🔊 “Want to go super deep into LLMs? Do an advanced project with Langchain”…
🤔 Said a post I saw a while back. Leaving aside what’s meant by advanced, & suggesting folks are better off grasping fundamentals like chunking, embedding, & transformer architectures, I’d ask “Why Langchain?”.
✨ OK, so Langchain is the go-to for LLM orchestration, boasting 62K GitHub stars. Great for PoCs with quick assembly of components into chains. Insightful for understanding different design patterns tailored for LLM tasks. The zero-shot ReAct framework for self-reasoning to a goal in 20 lines of code, powerful stuff!
âť—However, the more I use it, the more I find it very opinionated & difficult to wrangle into customized use cases. Also, the needless abstraction of simple model interfaces from likes of OpenAI. And about those abstractions. Being kept updated to really leverage state-of-the-art models?
🪲 Further, is it really production ready, with 2K open GitHub issues?
📜 Given this, some people advocate “roll your own” LLM orchestration in pure Python, while others say that’s just reinventing the wheel. Or controversial - use no-code first!
❓ Anyone found it’s easier, faster, & more reliable to develop LLM use cases without Langchain rather than with it? Would you go to production scale with it? Would love to hear your thoughts 🙏