Skip to main content

Functions

All cuVS functions consume parenthesized relation-valued subqueries and return ordinary relations that can be joined, filtered, aggregated, exported, or used by a later GPU SQL operator in the same statement. Their exact input and lifecycle contracts differ, so choose the function by the work it does within one evaluated query rather than by a generic "vector operation" label.

These are computation calls, not persistent artifact-management APIs: the current integration has no cross-statement cuVS cache for indexes, fitted models, transforms, or result relations. See cuVS SQL API for the bounded-preflight and server-cache boundary.

  • Brute-force kNN performs exact, query-local nearest-neighbor search. Use it for filtered candidate generation and exact batch reranking; it does not create a persistent ANN index.

Clustering​

  • KMeans fits and assigns clusters in one SQL statement. Its cluster_id values are query-local labels, not stable topic or business identifiers.

Dimensionality Reduction​

  • PCA fits and transforms its evaluated input in one SQL statement. The emitted pc_0 through pc_<n_components - 1> fields depend on the literal component count; no reusable transform is returned.

Every function page links to Vector Inputs for the shared ID/vector contract and to GPU Function Catalog API for metadata validation.