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.
When OpenAI originally released `stream_options` to enable token usage
during streaming, it was not supported in AzureOpenAI. It is now
supported.
Like the [OpenAI
SDK](f66d2e6fdc/src/openai/resources/completions.py (L68)),
ChatOpenAI does not return usage metadata during streaming by default
(which adds an extra chunk to the stream). The OpenAI SDK requires users
to pass `stream_options={"include_usage": True}`. ChatOpenAI implements
a convenience argument `stream_usage: Optional[bool]`, and an attribute
`stream_usage: bool = False`.
Here we extend this to AzureChatOpenAI by moving the `stream_usage`
attribute and `stream_usage` kwarg (on `_(a)stream`) from ChatOpenAI to
BaseChatOpenAI.
---
Additional consideration: we must be sensitive to the number of users
using BaseChatOpenAI to interact with other APIs that do not support the
`stream_options` parameter.
Suppose OpenAI in the future updates the default behavior to stream
token usage. Currently, BaseChatOpenAI only passes `stream_options` if
`stream_usage` is True, so there would be no way to disable this new
default behavior.
To address this, we could update the `stream_usage` attribute to
`Optional[bool] = None`, but this is technically a breaking change (as
currently values of False are not passed to the client). IMO: if / when
this change happens, we could accompany it with this update in a minor
bump.
---
Related previous PRs:
- https://github.com/langchain-ai/langchain/pull/22628
- https://github.com/langchain-ai/langchain/pull/22854
- https://github.com/langchain-ai/langchain/pull/23552
---------
Co-authored-by: Eugene Yurtsev <eyurtsev@gmail.com>
- Support thinking blocks in core's `convert_to_openai_messages` (pass
through instead of error)
- Ignore thinking blocks in ChatOpenAI (instead of error)
- Support Anthropic-style image blocks in ChatOpenAI
---
Standard integration tests include a `supports_anthropic_inputs`
property which is currently enabled only for tests on `ChatAnthropic`.
This test enforces compatibility with message histories of the form:
```
- system message
- human message
- AI message with tool calls specified only through `tool_use` content blocks
- human message containing `tool_result` and an additional `text` block
```
It additionally checks support for Anthropic-style image inputs if
`supports_image_inputs` is enabled.
Here we change this test, such that if you enable
`supports_anthropic_inputs`:
- You support AI messages with text and `tool_use` content blocks
- You support Anthropic-style image inputs (if `supports_image_inputs`
is enabled)
- You support thinking content blocks.
That is, we add a test case for thinking content blocks, but we also
remove the requirement of handling tool results within HumanMessages
(motivated by existing agent abstractions, which should all return
ToolMessage). We move that requirement to a ChatAnthropic-specific test.
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>
- **Description:** Before sending a completion chunk at the end of an
OpenAI stream, removing the tool_calls as those have already been sent
as chunks.
- **Issue:** -
- **Dependencies:** -
- **Twitter handle:** -
@ccurme as mentioned in another PR
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
The tokens I get are:
```
['', '\n\n', 'The', ' sun', ' was', ' setting', ' over', ' the', ' horizon', ',', ' casting', '']
```
so possibly an extra empty token is included in the output.
lmk @efriis if we should look into this further.
## 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>
From what I can tell response using SDK is not deterministic:
```python
import numpy as np
import openai
documents = ["disallowed special token '<|endoftext|>'"]
model = "text-embedding-ada-002"
direct_output_1 = (
openai.OpenAI()
.embeddings.create(input=documents, model=model)
.data[0]
.embedding
)
for i in range(10):
direct_output_2 = (
openai.OpenAI()
.embeddings.create(input=documents, model=model)
.data[0]
.embedding
)
print(f"{i}: {np.isclose(direct_output_1, direct_output_2).all()}")
```
```
0: True
1: True
2: True
3: True
4: False
5: True
6: True
7: True
8: True
9: True
```
See related discussion here:
https://community.openai.com/t/can-text-embedding-ada-002-be-made-deterministic/318054
Found the same result using `"text-embedding-3-small"`.
Given the current erroring behavior, every time we've moved a kwarg from
model_kwargs and made it its own field that was a breaking change.
Updating this behavior to support the old instantiations /
serializations.
Assuming build_extra_kwargs was not something that itself is being used
externally and needs to be kept backwards compatible
Chunking of the input array controlled by `self.chunk_size` is being
ignored when `self.check_embedding_ctx_length` is disabled. Effectively,
the chunk size is assumed to be equal 1 in such a case. This is
suprising.
The PR takes into account `self.chunk_size` passed by the user.
---------
Co-authored-by: Erick Friis <erick@langchain.dev>
- **Description:** This is a **one line change**. the
`self.async_client.with_raw_response.create(**payload)` call is not
properly awaited within the `_astream` method. In `_agenerate` this is
done already, but likely forgotten in the other method.
- **Issue:** Not applicable
- **Dependencies:** No dependencies required.
(If no one reviews your PR within a few days, please @-mention one of
baskaryan, efriis, eyurtsev, ccurme, vbarda, hwchase17.)
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
- [ ] **PR message**:
- **Description:** Compatible with other llm (eg: deepseek-chat, glm-4)
usage meta data
- **Issue:** N/A
- **Dependencies:** no new dependencies added
- [ ] **Add tests and docs**:
libs/partners/openai/tests/unit_tests/chat_models/test_base.py
```shell
cd libs/partners/openai
poetry run pytest tests/unit_tests/chat_models/test_base.py::test_openai_astream
poetry run pytest tests/unit_tests/chat_models/test_base.py::test_openai_stream
poetry run pytest tests/unit_tests/chat_models/test_base.py::test_deepseek_astream
poetry run pytest tests/unit_tests/chat_models/test_base.py::test_deepseek_stream
poetry run pytest tests/unit_tests/chat_models/test_base.py::test_glm4_astream
poetry run pytest tests/unit_tests/chat_models/test_base.py::test_glm4_stream
```
---------
Co-authored-by: hyman <hyman@xiaozancloud.com>
Co-authored-by: Erick Friis <erick@langchain.dev>
Hello.
First of all, thank you for maintaining such a great project.
## Description
In https://github.com/langchain-ai/langchain/pull/25123, support for
structured_output is added. However, `"additionalProperties": false`
needs to be included at all levels when a nested object is generated.
error from current code:
https://gist.github.com/fufufukakaka/e9b475300e6934853d119428e390f204
```
BadRequestError: Error code: 400 - {'error': {'message': "Invalid schema for response_format 'JokeWithEvaluation': In context=('properties', 'self_evaluation'), 'additionalProperties' is required to be supplied and to be false", 'type': 'invalid_request_error', 'param': 'response_format', 'code': None}}
```
Reference: [Introducing Structured Outputs in the
API](https://openai.com/index/introducing-structured-outputs-in-the-api/)
```json
{
"model": "gpt-4o-2024-08-06",
"messages": [
{
"role": "system",
"content": "You are a helpful math tutor."
},
{
"role": "user",
"content": "solve 8x + 31 = 2"
}
],
"response_format": {
"type": "json_schema",
"json_schema": {
"name": "math_response",
"strict": true,
"schema": {
"type": "object",
"properties": {
"steps": {
"type": "array",
"items": {
"type": "object",
"properties": {
"explanation": {
"type": "string"
},
"output": {
"type": "string"
}
},
"required": ["explanation", "output"],
"additionalProperties": false
}
},
"final_answer": {
"type": "string"
}
},
"required": ["steps", "final_answer"],
"additionalProperties": false
}
}
}
}
```
In the current code, `"additionalProperties": false` is only added at
the last level.
This PR introduces the `_add_additional_properties_key` function, which
recursively adds `"additionalProperties": false` to the entire JSON
schema for the request.
Twitter handle: `@fukkaa1225`
Thank you!
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
This fix is for #21726. When having other packages installed that
require the `openai_api_base` environment variable, users are not able
to instantiate the AzureChatModels or AzureEmbeddings.
This PR adds a new value `ignore_openai_api_base` which is a bool. When
set to True, it sets `openai_api_base` to `None`
Two new tests were added for the `test_azure` and a new file
`test_azure_embeddings`
A different approach may be better for this. If you can think of better
logic, let me know and I can adjust it.
---------
Co-authored-by: Erick Friis <erick@langchain.dev>
- add test for structured output
- fix bug with structured output for Azure
- better testing on Groq (break out Mixtral + Llama3 and add xfails
where needed)
- Refactor standard test classes to make them easier to configure
- Update openai to support stop_sequences init param
- Update groq to support stop_sequences init param
- Update fireworks to support max_retries init param
- Update ChatModel.bind_tools to type tool_choice
- Update groq to handle tool_choice="any". **this may be controversial**
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
Here we add `stream_usage` to ChatOpenAI as:
1. a boolean attribute
2. a kwarg to _stream and _astream.
Question: should the `stream_usage` attribute be `bool`, or `bool |
None`?
Currently I've kept it `bool` and defaulted to False. It was implemented
on
[ChatAnthropic](e832bbb486/libs/partners/anthropic/langchain_anthropic/chat_models.py (L535))
as a bool. However, to maintain support for users who access the
behavior via OpenAI's `stream_options` param, this ends up being
possible:
```python
llm = ChatOpenAI(model_kwargs={"stream_options": {"include_usage": True}})
assert not llm.stream_usage
```
(and this model will stream token usage).
Some options for this:
- it's ok
- make the `stream_usage` attribute bool or None
- make an \_\_init\_\_ for ChatOpenAI, set a `._stream_usage` attribute
and read `.stream_usage` from a property
Open to other ideas as well.
Adds `response_metadata` to stream responses from OpenAI. This is
returned with `invoke` normally, but wasn't implemented for `stream`.
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
ChatOpenAI supports a kwarg `stream_options` which can take values
`{"include_usage": True}` and `{"include_usage": False}`.
Setting include_usage to True adds a message chunk to the end of the
stream with usage_metadata populated. In this case the final chunk no
longer includes `"finish_reason"` in the `response_metadata`. This is
the current default and is not yet released. Because this could be
disruptive to workflows, here we remove this default. The default will
now be consistent with OpenAI's API (see parameter
[here](https://platform.openai.com/docs/api-reference/chat/create#chat-create-stream_options)).
Examples:
```python
from langchain_openai import ChatOpenAI
llm = ChatOpenAI()
for chunk in llm.stream("hi"):
print(chunk)
```
```
content='' id='run-8cff4721-2acd-4551-9bf7-1911dae46b92'
content='Hello' id='run-8cff4721-2acd-4551-9bf7-1911dae46b92'
content='!' id='run-8cff4721-2acd-4551-9bf7-1911dae46b92'
content='' response_metadata={'finish_reason': 'stop'} id='run-8cff4721-2acd-4551-9bf7-1911dae46b92'
```
```python
for chunk in llm.stream("hi", stream_options={"include_usage": True}):
print(chunk)
```
```
content='' id='run-39ab349b-f954-464d-af6e-72a0927daa27'
content='Hello' id='run-39ab349b-f954-464d-af6e-72a0927daa27'
content='!' id='run-39ab349b-f954-464d-af6e-72a0927daa27'
content='' response_metadata={'finish_reason': 'stop'} id='run-39ab349b-f954-464d-af6e-72a0927daa27'
content='' id='run-39ab349b-f954-464d-af6e-72a0927daa27' usage_metadata={'input_tokens': 8, 'output_tokens': 9, 'total_tokens': 17}
```
```python
llm = ChatOpenAI().bind(stream_options={"include_usage": True})
for chunk in llm.stream("hi"):
print(chunk)
```
```
content='' id='run-59918845-04b2-41a6-8d90-f75fb4506e0d'
content='Hello' id='run-59918845-04b2-41a6-8d90-f75fb4506e0d'
content='!' id='run-59918845-04b2-41a6-8d90-f75fb4506e0d'
content='' response_metadata={'finish_reason': 'stop'} id='run-59918845-04b2-41a6-8d90-f75fb4506e0d'
content='' id='run-59918845-04b2-41a6-8d90-f75fb4506e0d' usage_metadata={'input_tokens': 8, 'output_tokens': 9, 'total_tokens': 17}
```
OpenAI recently added a `stream_options` parameter to its chat
completions API (see [release
notes](https://platform.openai.com/docs/changelog/added-chat-completions-stream-usage)).
When this parameter is set to `{"usage": True}`, an extra "empty"
message is added to the end of a stream containing token usage. Here we
propagate token usage to `AIMessage.usage_metadata`.
We enable this feature by default. Streams would now include an extra
chunk at the end, **after** the chunk with
`response_metadata={'finish_reason': 'stop'}`.
New behavior:
```
[AIMessageChunk(content='', id='run-4b20dbe0-3817-4f62-b89d-03ef76f25bde'),
AIMessageChunk(content='Hello', id='run-4b20dbe0-3817-4f62-b89d-03ef76f25bde'),
AIMessageChunk(content='!', id='run-4b20dbe0-3817-4f62-b89d-03ef76f25bde'),
AIMessageChunk(content='', response_metadata={'finish_reason': 'stop'}, id='run-4b20dbe0-3817-4f62-b89d-03ef76f25bde'),
AIMessageChunk(content='', id='run-4b20dbe0-3817-4f62-b89d-03ef76f25bde', usage_metadata={'input_tokens': 8, 'output_tokens': 9, 'total_tokens': 17})]
```
Old behavior (accessible by passing `stream_options={"include_usage":
False}` into (a)stream:
```
[AIMessageChunk(content='', id='run-1312b971-c5ea-4d92-9015-e6604535f339'),
AIMessageChunk(content='Hello', id='run-1312b971-c5ea-4d92-9015-e6604535f339'),
AIMessageChunk(content='!', id='run-1312b971-c5ea-4d92-9015-e6604535f339'),
AIMessageChunk(content='', response_metadata={'finish_reason': 'stop'}, id='run-1312b971-c5ea-4d92-9015-e6604535f339')]
```
From what I can tell this is not yet implemented in Azure, so we enable
only for ChatOpenAI.
