← AI Terminology

Agentic RAG

Agentic RAG lets an agent decide iteratively what to retrieve, whether to re-query, which tools to use, and when enough evidence exists — rather than a single fixed retrieve-then-read pass.

It blends agents with retrieval for complex questions.
Why It Matters in AI
One-shot RAG fails on multi-hop and ambiguous queries. Agent loops that reformulate searches, compare sources, and fetch more context handle real research workflows — at higher latency and cost.
Key Points
Aspect Description
Loop Think → search → read → maybe search again → answer
Risk Runaway tool loops; need budgets
Tools Multiple indices, web search, SQL, calculators
Related ReAct, multi-hop reasoning, GraphRAG
Frameworks LangGraph, LlamaIndex agents, custom planners
Vs classic RAG Adaptive retrieval policy, not fixed k chunks
Simple Analogy
A researcher who keeps returning to the library catalogue with better queries until the paper pile actually answers the question — not one automatic shelf grab.
Common Usage Examples
  • Query rewrite then second retrieval
  • Agent chooses vector vs SQL tool
  • Stop when confidence or evidence quota met
  • Trace tool calls for debugging
Summary
In short: Agentic RAG makes retrieval an interactive agent loop — repeated, adaptive search instead of a single retrieve-then-generate pass.