RuVector-Postgres

High-Performance PostgreSQL Vector Similarity Search Extension

A drop-in replacement for pgvector, built in Rust with SIMD-optimized distance calculations, advanced indexing algorithms, and quantization support for memory-efficient vector storage.

Features

• pgvector API Compatibility - 100% compatible SQL interface, seamless migration
• SIMD Acceleration - AVX-512, AVX2, and ARM NEON optimized distance calculations (2-10x faster)
• Multiple Index Types - HNSW and IVFFlat indexes for approximate nearest neighbor search
• Quantization Support - Scalar, product, and binary quantization (up to 32x memory reduction)
• Multiple Vector Types - Dense (ruvector), half-precision (halfvec), and sparse (sparsevec)
• Zero-Copy Operations - Direct memory access for minimal overhead
• Neon Compatible - Designed for serverless PostgreSQL environments

Comparison with pgvector

Performance Benchmarks

Single distance calculation (1536 dimensions):

Batch processing (10K vectors x 384 dimensions):

Quick Start

Installation

Option 1: Quick Install Script

# Auto-detects platform and installs dependencies
curl -sSL https://raw.githubusercontent.com/ruvnet/ruvector/main/crates/ruvector-postgres/install/quick-start.sh | bash

Option 2: Full Installation

# Clone repository
git clone https://github.com/ruvnet/ruvector.git
cd ruvector/crates/ruvector-postgres

# Install with auto-detection
./install/install.sh --build-from-source

# Or specify PostgreSQL version
./install/install.sh --build-from-source --pg-version 16

See install/install.sh for all options including --dry-run, --verbose, and platform-specific configurations.

Basic Usage

-- Create the extension
CREATE EXTENSION ruvector;

-- Create a table with vector column
CREATE TABLE documents (
    id SERIAL PRIMARY KEY,
    content TEXT,
    embedding ruvector(1536)  -- OpenAI ada-002 dimensions
);

-- Insert vectors
INSERT INTO documents (content, embedding) VALUES
    ('First document', '[0.1, 0.2, 0.3, ...]'),
    ('Second document', '[0.4, 0.5, 0.6, ...]');

-- Create an HNSW index for fast similarity search
CREATE INDEX ON documents USING ruhnsw (embedding ruvector_l2_ops);

-- Find similar documents
SELECT content, embedding <-> '[0.15, 0.25, 0.35, ...]'::ruvector AS distance
FROM documents
ORDER BY distance
LIMIT 10;

Vector Types

ruvector(n) - Dense Vector

Standard 32-bit floating point vector for maximum precision.

CREATE TABLE items (embedding ruvector(1536));
-- Storage: 8 + (4 × dimensions) bytes

halfvec(n) - Half-Precision Vector

16-bit floating point for 50% memory savings with minimal accuracy loss.

CREATE TABLE items (embedding halfvec(1536));
-- Storage: 8 + (2 × dimensions) bytes

sparsevec(n) - Sparse Vector

For high-dimensional sparse data (BM25, TF-IDF).

CREATE TABLE items (embedding sparsevec(50000));
-- Storage: 12 + (8 × non_zero_elements) bytes
INSERT INTO items VALUES ('{1:0.5, 100:0.8, 5000:0.3}/50000');

Distance Operators

Index Types

HNSW (Hierarchical Navigable Small World)

Best for high recall and fast queries.

CREATE INDEX ON items USING ruhnsw (embedding ruvector_l2_ops)
WITH (m = 16, ef_construction = 64);

-- Tune search quality
SET ruvector.ef_search = 100;

IVFFlat (Inverted File Flat)

Best for memory-constrained environments and large datasets.

CREATE INDEX ON items USING ruivfflat (embedding ruvector_l2_ops)
WITH (lists = 100);

-- Tune search quality
SET ruvector.ivfflat_probes = 10;

When to Use Each Index

Tutorials

Semantic Search with OpenAI Embeddings

-- Create table for documents
CREATE TABLE documents (
    id SERIAL PRIMARY KEY,
    title TEXT,
    content TEXT,
    embedding ruvector(1536)
);

-- Create index
CREATE INDEX ON documents USING ruhnsw (embedding ruvector_cosine_ops);

-- Search (after inserting embeddings from OpenAI API)
SELECT title, content, embedding <=> $query_embedding AS similarity
FROM documents
ORDER BY similarity
LIMIT 5;

Image Similarity with CLIP Embeddings

-- CLIP produces 512-dimensional vectors
CREATE TABLE images (
    id SERIAL PRIMARY KEY,
    filename TEXT,
    embedding ruvector(512)
);

CREATE INDEX ON images USING ruhnsw (embedding ruvector_l2_ops)
WITH (m = 32, ef_construction = 200);

-- Find similar images
SELECT filename, embedding <-> $query_embedding AS distance
FROM images
ORDER BY distance
LIMIT 10;

Memory-Efficient Large-Scale Search

-- Use half-precision for 50% memory savings
CREATE TABLE large_dataset (
    id SERIAL PRIMARY KEY,
    embedding halfvec(1536)
);

-- IVFFlat for memory efficiency
CREATE INDEX ON large_dataset USING ruivfflat (embedding ruvector_l2_ops)
WITH (lists = 1000);

-- Increase probes for better recall
SET ruvector.ivfflat_probes = 20;

Hybrid Search (Vector + Text)

SELECT
    content,
    embedding <-> $query_vector AS vector_score,
    ts_rank(to_tsvector(content), to_tsquery($search_terms)) AS text_score,
    (0.7 * (1.0 / (1.0 + embedding <-> $query_vector)) +
     0.3 * ts_rank(to_tsvector(content), to_tsquery($search_terms))) AS combined
FROM documents
WHERE to_tsvector(content) @@ to_tsquery($search_terms)
ORDER BY combined DESC
LIMIT 10;

Configuration

GUC Variables

-- HNSW search quality (higher = better recall, slower)
SET ruvector.ef_search = 100;

-- IVFFlat probes (higher = better recall, slower)
SET ruvector.ivfflat_probes = 10;

Performance Tuning

-- Enable parallel index builds
SET maintenance_work_mem = '8GB';
SET max_parallel_maintenance_workers = 8;

-- Enable parallel queries
SET max_parallel_workers_per_gather = 4;

Installation Options

The install.sh script supports:

Platform-specific setup scripts are available in install/scripts/:

• setup-debian.sh - Debian/Ubuntu
• setup-rhel.sh - RHEL/CentOS/Fedora
• setup-macos.sh - macOS (Homebrew)

Documentation

Building from Source

Prerequisites

• Rust 1.70+ (install via rustup)
• PostgreSQL 14-17 with development headers
• Build tools (gcc/clang, make)

Build Steps

cd crates/ruvector-postgres

# Install pgrx
cargo install cargo-pgrx --version "0.12.9" --locked

# Initialize pgrx for your PostgreSQL version
cargo pgrx init --pg16 $(which pg_config)

# Build and install
cargo pgrx install --release

Running Tests

# Rust tests
cargo test

# SQL integration tests
psql -f tests/ivfflat_am_test.sql

Requirements

• PostgreSQL 14, 15, 16, or 17
• x86_64 (with AVX2/AVX-512) or ARM64 (with NEON)
• Linux, macOS, or Windows (via WSL)

License

MIT License - See LICENSE in the repository root.

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

Support

• Documentation: docs/
• Issues: GitHub Issues
• Examples: examples/