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…