vector

Overview

Upstream and source archive are at pgvector 0.8.6; indexed PGDG RPM and DEB packages remain at 0.8.5 for PostgreSQL 14-18.

Version

Build

You can build the RPM / DEB packages for pgvector using pig build:

pig build pkg pgvector         # build RPM / DEB packages

Install

You can install pgvector directly. First, make sure the PGDG repository is added and enabled:

pig repo add pgdg -u          # Add PGDG repo and update cache

Install the extension using pig or apt/yum/dnf:

pig install pgvector;          # Install for current active PG version

pig ext install -y pgvector -v 18  # PG 18
pig ext install -y pgvector -v 17  # PG 17
pig ext install -y pgvector -v 16  # PG 16
pig ext install -y pgvector -v 15  # PG 15
pig ext install -y pgvector -v 14  # PG 14

dnf install -y pgvector_18       # PG 18
dnf install -y pgvector_17       # PG 17
dnf install -y pgvector_16       # PG 16
dnf install -y pgvector_15       # PG 15
dnf install -y pgvector_14       # PG 14

apt install -y postgresql-18-pgvector   # PG 18
apt install -y postgresql-17-pgvector   # PG 17
apt install -y postgresql-16-pgvector   # PG 16
apt install -y postgresql-15-pgvector   # PG 15
apt install -y postgresql-14-pgvector   # PG 14

Create Extension:

CREATE EXTENSION vector;

Usage

Sources:

• pgvector v0.8.6 README
• pgvector v0.8.6 CHANGELOG
• Changes from v0.8.5 to v0.8.6

pgvector provides vector similarity search inside PostgreSQL. The extension name is vector, while Pigsty packages it as pgvector. It supports exact search, approximate nearest-neighbor search with HNSW and IVFFlat indexes, and multiple vector representations for dense, half-precision, binary, and sparse embeddings.

Version 0.8.6 is a focused correctness release. It retains the 0.8.x HNSW iterative-scan and maintenance improvements documented in the current README.

Create and Query Vectors

CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE items (
  id bigserial PRIMARY KEY,
  embedding vector(3)
);

INSERT INTO items (embedding)
VALUES ('[1,2,3]'), ('[4,5,6]');

SELECT *
FROM items
ORDER BY embedding <-> '[3,1,2]'
LIMIT 5;

Common distance operators:

• <-> for L2 distance
• <#> for negative inner product
• <=> for cosine distance
• <+> for L1 distance
• <~> for Hamming distance on binary vectors
• <%> for Jaccard distance on binary vectors

Because PostgreSQL indexes scan in ascending order, <#> returns the negative inner product; multiply by -1 when displaying the actual inner product.

Vector Types

CREATE TABLE embeddings (
  id bigserial PRIMARY KEY,
  dense      vector(768),
  half_dense halfvec(768),
  binary_sig bit(1024),
  sparse     sparsevec(100000)
);

vector is the standard single-precision type. Use halfvec to reduce storage and memory pressure, bit for binary signatures, and sparsevec for high-dimensional sparse vectors.

Aggregates such as avg() and sum() can be used with vector columns:

SELECT avg(embedding) FROM items;

HNSW Indexes

HNSW gives strong speed/recall tradeoffs and does not require a training step.

CREATE INDEX items_embedding_hnsw
ON items USING hnsw (embedding vector_l2_ops);

SET hnsw.ef_search = 100;

SELECT *
FROM items
ORDER BY embedding <-> '[3,1,2]'
LIMIT 10;

Choose the operator class that matches the distance:

CREATE INDEX ON items USING hnsw (embedding vector_ip_ops);
CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
CREATE INDEX ON items USING hnsw (embedding vector_l1_ops);
CREATE INDEX ON embeddings USING hnsw (half_dense halfvec_l2_ops);
CREATE INDEX ON embeddings USING hnsw (sparse sparsevec_l2_ops);
CREATE INDEX ON embeddings USING hnsw (binary_sig bit_hamming_ops);

Useful tuning settings include hnsw.ef_search, hnsw.iterative_scan, hnsw.max_scan_tuples, and hnsw.scan_mem_multiplier.

IVFFlat Indexes

IVFFlat requires representative data before index creation because it trains cluster lists at build time.

CREATE INDEX items_embedding_ivfflat
ON items USING ivfflat (embedding vector_l2_ops)
WITH (lists = 100);

SET ivfflat.probes = 10;

SELECT *
FROM items
ORDER BY embedding <-> '[3,1,2]'
LIMIT 10;

Increase lists for larger tables and increase ivfflat.probes for higher recall. For filtered queries, test whether an exact btree filter, a partial vector index, or partitioning gives better plans.

Filtering and Hybrid Search

Normal PostgreSQL filters can be combined with vector ordering:

ALTER TABLE items ADD COLUMN tenant_id bigint;
CREATE INDEX ON items (tenant_id);

SELECT *
FROM items
WHERE tenant_id = 42
ORDER BY embedding <=> '[0.1,0.2,0.3]'
LIMIT 20;

For hybrid search, combine pgvector with PostgreSQL full text search, trigram search, or an external ranking expression:

CREATE TABLE docs (
  id bigint PRIMARY KEY,
  body text NOT NULL,
  text_tsv tsvector GENERATED ALWAYS AS
    (to_tsvector('english', body)) STORED,
  embedding vector(3)
);

SELECT id,
       ts_rank_cd(text_tsv, plainto_tsquery('database')) AS text_score,
       1 - (embedding <=> '[0.1,0.2,0.3]') AS vector_score
FROM docs
WHERE text_tsv @@ plainto_tsquery('database')
ORDER BY vector_score DESC
LIMIT 20;

Maintenance

VACUUM items;
REINDEX INDEX CONCURRENTLY items_embedding_hnsw;
ANALYZE items;

HNSW indexes can be large and expensive to build. Use maintenance_work_mem for builds, monitor build notices, and schedule REINDEX when index bloat or recall drift matters.

Caveats

• Version 0.8.6 fixes an IVFFlat build overflow on 32-bit systems, enforcement of the nonzero-element limit when casting an array to sparsevec, and memory growth during IVFFlat scans inside nested loops. It does not add a new SQL feature surface. Run ALTER EXTENSION vector UPDATE after installing new extension files when the database reports an older SQL version.
• Use the operator class that matches the query operator. A cosine index will not accelerate an L2 ORDER BY.
• Approximate indexes trade exact recall for speed. Validate recall with representative data and query filters.
• Build IVFFlat after loading data. If data distribution changes substantially, rebuild the index.
• Keep pgvector updated when using HNSW with heavy writes and vacuum activity; the 0.8.x line includes important HNSW maintenance fixes.