Hey everyone! I've been lurking for a bit while learning the ropes of this whole LLM orchestration space, and I finally have a concrete project to talk about. I just finished a proof-of-concept for automatically routing inbound sales emails to the right internal team (like "enterprise," "SMB," or "technical support") based on the email content. My go-to stack is usually Python with some SQL for the data modeling side of things, so I naturally reached for LangChain to build the pipeline.
The good news is, it absolutely worked. I managed to string together a prompt template, an OpenAI call, and some output parsing to get a clean JSON with the predicted category and confidence. Getting something functional up and running was surprisingly fast thanks to the documentation and examples. However, I can't shake this feeling that it was a bit... heavy for what I was doing. My PoC is essentially one straightforward chain: text -> LLM -> decision. Bringing in all of LangChain felt like using a whole framework just to call an API and parse the result. I found myself wondering if I was using parts of the library "correctly" or if I was adding unnecessary abstraction layers.
I'm particularly cautious about locking into a tool early, especially with how fast this ecosystem moves. So I'd love to hear from others who've built similar lightweight applications. For those of you who have production experience:
* Did that "heaviness" feeling go away as you added more steps (like retrieval or tool use), or did it become more pronounced?
* At what point did you feel like LangChain's abstractions became truly necessary versus just using the raw OpenAI SDK and some helper functions?
* I'm drawn to open-source solutions—are there lighter alternatives you'd recommend for someone with a data engineering background who's comfortable with Python but new to the LLM app world?
My instinct, coming from data engineering, is to keep things as simple and transparent as possible, especially in the early stages. But I also don't want to reinvent the wheel if LangChain is the standard for a reason. Maybe I just need to build a more complex workflow to appreciate it? Any insights or shared experiences would be incredibly helpful as I decide whether to double down on LangChain for this project or switch gears.