If you have built anything with retrieval-augmented generation (RAG) in the last two years, you have lived its central frustration: You chop your documents into chunks, embed them, retrieve the top few that look similar to the question, and hand them to the model. For “What was our Q3 refund policy?