Does LangChain depend on OpenAI API too much

[bennylam]

I found LangChain is a very useful and convenient AI tool and framework for LLM application development. A rich set of API document, good sample codes and tutorials are sugar spots in attracting developers.

However, after trying many samples codes in the tutorials, I found most of these sample codes involved "import OpenAI..." API calls, which provided many key supplementary NLP functions like similarity search, embedding and keywords extraction...etc.

There are several concerns for such approach:

1. Since OpenAI APIs are mostly closed sourced now, sample code and tutorial will be broken if OpenAI change their API calls or behaviours.
2. Such tutorial and sample codes are useless for developers, who are not able or not prefer to apply for OpenAI API key due to various reasons e.g. OpenAI stringent approval process and limitations on free usage.
3. For LLM development for applications to be deployed on local server or internal/private network environment, the mandatory requirements of accessing the Internet and calling an unstable (sometimes) external API providers are not preferred or prohibited in many situations.

Since the main purpose of LangChain sample code and tutorial is to promote the functions and capabilities of LangChain and grow a wider user community and eco-system, I suggest more sample code and tutorial should be made without relying on OpenAI API calls only or a standard good practice of replacing OpenAI API call with at least one alternative (opensource is preferred) should be provided.

If there are already some good LangChain tutorials without relying on OpenAI API calls, please let me know.

Thanks
Benny

[GenerativeAI4Finance]

What I understand is
Either you are trying to setup a local LLM instance
Or
Want to use LLMs other than OpenAI

For latter, there are enough examples like for Hugging Face, Cohere. You may also look at AWS Sagemaker for more information.
You are right, for former, there is very less documentation.

We are also trying to use langchain with a local setup. And it seems, it's non-trivial task (If not too complex) to integrate langchain with local LLM & embeddings.

What is needed is to allow usage of Open source embedding models. Which should solve a lot of problems.
I am still in the process of doing it, but may be enough pointers are helpful to certain extent for you to try a few things.

[aronweiler]

I don't think that's what they're saying- I think it's a valid critique of the documentation that overly uses OpenAI examples.

The examples aside, I have found working with open-source LLMs and other non-OpenAI interfaces to be fairly easy, but I can see why it would present a challenge for anyone that isn't already a seasoned application developer.

I would second @bennylam and call for more documentation around other APIs besides OpenAI.

When I have a free moment, I will start looking at contributing some documentation related to this where I can.