LangChain

1 - I can move from having an idea to creating a working website in just a few hours, instead of spending days writing code from scratch. 2 - I don't have to search for separate libraries for document handling, vector databases, or managing prompts. Everything is available in LangChain package which is ready to use. Just I have to import it and start using that. 3 - If I want to try a different model, switch the vector store, or change how information is retrieved, i can easily do it with the help of LangChain.

It was one of the first agent oriented tools on the scene and for that we are thankful. It sped up our agent development and we were able to meet our milestones. We also liked how one could easily swap out and try different vector stores.

- LangChain provides a single way to interact with many LLMs and services, avoiding codebase bloat while trying different models. - LangChain is not just a text text generator, but it helps to easily connect LLMs with external APIs and tools. - It can be used to simplify building autonomous agents and multi-step workflows. - LangChain lets developers build powerful applications with minimal coding as it abstracts away complex details like prompt templating and document loading.

1 - Every time when a new update comes, new changes are introduced that break things. Code that worked last month might not work anymore. I've had to rewrite the same project several times just to keep up with the latest versions, which is really annoying. 2 - For simple tasks like making an API call or formatting a prompt, I end up writing a lot of classes and layers of code. it feels like overkill when I could have just written 10 lines of plain Python code. 3 - When something breaks, it's hard to find out where the problem is because there are so many layers between my code and what's happening. The error messages are not always helpful.

Being one of the earliest tools in this space, in hindsight some design decisions were obviously not the right choice. Also, the devs kept pushing breaking changes as it was updated and changed abstractions and apis; causing grief for us and other users. Nowadays looking back we can see that LangChain was over abstracted and made some not-optimal design choices. This obviously leads to debugging problems, even when we set verbose=True an error with LangChain was impossible to debug. Overall I would say the three primary cons of LangChain are: 1) Weak or incorrect documentation (partly due to the constant breaking changes and it not being kept up to date). 2) Hard to make anything custom. If you need anything outside of workflows it is difficult to accommodate, even with custom agents. 3) Overabstraction. For a specific example, see how it handles prompt templates. This could have been easily handled via a simple function and a formatted string but not when doing it the LangChain way.

- LangChain has too many layers of abstraction which makes simple tasks complicated compared to using raw LLM APIs. - Troubleshooting in LangChain is time consuming if there is an API timeout because it is difficult to trace the error across the numerous nested layers of LangChain. - Maintaining a stable codebase is a challenge using LangChain as the framework encounters frequent breaking changes.