Vector index

pkg/sql/vecindex/ adds an approximate-nearest-neighbor (ANN) index for the VECTOR SQL type, integrated with the optimizer and execution engines.

Surface

CREATE TABLE docs (id UUID PRIMARY KEY, embedding VECTOR(1536));
CREATE VECTOR INDEX docs_emb_idx ON docs (embedding);

SELECT id FROM docs ORDER BY embedding <-> '[…]'::VECTOR(1536) LIMIT 10;

The optimizer recognizes the <->/<#>/<=> operators and rewrites the query to use the vector index when one is available.

Implementation

• pkg/util/vector/ — the data type (single-precision floats; cosine, L2, dot product distance).
• pkg/sql/vecindex/ — the index. Internally a multi-level k-means tree that partitions the vector space; each partition becomes its own KV row with a small SST-friendly layout.
• pkg/sql/vector_search.go — query-time scanner that walks the tree.
• pkg/sql/vecindex/cspann/ — control-plane code for re-balancing partitions.
• pkg/cmd/vecbench/ — benchmark harness (recall and latency sweeps).

The index is online: writes to the base table are journaled and asynchronously absorbed into the partitioning tree. Reads tolerate slight stale-ness in exchange for low write amplification.

Configuration

A few cluster settings under sql.vector_index.* control the partitioning fan-out, the journal size, and the recall target.

Related pages

• SQL — optimizer and execution integration.
• features/workload — vecindex workload exercises the index.