Vector search

NYXDB treats embeddings as first-class data: a vector(N) column stores them next to the source rows, an HNSW index makes nearest-neighbor search fast, and an ANN query is ordinary SQL.

The vector(N) type

CREATE TABLE docs (
  id        UInt64,
  embedding vector(768),
  PRIMARY KEY (id)
);

vector(N) is a fixed-dimension dense numeric vector — N is 1–65535, element type float32 (default) or float64. See data types.

HNSW indexes

Declare an HNSW index on a vector column. A bare INDEX … TYPE hnsw defaults to the cosine metric. The index is built per part at flush (vendored usearch) and rebuilt on compaction, like every other secondary index.

-- inline on a standalone table
INDEX ix emb TYPE hnsw

On attribute tables, declare INDEX hnsw directly on the attribute; the index is built on the projection:

CREATE TABLE emb_t (
  id UInt64,
  ATTRIBUTE (emb vector(8) INDEX hnsw)
);

ANN search → VectorTopK

An ORDER BY <distance>(col, $q) LIMIT k query rewrites to the VectorTopK operator over the HNSW index (the AnnRewriteRule). The query vector is a '[...]' literal:

SELECT id, cosine_distance(emb, '[1,0,0,0,0,0,0,0]')
FROM emb_t
ORDER BY cosine_distance(emb, '[1,0,0,0,0,0,0,0]')
LIMIT 3;

EXPLAIN shows the rewrite:

EXPLAIN SELECT id, cosine_distance(emb, '[1,0,0,0,0,0,0,0]')
FROM emb_t
ORDER BY cosine_distance(emb, '[1,0,0,0,0,0,0,0]')
LIMIT 3;
-- plan contains: VectorTopK  column: emb  metric: cosine

For a default-latest read over an attribute table, the latest-projection rewrite runs first, then the ANN rewrite fires over the projection's HNSW index — so a KNN over current per-entity state is a VectorTopK, not a full scan.

Distance functions

cosine_distance, cosine_similarity, l2_distance, inner_product, dot_product, negative_inner_product. See Functions → vector / distance.