Vector Search

TidesDB supports approximate nearest-neighbor search through MariaDB’s built-in MHNSW vector index. The server handles graph construction and search, and TidesDB provides the storage for both the table data and the hidden MHNSW graph. The VECTOR type and its index need MariaDB 11.7 or later, since older servers have no VECTOR type.

CREATE TABLE embeddings (
  id INT NOT NULL PRIMARY KEY, title VARCHAR(200),
  v VECTOR(384) NOT NULL, VECTOR INDEX (v)
) ENGINE=TIDESDB;

INSERT INTO embeddings VALUES (1, 'cat picture', Vec_FromText('[0.1, 0.9, ...]'));

Searching

Vector search uses ORDER BY VEC_DISTANCE_EUCLIDEAN() or VEC_DISTANCE_COSINE() with a LIMIT. The MHNSW index returns approximate neighbors without scanning the whole table. MariaDB uses the index only when the distance function matches the index’s DISTANCE metric, so the embeddings table above, whose index uses the default euclidean metric, is searched with VEC_DISTANCE_EUCLIDEAN():

SELECT id, title, VEC_DISTANCE_EUCLIDEAN(v, Vec_FromText('[0.15, 0.85, ...]')) AS dist
FROM embeddings ORDER BY dist LIMIT 5;

A query using VEC_DISTANCE_COSINE() against that table is answered with a full scan. To search by cosine distance, create the index with DISTANCE='cosine', as in the docs table under Index options below, and query it with the matching function:

SELECT id, VEC_DISTANCE_COSINE(v, Vec_FromText('[0.15, 0.85, ...]')) AS dist
FROM docs ORDER BY dist LIMIT 5;

Index options

The MHNSW index accepts two optional parameters, both handled by the server:

CREATE TABLE docs (
  id INT PRIMARY KEY, v VECTOR(128) NOT NULL,
  VECTOR INDEX (v) M=12 DISTANCE='cosine'
) ENGINE=TIDESDB;

M is the number of neighbors per graph node, default 6, range 3 to 200. Higher values improve recall at the cost of slower inserts and more memory. DISTANCE selects the metric, euclidean (default) or cosine.

DML and limitations

All DML works on vector-indexed tables. INSERT adds the vector to the graph, DELETE removes it, and UPDATE on the vector column invalidates the old graph node and inserts a new one. The engine handles the interleaved record[0]/record[1] access pattern that the MHNSW maintenance uses for BLOB-backed vector data.

A partitioned table cannot carry a vector index, and MariaDB’s MHNSW implementation supports one vector index per table.