On this page
Dense vectors — supported#
CREATE TABLE docs (
id BIGINT PRIMARY KEY,
embedding VECTOR(FLOAT32, 3)
);VECTOR(element, dimensions) declares a dense vector column. The shorthand VECTOR(n) means VECTOR(FLOAT32, n).
| Element type | Stores | Notes |
|---|---|---|
FLOAT32 |
32-bit floats | Default; use unless you have a reason not to. |
FLOAT16 |
16-bit floats | Half the storage of FLOAT32; small precision trade-off. |
BFLOAT16 |
Brain float 16 | Same size as FLOAT16, larger range, less precision. |
INT8 |
Signed bytes (−128…127) | Quantized embeddings; smaller storage. |
UINT8 |
Unsigned bytes (0…255) | Quantized embeddings. |
INT4 |
4-bit signed nibbles | Smallest numeric form; accepted by some index paths via promotion — see Indexes. |
Integer elements must be whole numbers within range; non-finite values (NaN, Infinity) are rejected at insert.
Dimension limits#
- Dimensions: 1 to 65536.
- A single vector value is capped at 64 MiB.
- The dimension is fixed when the table is created; every row in that column must match it. A mismatch is a hard error, not a silent truncation:
-- ERROR: vector dimension mismatch
SELECT [1.0, 0.0] <-> [1.0, 0.0, 0.0];Literals and input forms#
Dense vectors are written as bracket literals of numbers:
INSERT INTO docs VALUES (4, 'delta', 'guide', [0.5, 0.5, 0.0]);
SELECT [1.0, 0.0, 0.0] <-> [0.0, 1.0, 0.0]; -- works on bare literals tooInteger-element columns accept integer literals ([1, 2, 3] into VECTOR(INT8, 3)). Values outside the element range are rejected.
NULL vectors are allowed (column without NOT NULL) and propagate as NULL distances — a NULL embedding never matches; it sorts per normal NULL ordering. Zero-norm vectors (e.g. [0, 0, 0]) yield NULL for cosine distance rather than a division-by-zero error.
Binary vectors — limited#
CREATE TABLE bdocs (
id BIGINT PRIMARY KEY,
embedding VECTOR(BINARY, 8)
);VECTOR(BINARY, n) stores bit strings of exactly n bits. Literals use the quoted bit-string form:
INSERT INTO bdocs VALUES (1, '10101010');
SELECT embedding <~> '11101010' AS hamming FROM bdocs;Use <~> (Hamming) or <%> (Jaccard) with binary columns — L2/cosine/dot are numeric-only. Binary ANN indexes fall back to exact computation in v0.1.0; results are correct, just not index-accelerated. See Limitations.
Sparse and multivector types — experimental#
SPARSE_VECTOR(element, dimensions) and MULTI_VECTOR(element, max_vectors, dimensions) parse in DDL, but there is no verified literal or insert path through SQL in v0.1.0. Do not build on them yet; they are tracked under Limitations.
Choosing#
- Text/image embeddings from standard models →
VECTOR(FLOAT32, <model dim>). - Tight storage with acceptable quantization error →
INT8/UINT8. - Bit-hash / fingerprint workloads →
VECTOR(BINARY, n)with Hamming distance.