Letting the model search several times, look at what came back, and search again with a better query, instead of doing one search and hoping. It handles questions that need information from several places, and it costs more and takes longer, so it suits research-style tasks rather than quick lookups.
Technical view
An iterative loop where the model issues queries, evaluates retrieved evidence for sufficiency, and reformulates or drills down until it can answer or hits a stopping condition. Suits multi-hop and comparative questions that single-shot retrieval cannot serve. Requires iteration and token budget caps, and evaluation must assess the trajectory, since a correct final answer can follow retrieval that missed the relevant evidence entirely.
Related
Part of AI Terminology Index, section: Retrieval and Knowledge.