I will debug your rag chatbot and fix why it gives wrong answers

Your RAG chatbot worked in the demo. In production it hallucinates pricing, invents policies, or confidently answers with the wrong document. Everyone tells you to add reranking or rewrite the prompt that's step 4, and it's rarely where the bug actually is.

I debug RAG systems the way you'd debug any other production system: top-down, starting with what was actually retrieved, not what the model said. Most wrong answers trace back to retrieval the right chunk never reached the model not the LLM being "dumb." I separate the two before touching anything, because the fix for each is completely different and guessing wastes your time and money.

I work on production LLM pipelines professionally multi-model validation, schema-enforced outputs, retry-capped regeneration loops so I'm used to finding the actual failure mode instead of throwing fixes at a wall.

What I check: retrieval accuracy, chunking boundaries, context assembly, prompt truncation, and whether the model is misreading correct context vs. never seeing it at all.

Send me 5-10 of your worst wrong answers before you order I'll tell you honestly whether this is a quick fix or a deeper rebuild.