Self contained passages

Passages, not whole pages, are what gets matched. Google's passage ranking system identifies individual sections (passages) of a page to judge how relevant the page is to a search; announced in October 2020 as improving 7 percent of queries across all languages, it lets a highly relevant passage buried in a long page make that page relevant to a specific question. Retrieval-augmented systems work at the same grain: Dense Passage Retrieval and RAG split Wikipedia into disjoint 100-word passages (21M in RAG's index) and retrieve those. Google says its systems can understand multiple topics on one page and show the relevant piece to users, and that breaking content into tiny pieces is not required. Inference from passage-level retrieval, not a measured result: a passage that names its subject and states the answer in its own first sentences can be matched and quoted alone, while one that leans on pronouns or earlier context carries less of the query's meaning.