Query fan-out rewrites a single request into several targeted sub-queries, runs them concurrently against the index or the web, and fuses the returned passages. It widens recall for complex questions — and it is how AI search engines dispatch one user prompt across many searches before synthesizing an answer.
Worked example: expand one user query into several sub-queries (rephrasings, decompositions, per-facet searches), retrieve for each, and merge the results — so a broad or multi-part question is not answered from a single narrow search. Gotcha: it raises recall (you catch documents a single phrasing missed) at the cost of more retrieval calls and latency, and the merge/rerank step becomes critical or you flood the context with near-duplicates; fan out when the query is complex, not for every lookup.