Agentic RAGAgentic RAG
Agentic RAG
让 agent 自己决定要不要检索、检索几次、怎么改写 query，而不是一次检索就生成。Let the agent decide whether to retrieve, how many times, and how to rewrite the query, instead of generating after one retrieval.
Agentic RAG 把检索当成 agent 的工具：它可以多轮改写查询、选择不同索引、判断资料是否足够，再决定是否继续。
与固定 RAG 链路相比，它更灵活，但步数、成本与不可控性也更高。
Agentic RAG treats retrieval as an agent tool: it can rewrite queries over multiple turns, choose different indexes, judge whether the material is sufficient and decide whether to continue.
Compared with a fixed RAG pipeline it is more flexible, but also less predictable and more expensive in steps.
复杂问题常常一次检索不够；理解这一层，才能判断「RAG 效果不好」到底是检索器的问题，还是流程设计的问题。Complex questions often need more than one retrieval. Understanding this layer tells you whether poor RAG results come from the retriever or from workflow design.