Vector Types

In the previous sections Indexing and Search, we encountered the vector type. This section will explore additional vector types and their applications in the re

In the previous sections Indexing and Search, we encountered the vector type. This section will explore additional vector types and their applications in the realm of vector search.

Half-Precision Vector
8-Bit Integer Vector since v0.3.0
Sparse Vector since v0.3.0
Binary Vector since v0.3.0

Low-precision Indexing

since v0.3.0

To improve the speed of queries on the vector column, consider creating a low-precision index using vecf16, veci8, or bvector types instead of a vector type.

CREATE TABLE t (val vector(3));
INSERT INTO t (val) SELECT ARRAY[random(), random(), random()]::real[] FROM generate_series(1, 1000);

-- Create bvector index at vector column
-- Warning: Float32 vectors are truncated to 
--          integers of either 0 or 1 during index construction
CREATE INDEX lp_bin ON t USING vectors ((binarize(val)::bvector(3)) bvector_l2_ops);

-- Create vecf16 index at vector column
CREATE INDEX lp_f16 ON t USING vectors ((val::vecf16(3)) vecf16_l2_ops);

-- Create veci8 index at vector column
-- Warning: Float32 vectors are truncated to 
--          integers ranging from -128 to 127 during index construction
CREATE INDEX lp_i8 ON t USING vectors ((val::veci8(3)) veci8_l2_ops);

Shortening Embedding

since v0.3.0

Certain embedding models like OpenAI text-embedding-3-small support shortening embeddings, which involves removing some numbers from the end of the sequence without losing the embedding’s concept-representing properties.

In some cases, it may be necessary to change the embedding dimension after generation. The pgvecto.rs offers an approach to query by shortened embeddings.

CREATE TABLE t (val vector(3));
INSERT INTO t (val) SELECT ARRAY[random(), random(), random()]::real[] FROM generate_series(1, 1000);
CREATE INDEX ON t USING vectors (val vector_l2_ops);

-- Measure distance using only the first two dimensions
SELECT * from t ORDER BY (val[:2])::vector(2) <-> '[3, 2]'::vector;

Shortened embeddings are supported by all data types in pgvecto.rs since version v0.3.0.

Casts

This diagram shows the conversion between different data types, where the types connected by arrows can be cast to each other:

Among them, ARRAY is a native type of postgreSQL, others are types defined by pgvecto.rs.

Here are some examples of vector type casting:

-- OK: String ⇒ vector
SELECT '{1,2,3}'::vector;

-- OK: vector ⇒ ARRAY
SELECT '{1,2,3}'::vector::real[];

-- OK: ARRAY ⇒ vector ⇒ bvector
SELECT ARRAY[1.0 ,0 ,0]::real[]::vector::bvector;

-- ERROR: cannot cast type vecf16 to veci8
SELECT '[3, 2]'::vecf16::veci8;

-- OK: vecf16 ⇒ vector ⇒ veci8
SELECT '[3, 2]'::vecf16::vector::veci8;

Half-Precision Vector

Stored as a half-precision number format, vecf16 takes advantage of the 16-bit float, which uses half the memory and bandwidth compared to vector.

8-Bit Integer Vector

Int8 is a data type used to represent signed integers. It stands for “integer 8-bit” and typically uses 8 bits of memory to store a value. This allows it to rep

Sparse Vector

Unlike dense vectors, sparse vectors are very high-dimensional but contain few non-zero values.

Binary Vector

The bvector type is a binary vector type in pgvecto.rs. It represents a binary vector, which is a vector where each component can take on two possible values, t