
- return _construct_lc_result_from_responses_api(
- response,
- schema=original_schema_obj,
- metadata=generation_info,
- output_version=self.output_version,
- )
- elif self.include_response_headers:
- raw_response = self.client.with_raw_response.create(**payload)
- response = raw_response.parse()
+ raw_response = self.root_client.responses.with_raw_response.create(
+ **payload
+ )
+ response = raw_response.parse()
+ if self.include_response_headers:
+ generation_info = {"headers": dict(raw_response.headers)}
+ return _construct_lc_result_from_responses_api(
+ response,
+ schema=original_schema_obj,
+ metadata=generation_info,
+ output_version=self.output_version,
+ )
+ else:
+ raw_response = self.client.with_raw_response.create(**payload)
+ response = raw_response.parse()
+ except Exception as e:
+ if raw_response is not None and hasattr(raw_response, "http_response"):
+ e.response = raw_response.http_response # type: ignore[attr-defined]
+ raise e
+ if (
+ self.include_response_headers
+ and raw_response is not None
+ and hasattr(raw_response, "headers")
+ ):
generation_info = {"headers": dict(raw_response.headers)}
- else:
- response = self.client.create(**payload)
return self._create_chat_result(response, generation_info)
def _use_responses_api(self, payload: dict) -> bool:
@@ -1375,46 +1384,55 @@ class BaseChatOpenAI(BaseChatModel):
return await agenerate_from_stream(stream_iter)
payload = self._get_request_payload(messages, stop=stop, **kwargs)
generation_info = None
- if "response_format" in payload:
- if self.include_response_headers:
- warnings.warn(
- "Cannot currently include response headers when response_format is "
- "specified."
- )
- payload.pop("stream")
- try:
- response = await self.root_async_client.beta.chat.completions.parse(
- **payload
- )
- except openai.BadRequestError as e:
- _handle_openai_bad_request(e)
- elif self._use_responses_api(payload):
- original_schema_obj = kwargs.get("response_format")
- if original_schema_obj and _is_pydantic_class(original_schema_obj):
- response = await self.root_async_client.responses.parse(**payload)
- else:
- if self.include_response_headers:
+ raw_response = None
+ try:
+ if "response_format" in payload:
+ payload.pop("stream")
+ try:
+ raw_response = await self.root_async_client.chat.completions.with_raw_response.parse( # noqa: E501
+ **payload
+ )
+ response = raw_response.parse()
+ except openai.BadRequestError as e:
+ _handle_openai_bad_request(e)
+ elif self._use_responses_api(payload):
+ original_schema_obj = kwargs.get("response_format")
+ if original_schema_obj and _is_pydantic_class(original_schema_obj):
raw_response = (
- await self.root_async_client.with_raw_response.responses.create(
+ await self.root_async_client.responses.with_raw_response.parse(
**payload
)
)
- response = raw_response.parse()
- generation_info = {"headers": dict(raw_response.headers)}
else:
- response = await self.root_async_client.responses.create(**payload)
- return _construct_lc_result_from_responses_api(
- response,
- schema=original_schema_obj,
- metadata=generation_info,
- output_version=self.output_version,
- )
- elif self.include_response_headers:
- raw_response = await self.async_client.with_raw_response.create(**payload)
- response = raw_response.parse()
+ raw_response = (
+ await self.root_async_client.responses.with_raw_response.create(
+ **payload
+ )
+ )
+ response = raw_response.parse()
+ if self.include_response_headers:
+ generation_info = {"headers": dict(raw_response.headers)}
+ return _construct_lc_result_from_responses_api(
+ response,
+ schema=original_schema_obj,
+ metadata=generation_info,
+ output_version=self.output_version,
+ )
+ else:
+ raw_response = await self.async_client.with_raw_response.create(
+ **payload
+ )
+ response = raw_response.parse()
+ except Exception as e:
+ if raw_response is not None and hasattr(raw_response, "http_response"):
+ e.response = raw_response.http_response # type: ignore[attr-defined]
+ raise e
+ if (
+ self.include_response_headers
+ and raw_response is not None
+ and hasattr(raw_response, "headers")
+ ):
generation_info = {"headers": dict(raw_response.headers)}
- else:
- response = await self.async_client.create(**payload)
return await run_in_executor(
None, self._create_chat_result, response, generation_info
)
diff --git a/libs/partners/openai/tests/cassettes/test_schema_parsing_failures.yaml.gz b/libs/partners/openai/tests/cassettes/test_schema_parsing_failures.yaml.gz
new file mode 100644
index 0000000000..3f58e1a81b
Binary files /dev/null and b/libs/partners/openai/tests/cassettes/test_schema_parsing_failures.yaml.gz differ
diff --git a/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_async.yaml.gz b/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_async.yaml.gz
new file mode 100644
index 0000000000..e9f29056c5
Binary files /dev/null and b/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_async.yaml.gz differ
diff --git a/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_responses_api.yaml.gz b/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_responses_api.yaml.gz
new file mode 100644
index 0000000000..f833103aa3
Binary files /dev/null and b/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_responses_api.yaml.gz differ
diff --git a/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_responses_api_async.yaml.gz b/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_responses_api_async.yaml.gz
new file mode 100644
index 0000000000..344e172f79
Binary files /dev/null and b/libs/partners/openai/tests/cassettes/test_schema_parsing_failures_responses_api_async.yaml.gz differ
diff --git a/libs/partners/openai/tests/integration_tests/chat_models/test_base.py b/libs/partners/openai/tests/integration_tests/chat_models/test_base.py
index b914af07cb..67177a2ef5 100644
--- a/libs/partners/openai/tests/integration_tests/chat_models/test_base.py
+++ b/libs/partners/openai/tests/integration_tests/chat_models/test_base.py
@@ -27,7 +27,7 @@ from langchain_tests.integration_tests.chat_models import (
_validate_tool_call_message,
magic_function,
)
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, field_validator
from langchain_openai import ChatOpenAI
from tests.unit_tests.fake.callbacks import FakeCallbackHandler
@@ -1155,3 +1155,62 @@ def test_prompt_cache_key_usage_methods_integration() -> None:
response_model_level = chat_model_level.invoke(messages)
assert isinstance(response_model_level, AIMessage)
assert isinstance(response_model_level.content, str)
+
+
+class BadModel(BaseModel):
+ response: str
+
+ @field_validator("response")
+ @classmethod
+ def validate_response(cls, v: str) -> str:
+ if v != "bad":
+ raise ValueError('response must be exactly "bad"')
+ return v
+
+
+# VCR can't handle parameterized tests
+def test_schema_parsing_failures() -> None:
+ llm = ChatOpenAI(model="gpt-5-nano", use_responses_api=False)
+ try:
+ llm.invoke("respond with good", response_format=BadModel)
+ except Exception as e:
+ assert e.response is not None # type: ignore[attr-defined]
+ else:
+ assert False
+
+
+# VCR can't handle parameterized tests
+def test_schema_parsing_failures_responses_api() -> None:
+ llm = ChatOpenAI(model="gpt-5-nano", use_responses_api=True)
+ try:
+ llm.invoke("respond with good", response_format=BadModel)
+ except Exception as e:
+ assert e.response is not None # type: ignore[attr-defined]
+ else:
+ assert False
+
+
+# VCR can't handle parameterized tests
+async def test_schema_parsing_failures_async() -> None:
+ llm = ChatOpenAI(model="gpt-5-nano", use_responses_api=False)
+ try:
+ await llm.ainvoke("respond with good", response_format=BadModel)
+ except Exception as e:
+ assert e.response is not None # type: ignore[attr-defined]
+ else:
+ assert False
+
+
+# VCR can't handle parameterized tests
+async def test_schema_parsing_failures_responses_api_async() -> None:
+ llm = ChatOpenAI(model="gpt-5-nano", use_responses_api=True)
+ try:
+ await llm.ainvoke("respond with good", response_format=BadModel)
+ except Exception as e:
+ assert e.response is not None # type: ignore[attr-defined]
+ else:
+ assert False
diff --git a/libs/partners/openai/tests/unit_tests/chat_models/test_base.py b/libs/partners/openai/tests/unit_tests/chat_models/test_base.py
index 68caee5d63..2c33fe062b 100644
--- a/libs/partners/openai/tests/unit_tests/chat_models/test_base.py
+++ b/libs/partners/openai/tests/unit_tests/chat_models/test_base.py
@@ -601,7 +601,7 @@ def test_openai_invoke(mock_client: MagicMock) -> None:
# headers are not in response_metadata if include_response_headers not set
assert "headers" not in res.response_metadata
- assert mock_client.create.called
+ assert mock_client.with_raw_response.create.called
async def test_openai_ainvoke(mock_async_client: AsyncMock) -> None:
@@ -613,7 +613,7 @@ async def test_openai_ainvoke(mock_async_client: AsyncMock) -> None:
# headers are not in response_metadata if include_response_headers not set
assert "headers" not in res.response_metadata
- assert mock_async_client.create.called
+ assert mock_async_client.with_raw_response.create.called
@pytest.mark.parametrize(
@@ -638,7 +638,7 @@ def test_openai_invoke_name(mock_client: MagicMock) -> None:
with patch.object(llm, "client", mock_client):
messages = [HumanMessage(content="Foo", name="Katie")]
res = llm.invoke(messages)
- call_args, call_kwargs = mock_client.create.call_args
+ call_args, call_kwargs = mock_client.with_raw_response.create.call_args
assert len(call_args) == 0 # no positional args
call_messages = call_kwargs["messages"]
assert len(call_messages) == 1
@@ -678,7 +678,7 @@ def test_function_calls_with_tool_calls(mock_client: MagicMock) -> None:
]
with patch.object(llm, "client", mock_client):
_ = llm.invoke(messages)
- _, call_kwargs = mock_client.create.call_args
+ _, call_kwargs = mock_client.with_raw_response.create.call_args
call_messages = call_kwargs["messages"]
tool_call_message_payload = call_messages[1]
assert "tool_calls" in tool_call_message_payload
@@ -688,7 +688,7 @@ def test_function_calls_with_tool_calls(mock_client: MagicMock) -> None:
cast(AIMessage, messages[1]).tool_calls = []
with patch.object(llm, "client", mock_client):
_ = llm.invoke(messages)
- _, call_kwargs = mock_client.create.call_args
+ _, call_kwargs = mock_client.with_raw_response.create.call_args
call_messages = call_kwargs["messages"]
tool_call_message_payload = call_messages[1]
assert "function_call" in tool_call_message_payload
@@ -2326,8 +2326,9 @@ def test_mcp_tracing() -> None:
tracer = FakeTracer()
mock_client = MagicMock()
- def mock_create(*args: Any, **kwargs: Any) -> Response:
- return Response(
+ def mock_create(*args: Any, **kwargs: Any) -> MagicMock:
+ mock_raw_response = MagicMock()
+ mock_raw_response.parse.return_value = Response(
id="resp_123",
created_at=1234567890,
model="o4-mini",
@@ -2349,8 +2350,9 @@ def test_mcp_tracing() -> None:
)
],
)
+ return mock_raw_response
- mock_client.responses.create = mock_create
+ mock_client.responses.with_raw_response.create = mock_create
input_message = HumanMessage("Test query")
tools = [
{