Quantization

Quantization compresses vectors to reduce storage and increase search speed with minimal accuracy loss.

Product Quantization (PQ)

PQ divides vectors into subvectors and quantizes each subvector separately.

Create PQ Index

Copy
-- Create product quantization index
SELECT pq_create_index(
    'documents',
    'embedding',
    'doc_pq_idx',
    8,   -- number of subvectors
    256  -- codebook size per subvector
);

Search with PQ

Copy
-- Search using quantized index
SELECT id, content
FROM documents
ORDER BY embedding <-> embed_text('query text')
LIMIT 10;

Optimized Product Quantization (OPQ)

OPQ adds rotation before quantization for better compression.

Copy
-- Create OPQ index
SELECT opq_create_index(
    'documents',
    'embedding',
    'doc_opq_idx',
    8,   -- subvectors
    256  -- codebook size
);

Compression Ratio

PQ and OPQ can achieve 2x-32x compression:

• 2x: High accuracy, minimal compression
• 32x: Maximum compression, slight accuracy trade-off

Learn More

For detailed documentation on quantization techniques, choosing parameters, accuracy trade-offs, and hybrid quantization, visit:

Quantization Documentation

Related Topics

• Vector Types - Vector data types
• Indexing - Creating indexes
• Distance Metrics - Similarity measures