AzureChatOpenAI to langchain_community/llms/__init__.py

[a107029]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Add
AzureChatOpenAi to langchain_community/llms/__init__.py which would allow us to instantiate chains from configs using loaders such as: _load_stuff_documents_chain

Motivation

Azure is a big necessity for making usage of LLMs in the private-sector, including, but not limited to those offered by OpenAI, as many corporations have existing partnerships with Microsoft. This makes Azure the first choice for deployment of LLMs as it comes with the backing of Microsoft Azure for cloud security.

Currently, AzureChatOpenAi is missing from langchain_community/llms/init.py, making it impossible to load a chain from configs if it makes usage of AzureChatOpenAI where we are getting the value error:

ValueError: Loading azure-openai-chat LLM not supported

when attempting to, for example, use _load_stuff_documents_chain.

Proposal (If applicable)

At face-value, the only modification that should need to happen is in langchain_community/llms/__init__.py.. I believe this would be:

.
.
# The uncommented is preferable as it is more general and follows the existing pattern
# from langchain_openai.chat_models.base import BaseChatOpenAI
from langchain_core.language_models.chat_models import BaseChatModel
.
.
def _import_azure_openai_chat() -> BaseChatModel:
    from langchain_openai.chat_models import AzureChatOpenAI

    return AzureChatOpenAI

If there is any nuance beyond what I mentioned, or existing work in the pipeline to accomplish this I apologize. I did my best to check ahead of time and found nothing. I am eager to learn more and kick off a discussion to help make Langchain that much better.

[eyurtsev]

Use the langchain_openai partner package

[a107029]

Hey. Thank you for the response @eyurtsev . I am, and I think you've missed my point here or I failed to adequately describe it. Can you please elaborate on what you mean? Your response was too short for me to make sense of it. I'll include some pseudo-code below that will result in the error per my original suggestion. Let me know if this helps.

[a107029]

Here is a minimal example that will fail with the error I mentioned above:

from langchain.chains.combine_documents.stuff import StuffDocumentsChain
from langchain.chains.llm import LLMChain
from langchain.chains.loading import _load_stuff_documents_chain
from langchain_openai.chat_models import AzureChatOpenAI


AZURE_GPT_KWARGS = {
    "model_name": "gpt-4-32k",
    "azure_endpoint": "https://<some_subscription>.openai.azure.com/",
    "openai_api_version": "2023-07-01-preview",
    "openai_api_key": get_credential("OPENAI_API_KEY"),
    "deployment_name": "gpt-4-32k-2013june",
    "temperature": 0,
}

llm = AzureChatOpenAI(**AZURE_GPT_KWARGS)
llm_chain = LLMChain(llm=llm, prompt=some_prompt)
sdc = StuffDocumentsChain(llm_chain=llm_chain, document_variable_name="alternate_name")

# Ignore that this configuration may not be appropriate
configs = sdc.dict()

new_sdc = _load_stuff_documents_chain(configs)