Recommendations

Recommendations are a first-class feature on Flexible Vector Database. The pattern is two calls: record that the user viewed a document, then send a search that biases results toward what the user has been looking at.

Recording a view

Call POST /flexible-vector-databases/{id}/view-element with the document id the user interacted with. The platform returns a recommend token that identifies a vector of view signals the search engine can blend with the query embedding.

Send the token on every subsequent search that should be biased. The token is short-lived and single-user: do not share it across accounts and treat it like a session identifier.

Biased search

Add a recommend block to the search payload (see Searching). The block carries the token and a bias float in the [0, 1] range. A bias of 0 returns pure query results; a bias of 1 returns pure recommendation results; values in between blend the two.

The blending happens inside the search engine: results that match both the query and the user's view history are surfaced first, and the top_k ceiling is preserved.