📄 中文摘要
在调试 RAG 流水线时，发现如果 RAG 特性不断返回看似合理但错误的答案，首先应检查检索环节，而不是立即调整提示。经过多次重写提示、增加约束、收紧措辞并要求模型更贴近提供的上下文，虽然答案听起来更好，但依然错误。最终的解决方案并不是更智能的提示，而是清理数据路径，包括移除过时文档、改变块边界、添加可用元数据以及检查实际返回的检索结果。这一经验强调了数据质量在 RAG 系统中的重要性。
RAG 首先是一个数据问题，而不是提示问题
出处: RAG Is a Data Problem Before It’s a Prompt Problem
发布: 2026年3月16日
📄 English Summary
RAG Is a Data Problem Before It’s a Prompt Problem
During the debugging of a RAG pipeline, it was observed that when the RAG feature consistently returns plausible yet incorrect answers, the retrieval aspect should be inspected before adjusting the prompt. After multiple rewrites of the prompt, adding constraints, tightening the wording, and instructing the model to stay closer to the provided context, the answers sounded better but remained incorrect. The ultimate fix was not a smarter prompt but rather cleaning the data path by removing stale documents, changing chunk boundaries, adding usable metadata, and checking what retrieval actually returned. This experience highlights the importance of data quality in RAG systems.