Definition
Agentic RAG
RAG in which an AI agent decides how to search, for example rewriting the query, choosing sources or searching again, instead of running a single fixed retrieval step.
Classic RAG retrieves once and answers. Agentic RAG puts a reasoning loop around retrieval: the agent can break a question into sub-questions, pick which index or tool to query, judge whether the results are sufficient and search again if they are not.
This helps with multi-part questions, questions that span several sources, and cases where the first query is a poor match for how documents are written. The trade-off is more model calls, so higher latency and cost, plus more behavior to evaluate.
A practical approach is to start with strong single-step retrieval and add agentic steps only where evaluation shows they improve answers.
Related services
Related terms