For RAG systems that return wrong, vague or made-up answers. Describe your chunking, embeddings, vector store, retrieval settings and context prompt, ideally with a few bad answers, and the review traces each bad answer to the stage that caused it. You get the likely causes, a table of fixes with actual settings (chunk size, top-k, reranking, hybrid search), a rewritten context prompt that requires citations, and three questions to confirm the fix worked.