Yeah I've got a few thousand lines of langchain code now for a data cleaning pipeline... I've been fighting it every step of the way. Trying to replace sections of the pipeline to use a local LLM instead of OpenAI has had me have to replace the templates entirely, the chat based templates won't allow me to assign the proper user/assistant names, so the performance for the local LLM is terrible (stupid). They have zero actual composibility when you look at it slightly differently than they expect.
It's a useless abstraction for every single purpose I've actually tried it for.
Will be extricating it from my code base as soon as I find something else that works any better.
I have an attempt in the same domain, would love feedback
We think - Generative AI is config management. We model it on top of config management grammar that is proven to work at K8s/Terraform scale - jsonnet.
Prompts live outside the code. We didnt invent a new markup - we used jsonnet which is used in large scale kubernetes and has a grammar that has been well tested for config mgmt.
We thought the same thing with https://neum.ai - config management/infra as code for your LLM app, using plain JSON, though we are abstracting even more what LangChain does. Not suitable for the audience here given the level of customizations some folks want.
It's a useless abstraction for every single purpose I've actually tried it for.
Will be extricating it from my code base as soon as I find something else that works any better.