什么时候不该用 RAGWhen Not to Use RAG
微型指南 · 约 7 分钟Guide · about 7 min · 引用核验 2026-11-03（仅引用链接与日期） · refs checked 2026-11-03 (reference links and dates only)
RAG 不是默认选项。它适合「知识会变、要出处」的场景，不适合另一些场景，硬上反而更差。RAG is not the default. It fits "knowledge changes, citations required"; it hurts other scenarios, and forcing it makes things worse.
知识量小且稳定：直接放进上下文更简单，检索层是多余复杂度。
要求精确算术或推理：检索帮不上，问题在模型本身。
文档质量差：垃圾进垃圾出，先修文档再谈检索。
延迟预算紧：检索加重排加生成，链路长，实时场景要掂量。
需要严格一致性：同一问题两次检索结果可能不同，合规场景要评估。
先试基线：不加检索直接问，效果够好就不用搭 RAG。
Small, stable knowledge: put it straight in the context; a retrieval layer is needless complexity.
Exact arithmetic or reasoning required: retrieval does not help; the limit is the model itself.
Poor document quality: garbage in, garbage out — fix the documents before talking about retrieval.
Tight latency budget: retrieval plus reranking plus generation is a long chain; weigh it for real-time use.
Strict consistency required: two runs of the same question may retrieve differently; assess this for compliance settings.
Try the baseline first: ask without retrieval; if it is good enough, do not build RAG.
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
只有一页稳定的产品说明时，先试把说明直接作为上下文。若需求只是按规则算总价，应考虑确定性的计算逻辑。是否增加检索层，取决于材料规模、更新、权限和出处要求，不由技术热度决定。For one stable product sheet, first try supplying it directly. For rule-based totals, consider deterministic calculation. Add retrieval based on volume, updates, permissions and citation needs, rather than novelty.
联网 / 检索增强Retrieval-Augmented Generation, RAG 上下文窗口Context Window 能力边界Capability Boundary
怎么向 AI 提问How to Ask AI · 怎么判断 AI 新闻是不是炒作How to Tell if AI News Is Hype · 怎么给文档做分块How to Chunk Documents for Retrieval · 怎么给检索结果重排How to Rerank Retrieval Results