This PR addresses the common issue where users struggle to pass custom
parameters to OpenAI-compatible APIs like LM Studio, vLLM, and others.
The problem occurs when users try to use `model_kwargs` for custom
parameters, which causes API errors.
## Problem
Users attempting to pass custom parameters (like LM Studio's `ttl`
parameter) were getting errors:
```python
# ❌ This approach fails
llm = ChatOpenAI(
base_url="http://localhost:1234/v1",
model="mlx-community/QwQ-32B-4bit",
model_kwargs={"ttl": 5} # Causes TypeError: unexpected keyword argument 'ttl'
)
```
## Solution
The `extra_body` parameter is the correct way to pass custom parameters
to OpenAI-compatible APIs:
```python
# ✅ This approach works correctly
llm = ChatOpenAI(
base_url="http://localhost:1234/v1",
model="mlx-community/QwQ-32B-4bit",
extra_body={"ttl": 5} # Custom parameters go in extra_body
)
```
## Changes Made
1. **Enhanced Documentation**: Updated the `extra_body` parameter
docstring with comprehensive examples for LM Studio, vLLM, and other
providers
2. **Added Documentation Section**: Created a new "OpenAI-compatible
APIs" section in the main class docstring with practical examples
3. **Unit Tests**: Added tests to verify `extra_body` functionality
works correctly:
- `test_extra_body_parameter()`: Verifies custom parameters are included
in request payload
- `test_extra_body_with_model_kwargs()`: Ensures `extra_body` and
`model_kwargs` work together
4. **Clear Guidance**: Documented when to use `extra_body` vs
`model_kwargs`
## Examples Added
**LM Studio with TTL (auto-eviction):**
```python
ChatOpenAI(
base_url="http://localhost:1234/v1",
api_key="lm-studio",
model="mlx-community/QwQ-32B-4bit",
extra_body={"ttl": 300} # Auto-evict after 5 minutes
)
```
**vLLM with custom sampling:**
```python
ChatOpenAI(
base_url="http://localhost:8000/v1",
api_key="EMPTY",
model="meta-llama/Llama-2-7b-chat-hf",
extra_body={
"use_beam_search": True,
"best_of": 4
}
)
```
## Why This Works
- `model_kwargs` parameters are passed directly to the OpenAI client's
`create()` method, causing errors for non-standard parameters
- `extra_body` parameters are included in the HTTP request body, which
is exactly what OpenAI-compatible APIs expect for custom parameters
Fixes#32115.
<!-- START COPILOT CODING AGENT TIPS -->
---
💬 Share your feedback on Copilot coding agent for the chance to win a
$200 gift card! Click
[here](https://survey.alchemer.com/s3/8343779/Copilot-Coding-agent) to
start the survey.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: mdrxy <61371264+mdrxy@users.noreply.github.com>
Co-authored-by: Mason Daugherty <github@mdrxy.com>
Co-authored-by: Mason Daugherty <mason@langchain.dev>
Some providers include (legacy) function calls in `additional_kwargs` in
addition to tool calls. We currently unpack both function calls and tool
calls if present, but OpenAI will raise 400 in this case.
This can come up if providers are mixed in a tool-calling loop. Example:
```python
from langchain.chat_models import init_chat_model
from langchain_core.messages import HumanMessage
from langchain_core.tools import tool
@tool
def get_weather(location: str) -> str:
"""Get weather at a location."""
return "It's sunny."
gemini = init_chat_model("google_genai:gemini-2.0-flash-001").bind_tools([get_weather])
openai = init_chat_model("openai:gpt-4.1-mini").bind_tools([get_weather])
input_message = HumanMessage("What's the weather in Boston?")
tool_call_message = gemini.invoke([input_message])
assert len(tool_call_message.tool_calls) == 1
tool_call = tool_call_message.tool_calls[0]
tool_message = get_weather.invoke(tool_call)
response = openai.invoke( # currently raises 400 / BadRequestError
[input_message, tool_call_message, tool_message]
)
```
Here we ignore function calls if tool calls are present.
When calling `embed_documents` and providing a `chunk_size` argument,
that argument is ignored when `OpenAIEmbeddings` is instantiated with
its default configuration (where `check_embedding_ctx_length=True`).
`_get_len_safe_embeddings` specifies a `chunk_size` parameter but it's
not being passed through in `embed_documents`, which is its only caller.
This appears to be an oversight, especially given that the
`_get_len_safe_embeddings` docstring states it should respect "the set
embedding context length and chunk size."
Developers typically expect method parameters to take effect (also, take
precedence) when explicitly provided, especially when instantiating
using defaults. I was confused as to why my API calls were being
rejected regardless of the chunk size I provided.
This bug also exists in langchain_community package. I can add that to
this PR if requested otherwise I will create a new one once this passes.
Related to https://github.com/langchain-ai/langchain/issues/30344https://github.com/langchain-ai/langchain/pull/30542 introduced an
erroneous test for token counts for o-series models. tiktoken==0.8 does
not support o-series models in
`tiktoken.encoding_for_model(model_name)`, and this is the version of
tiktoken we had in the lock file. So we would default to `cl100k_base`
for o-series, which is the wrong encoding model. The test tested against
this wrong encoding (so it passed with tiktoken 0.8).
Here we update tiktoken to 0.9 in the lock file, and fix the expected
counts in the test. Verified that we are pulling
[o200k_base](https://github.com/openai/tiktoken/blob/main/tiktoken/model.py#L8),
as expected.
Resolves https://github.com/langchain-ai/langchain/issues/29003,
https://github.com/langchain-ai/langchain/issues/27264
Related: https://github.com/langchain-ai/langchain-redis/issues/52
```python
from langchain.chat_models import init_chat_model
from langchain.globals import set_llm_cache
from langchain_community.cache import SQLiteCache
from pydantic import BaseModel
cache = SQLiteCache()
set_llm_cache(cache)
class Temperature(BaseModel):
value: int
city: str
llm = init_chat_model("openai:gpt-4o-mini")
structured_llm = llm.with_structured_output(Temperature)
```
```python
# 681 ms
response = structured_llm.invoke("What is the average temperature of Rome in May?")
```
```python
# 6.98 ms
response = structured_llm.invoke("What is the average temperature of Rome in May?")
```
Some o-series models will raise a 400 error for `"role": "system"`
(`o1-mini` and `o1-preview` will raise, `o1` and `o3-mini` will not).
Here we update `ChatOpenAI` to update the role to `"developer"` for all
model names matching `^o\d`.
We only make this change on the ChatOpenAI class (not BaseChatOpenAI).
This fix ensures that the chunk size is correctly determined when
processing text embeddings. Previously, the code did not properly handle
cases where chunk_size was None, potentially leading to incorrect
chunking behavior.
Now, chunk_size_ is explicitly set to either the provided chunk_size or
the default self.chunk_size, ensuring consistent chunking. This update
improves reliability when processing large text inputs in batches and
prevents unintended behavior when chunk_size is not specified.
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
## Goal
Solve the following problems with `langchain-openai`:
- Structured output with `o1` [breaks out of the
box](https://langchain.slack.com/archives/C050X0VTN56/p1735232400232099).
- `with_structured_output` by default does not use OpenAI’s [structured
output
feature](https://platform.openai.com/docs/guides/structured-outputs).
- We override API defaults for temperature and other parameters.
## Breaking changes:
- Default method for structured output is changing to OpenAI’s dedicated
[structured output
feature](https://platform.openai.com/docs/guides/structured-outputs).
For schemas specified via TypedDict or JSON schema, strict schema
validation is disabled by default but can be enabled by specifying
`strict=True`.
- To recover previous default, pass `method="function_calling"` into
`with_structured_output`.
- Models that don’t support `method="json_schema"` (e.g., `gpt-4` and
`gpt-3.5-turbo`, currently the default model for ChatOpenAI) will raise
an error unless `method` is explicitly specified.
- To recover previous default, pass `method="function_calling"` into
`with_structured_output`.
- Schemas specified via Pydantic `BaseModel` that have fields with
non-null defaults or metadata (like min/max constraints) will raise an
error.
- To recover previous default, pass `method="function_calling"` into
`with_structured_output`.
- `strict` now defaults to False for `method="json_schema"` when schemas
are specified via TypedDict or JSON schema.
- To recover previous behavior, use `with_structured_output(schema,
strict=True)`
- Schemas specified via Pydantic V1 will raise a warning (and use
`method="function_calling"`) unless `method` is explicitly specified.
- To remove the warning, pass `method="function_calling"` into
`with_structured_output`.
- Streaming with default structured output method / Pydantic schema no
longer generates intermediate streamed chunks.
- To recover previous behavior, pass `method="function_calling"` into
`with_structured_output`.
- We no longer override default temperature (was 0.7 in LangChain, now
will follow OpenAI, currently 1.0).
- To recover previous behavior, initialize `ChatOpenAI` or
`AzureChatOpenAI` with `temperature=0.7`.
- Note: conceptually there is a difference between forcing a tool call
and forcing a response format. Tool calls may have more concise
arguments vs. generating content adhering to a schema. Prompts may need
to be adjusted to recover desired behavior.
---------
Co-authored-by: Jacob Lee <jacoblee93@gmail.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
- Convert developer openai messages to SystemMessage
- store additional_kwargs={"__openai_role__": "developer"} so that the
correct role can be reconstructed if needed
- update ChatOpenAI to read in openai_role
---------
Co-authored-by: Erick Friis <erick@langchain.dev>
