Langchain, with its kitchen-sink approach is great for newbies playing around with language models, to get a feel for the different things possible. The uniform interfaces make it very easy to snap together pieces and try things out (iff you don’t care to understand what’s happening under the hood).
Once you know exactly what you need (which might be 10% of Langchain’s capabilities), it might make sense to avoid all the cruft and build your own simple wrappers suited to the task at hand — especially if you need a robust and transparent/debuggable system. Langchain has a little too much indirection, and clumsy abstractions, to be fit for this purpose.
Any examples? How can I see the exact interactions behind the scenes that it's performing in order to DIY? Setting verbose to true doesn't seem to show everything
Sort of. I’ve found it’s value , currently, is not in its intended purpose to build “chains” of tools. The moment you have to do something non-standard you have to build from scratch which is more difficult with all of the abstraction involved.
Where I do find it useful is the many tools it saves me from having to build from scratch. For example I use a page scraper and an embedding service with retries are two things I use it for with the bot I built for my companies slack and discord. In theory I could see using it more in my project eventually. https://github.com/ShelbyJenkins/shelby-as-a-service
Yes, the map-reduce chains for summarisation, to work around context size limits of LLM. It works while there is a lot more debugging to do why is it much slower than expected (but it is fortunately fast enough for the task).
Langchain has a lot of surface area and it is still on 0.0.x, so also much advised to pin the versions hard.
We do, but mostly as a fairly nice and fairly well-documented interface for plugging together tools like vector DBs and memory with LLMs.
My feeling is that with the agentic/AutoGPT hype fading and OpenAI adding the functions API there's a lot less value in LangChain's abstractions, at least for production use cases. They're still cool for hacking/toys.
Some of the abstractions seem a bit forced, but it seems nice to be abstracted away from how you'd handle vector storage or the LLM that get used (maybe someone launches something that does better than GPT4 on some sets of tasks and a few line change to swap over would be nice).
I think that’s the wrong goal.
Langchain, with its kitchen-sink approach is great for newbies playing around with language models, to get a feel for the different things possible. The uniform interfaces make it very easy to snap together pieces and try things out (iff you don’t care to understand what’s happening under the hood).
Once you know exactly what you need (which might be 10% of Langchain’s capabilities), it might make sense to avoid all the cruft and build your own simple wrappers suited to the task at hand — especially if you need a robust and transparent/debuggable system. Langchain has a little too much indirection, and clumsy abstractions, to be fit for this purpose.