My "simple" RAG chatbot took 6 weeks to stop lying about refrigerator manuals
Everyone in the AI Slack groups kept saying you can stand up a working RAG pipeline over a weekend with a handful of API calls, so I figured two weeks max for a small client project. Six weeks later, after 40 test questions and a pile of answers that mixed up defrost timers from three different fridge models, I finally accepted the real problem wasn't the model or the vector store. It was my chunking. Splitting a wiring diagram from its own footnote is basically teaching the thing to guess, and no amount of prompt tweaking fixes bad chunks. A guy named Marcus on a Discord server asked me one question about chunk overlap and I had the whole thing working that afternoon. How are you all handling tables and diagrams in your documents, or do you just avoid PDFs altogether?