Beyond the "extreme popularity of LangChain is warping the entire AI ecosystem around the workflows to the point of harming it", hasn't it recently become an attractor for substantial amount of investment money? I'm not saying you should be an ass about it, but the ecosystem will keep getting warped further if knowledgeable people won't speak up, and LangChain doesn't seem to be a random small open source project anymore.
EDIT:
From https://news.ycombinator.com/item?id=36647616:
> Remember, this is the project that raised ~$30m from Benchmark and Sequoia.
I think we're past the point of good-faith, small, volunteer-made open source project. If it can take $30M, it can stand some informed criticism too.
We’ve received a lot of commentary on our unwillingness to use it, and I don’t blame you for being hesitant. I don’t want to be the open-source project that says it’s not good when it’s not suitable for our uses.
I agree with your restraint, this feels like it might be more productive in another format. Ultimately this either needs to be broached with the maintainers or an alternative should be started.
(A while ago) I tried using LangChain and shortly gave up after not finding any way whatsoever to actually debug what’s going on under the hood (eg. see the actual prompts, LLM queries). It’s pretty ridiculous that this isn’t basic functionality, or at least it isn’t very discoverable.
I cannot imagine spending extended time with a framework without knowing what the internals are doing. I do realize this isn’t achievable on all levels with LLMs, but introducing more black boxes on top of existing ones isn’t solving any problems.
Please release it. People need to see these things BEFORE they get sucked into building an entire product around it.
A few bullet points:
- LangChain encourages tool lock-in for little developer benefit, as noted in the OP. There is no inherent advantage into using them, and some have suboptimal implementations.
- The current implementations of the ReAct workflow and prompt engineering are based on InstructGPT (text-davinci-003), and are extremely out of date compared to what you can do with ChatGPT/GPT-4.
- Debugging a LangChain error is near impossible, even with verbose=True.
- If you need anything outside the workflows in the documentation, it's extremely difficult to hack, even with Custom Agents.
- The documentation is missing a lot of relevant detail (e.g. the difference between Agent types) that you have to go diving into the codebase for.
- The extreme popularity of LangChain is warping the entire AI ecosystem around the workflows to the point of harming it. Recent releases by Hugging Face and OpenAI recontextualize themselves around LangChain's "it's just magic AI" to the point of hurting development and code clarity.
Part of the reason I'm hesitant to release said blog post is because I don't want to be that asshole who criticizes open source software that's operating in good faith.