I've been working on a simple Q&A bot for our internal knowledge base, using LangChain's `ConversationalRetrievalChain`. My goal was just to have it remember the last few exchanges for follow-up questions, a pretty standard need.
Following the official docs and tutorials, I ended down a rabbit hole of `ChatMessageHistory`, `ConversationBufferMemory`, and `ConversationSummaryBufferMemory`. The examples involve multiple classes, custom callbacks, and what feels like a lot of boilerplate just to maintain a simple list of previous messages. It seems like there's a focus on highly scalable, production-ready patterns even in the introductory material.
Has anyone else felt this way? I'm curious about two things:
* Is there a simpler, more straightforward pattern for basic chat history that I might have missed?
* For those who started with a complex setup, did you find the complexity paid off later, or did you eventually refactor to something simpler?
From my background in marketing automation, I'm thinking about it like lead scoring: sometimes a simple point-based system is more maintainable than a full Bayesian model, even if the latter is more "powerful." I'm wondering if the same principle applies here.