aidial-integration-langchain 0.2.1

Utilities for integration of AI DIAL API with Langchain library

Overview

The package provides helpers facilitating integration of DIAL API with Langchain library.

Passing DIAL-specific extra fields in Langchain requests/responses

Unlike openai library, langchain-openai library doesn't allow to pass extra request/response parameters to/from the upstream model.

The minimal example highlighting the issue could be found in the example folder:

cd example
python -m venv .venv
source .venv/bin/activate
pip install -q -r requirements.txt
python -m app

Received the following extra fields:
(1) ☑ request.extra_field
(2) ☑ request.tools[0].extra_field
(3) ☐ request.messages[0].extra_field
(4) ☐ response.message.extra_field
(5) ☐ response.extra_field

langchain-openai ignores certain extra fields, meaning that the upstream endpoint won't receive (2) and the client won't receive (4) and (5) if they were sent by the upstream.

Note that top-level request extra fields and tool definition extra fields do actually reach the upstream.

Solution

One way to fix the issue, is to modify the Langchain methods that ignore these extra fields so that they are taken into account.

This is achieved via monkey-patching certain private methods in langchain-openai that do the conversion from the Langchain datatypes to dictionaries and vice versa.

Usage

Install the aidial-integration-langchain package as a dependency in your project:

pip install aidial-integration-langchain

Then import aidial_integration_langchain before importing any Langchain module to apply the patches:

import aidial_integration_langchain.patch # isort:skip  # noqa: F401 # type: ignore

from langchain_openai import AzureChatOpenAI
...

The same example as above, but with the patch applied:

cd example
python -m venv .venv
source .venv/bin/activate
pip install -q -r requirements.txt
pip install -q aidial-integration-langchain
python -m app patch

Received the following extra fields:
(1) ☑ request.extra_field
(2) ☑ request.tools[0].extra_field
(3) ☑ request.messages[0].extra_field
(4) ☑ response.message.extra_field
(5) ☑ response.extra_field

Supported Langchain versions

The following langchain-openai versions have been tested for Python 3.9, 3.10, 3.11, 3.12, and 3.13:

Version	Request per-message	Response per-message	Response top-level
>=0.1.1,<=0.1.22	🟢	🟢	🔴
>=0.1.23,<=0.1.25	🟢	🟢	🟢
>=0.2.0,<=0.2.14	🟢	🟢	🟢
>=0.3.0,<=0.3.18	🟢	🟢	🟢

[!NOTE] The patched langchain-openai<=0.1.22 doesn't support response top-level extra fields, since the structure of the code back then was not very amicable for monkey-patching in this particular respect.

Configuration

The list of extra fields that are allowed to pass-through is controlled by the following environment variables.

Name	Default	Description
LC_EXTRA_REQUEST_MESSAGE_FIELDS	custom_content	A comma-separated list of extra message fields allowed to pass-through in chat completion requests.
LC_EXTRA_RESPONSE_MESSAGE_FIELDS	custom_content	A comma-separated list of extra message fields allowed to pass-through in chat completion responses.
LC_EXTRA_RESPONSE_FIELDS	statistics	A comma-separated list of extra fields allowed to pass-through on the top-level of the chat completion responses.