Your RAG system scores 0.89 on faithfulness. Users keep escalating. The answers are grounded — but they're answering the wrong question. The problem isn't the generation. It's what you handed the model.