Top-k retrieval is the step in which the system selects only a certain number of the best passages from the candidates found and sends them to the model as context. The number tends to be surprisingly small – on the order of single digits to a few dozen passages per sub-query. Everything else, even if relevant, does not make it into the answer at all. This gives rise to the harsh logic of generative search: there is no second page of results to reach. Either your passage is among those selected, or you do not exist for that question. The practical consequence is that it is worth focusing on the precision of the answer to a specific question rather than broad coverage of a topic in one long text, and having several targeted passages on a single topic that can each be selected for different sub-queries.
See also: Reranking, Query fan-out, Content chunking.
