Two Years Of Vector Search At Notion

Notion reduced repeated search-indexing work by distinguishing text changes from changes to a page's properties. Its search system divides pages into passages and turns each passage into a numerical representation of meaning. Previously, even a small edit caused every passage on the page to be processed again, according to its engineering retrospective.

The revised system stores separate fingerprints for passage text and metadata, such as permissions. It regenerates representations only for changed text; a permissions-only edit updates metadata instead. Notion reports a 70% reduction in data volume from these changes, reducing both embedding work and database writes.