agentic retrieval-augmented generation (agentic RAG)

Agentic RAG upgrades retrieval-augmented generation from a single fetch-then-answer step into an active loop driven by the model. Classic RAG retrieves once on the original question and hopes the passages contain the answer; agentic RAG instead lets the agent plan queries, read what comes back, judge whether it is sufficient, and decide to search again, reformulate, or consult a different source before answering. It is the difference between one library lookup and a researcher who follows leads.

Concretely the model treats retrieval as a tool inside a reasoning–acting loop: it decides when to query, writes the query, inspects the returned documents, and may decompose a multi-hop question into sub-queries whose answers chain together. Variants add a relevance grader that filters or triggers re-retrieval, query rewriting when results are poor, and routing across multiple indexes or tools. Memory of what has already been retrieved prevents redundant searches and lets the agent build up evidence over several turns.

Iteration buys robustness on hard, multi-hop, or ambiguous questions that a single retrieval cannot satisfy, at the cost of more latency and model calls per answer, and more places to go wrong — bad sub-queries, premature stopping, or looping. As with all retrieval, the answer is only as trustworthy as the sources, so grounding, citation, and a stop condition matter.