Definition
What is agentic RAG?
Last updated
Definition
Agentic RAG is a pattern where an AI agent decides at runtime which retrieval queries to issue, iterates on results, and synthesizes across multiple retrieval rounds — instead of doing one fixed retrieval step.
Classical RAG retrieves once and generates once. Agentic RAG treats retrieval as a tool the agent can call multiple times: the agent reads the question, decides what to search for, evaluates the results, refines the query, and synthesizes across rounds. This handles complex multi-hop questions that one-shot RAG fails on. The cost is more LLM calls and latency; the benefit is materially higher accuracy on hard retrieval tasks.
When agentic RAG wins
- Multi-hop questions that need information from multiple sources
- Questions where the right query isn’t obvious from the user’s phrasing
- Tasks that need fact-checking against retrieved sources
When one-shot RAG is enough
Single-fact lookups, FAQ-style questions, anything where the relevant document is one keyword away. Agentic RAG’s overhead isn’t justified.