Agentic RAG

Agentic RAG transforms retrieval from an automatic preparatory step into a tool managed by the agent in a ReAct loop: think → retrieve → assess sufficiency → refine query and retrieve again.

When to choose:

• simple questions with a clear need – classic RAG is faster and no worse;
• complex multi-step questions – agentic RAG is noticeably better: breaks down the question, iteratively searches, cross-checks.

Safety boundaries: retrieved documents are a channel for indirect prompt injection; the retrieved content should not directly trigger risky operations.

Related: RAG, ReAct loop, Prompt injection