Payload filtering is metadata filtering in systems where the structured data attached to each vector is called its payload — typically a JSON object of attributes stored alongside the vector. Filtering on the payload lets a query return only vectors whose attached data meets specified conditions, combining similarity with exact constraints.
The payload can hold any structured information about an item: category, price, timestamp, tags, permissions, tenant ID, and more. A payload filter expresses conditions over these fields — price below a threshold, category in a set, date after some point — and the database returns only the similar vectors whose payloads satisfy them. This is how vector search is constrained to respect real business rules.
Efficient payload filtering depends on how the database indexes and applies these conditions. Building payload indexes and using techniques like bitmap filtering keeps even complex or high-cardinality conditions fast, avoiding the recall cliff that naive filtering can cause. Strong payload filtering is a defining strength of vector databases built around rich metadata, where the payload is treated as a first-class part of every record rather than an afterthought.