#Is langchain useless?

12 messages · Page 1 of 1 (latest)

sage gate
#

I have been using langchain in my project for the better part of 6 months, what I am using it for is not generating duplicate data, so I pull data from my vector database save it onto a document and use that to generate new questions using langchain. Would it be any different if I just add the data from the vector database directly onto the "system" part of the prompt in the openai api?

torn widget
bronze topaz
#

What are the options?

static seal
bronze topaz
#

ok

#

But Lang Chain is Python; what is wrong with that? I am interested in your analysis.

static seal
#

The reason is langchain does not have pythonic syntax. And the whole thing is just an abstraction over basic python code.

#

So either a developer who knows python basics can just write the code or he wastes time learning a new library which does what he already knows.

#

Actually you might be able to make better code than langchain.

#

Almost every person here writes his own RAG code or other pieces of code.

#

It doesn't make sense to learn extra things.

glass matrix
#

Also depends on the use case. Langchain works well with small scale projects. If you were to have a knowledge base formatted in some special XML or JSONL format or you need a custom vector DB with millions of entries in it, langchain falls apart. That's where building your own RAG system comes into play.

To build on top of Ian and ArRad, it also sucks to read and parse. The code itself is 5 levels of useless abstraction straight out of OOP hell and the documentation serves better to give examples for 1 use case rather than cover all the features (for instance, lot of sample code on using OpenAI but very little on using local models from Huggingface).

That said, you don't have to avoid langchain. It's nifty for the people who just want their LLMs to chat with their documents on their machines.