Last updated: 2026-05-11
Agentic RAG Explained
Agentic RAG lets an agent plan and refine retrieval instead of accepting one static context bundle.
Definition
Agentic RAG combines retrieval, reasoning, and control flow so an agent can ask follow-up retrieval questions before answering or acting.
Why it matters
Many RAG failures are retrieval failures. Letting the agent inspect missing evidence can improve answers, but it also adds cost and latency.
Problems it solves
- Multi-hop knowledge questions
- Uncertain first-pass retrieval
- Source-aware research workflows
Common misconceptions
- Agentic RAG is not always better than simple RAG.
- More retrieval steps can increase hallucination if evidence is noisy.
- Evaluation must include retrieval metrics, not just final answer scores.
Minimal example
Run a first retrieval, ask the agent whether evidence is sufficient, then allow one bounded follow-up retrieval with a trace.
Next step: Start with retrieval evals, then add an agent loop where the failure cases justify it.