What is Agentic RAG?
Agentic RAG is retrieval-augmented generation where an agent actively decides what to retrieve, evaluates whether the results are good enough, and retrieves again — rather than doing a single fixed lookup before answering. It closes the biggest failure mode of classic RAG: answering confidently from the wrong or incomplete documents.
Classic RAG vs agentic RAG
Classic: query → embed → top-k search → stuff into prompt → answer
Agentic: query → plan → search → grade results → (rewrite query / search again
/ decompose into sub-questions) → synthesise → self-check → answer
What the agent adds
| Step | Purpose |
|---|---|
| Query rewriting | Turn a vague question into good search terms |
| Result grading | Discard low-relevance chunks before they poison the context |
| Decomposition | Break a multi-part question into separate retrievals |
| Multi-source routing | Pick vector search vs SQL vs web per sub-question |
| Self-verification | Check the draft answer is supported by the retrieved text |
The cost is latency and tokens — several model calls instead of one — so agentic RAG is used where accuracy matters more than speed: research, compliance, support over large knowledge bases.
END OF ANALYSIS