UserWarning: Pydantic serializer warnings when streaming with structured output

[Danstiv]

Submission checklist

• This is a bug, not a usage question.
• I added a clear and descriptive title that summarizes this issue.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).
• This is not related to the langchain-community package.
• I posted a self-contained, minimal, reproducible example. A maintainer can copy it and run it AS IS.

Package (Required)

• langchain
• langchain-openai
• langchain-anthropic
• langchain-classic
• langchain-core
• langchain-model-profiles
• langchain-tests
• langchain-text-splitters
• langchain-chroma
• langchain-deepseek
• langchain-exa
• langchain-fireworks
• langchain-groq
• langchain-huggingface
• langchain-mistralai
• langchain-nomic
• langchain-ollama
• langchain-openrouter
• langchain-perplexity
• langchain-qdrant
• langchain-xai
• Other / not sure / general

Related Issues / PRs

#35538

Reproduction Steps / Example Code (Python)

from langchain_openai import ChatOpenAI
from pydantic import BaseModel

llm = ChatOpenAI(model='gpt-5.4', streaming=True)

class ModelOutput(BaseModel):
    output: str


print(llm.with_structured_output(ModelOutput).invoke("What is the capital of France?"))

Error Message and Stack Trace (if applicable)

D:\temp\.venv\Lib\site-packages\pydantic\main.py:464: UserWarning: Pydantic serializer warnings:
  PydanticSerializationUnexpectedValue(Expected `none` - serialized value may not be as expected [field_name='parsed', input_value=ModelOutput(output='Paris.'), input_type=ModelOutput])
  return self.__pydantic_serializer__.to_python(
output='Paris.'

Description

Enabling streaming causes warning demonstrated in the example
It happens here

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

Line 1673 in ce21bf4

chunk = chunk.model_dump()

Warning when model invoked without streaming was fixed in #35543, but for streaming issue remains.

System Info

System Information

OS: Windows
OS Version: 10.0.26200
Python Version: 3.14.6 (tags/v3.14.6:c63aec6, Jun 10 2026, 10:26:10) [MSC v.1944 64 bit (AMD64)]

Package Information

langchain_core: 1.4.3
langchain: 1.3.7
langsmith: 0.8.14
langchain_openai: 1.3.0
langchain_protocol: 0.0.16
langgraph_sdk: 0.4.2

Optional packages not installed

deepagents
deepagents-cli

Other Dependencies

httpx: 0.28.1
jsonpatch: 1.33
langgraph: 1.2.4
openai: 2.41.1
orjson: 3.11.9
packaging: 26.2
pydantic: 2.13.4
pyyaml: 6.0.3
requests: 2.34.2
requests-toolbelt: 1.0.0
tenacity: 9.1.4
tiktoken: 0.13.0
typing-extensions: 4.15.0
uuid-utils: 0.16.0
websockets: 15.0.1
xxhash: 3.7.0
zstandard: 0.25.0

[gitbalaji]

Hi, I'd like to work on this. Could a maintainer please assign it to me?

[AyushDas4890]

Confirmed on current master — and there are two call sites, not one: the sync stream (base.py:1601) and the async stream (base.py:1866) both do a bare chunk.model_dump().

The non-streaming fix from #35543 is visible in the same file at ~1748:

response.model_dump(
    exclude={"choices": {"__all__": {"message": {"parsed"}}}}
)

with the typed parsed copied back from the response object afterwards. The streaming fix can mirror that pattern — exclude parsed at its chunk location (under delta rather than message) and reattach the typed object — applied to both the sync and async generators.

A regression test probably wants streaming=True + with_structured_output(...) inside warnings.catch_warnings(record=True) asserting nothing is emitted, since the existing non-streaming test from #35543 can't catch this path.

[xy200303]

I would be happy to work on this if the direction looks good.

I traced the current streaming path locally and agree that this is not limited to a single call site. The warning seems to come from the legacy streaming aggregation path in libs/core/langchain_core/language_models/chat_models.py, where both the sync and async branches serialize chunks with a bare model_dump() while structured output can carry a typed parsed payload.

My proposed fix is to mirror the non-streaming pattern from #35543 in the streaming path as well:

• handle both the sync _stream(...) branch and the async _astream(...) branch,
• exclude the typed parsed field from the serialized payload at the chunk/delta location that triggers the warning, and
• preserve the typed parsed object on the LangChain-side message object so behavior stays the same apart from removing the warning.

For verification, I would add a focused regression test that uses streaming=True together with with_structured_output(...) under warnings.catch_warnings(record=True) and asserts that the streaming path emits no Pydantic serializer warnings. If this approach aligns with what maintainers want, could someone please assign the issue?