Overview
Vector quantization techniques compress high-precision embeddings into lower-precision representations, dramatically reducing storage and memory requirements.
Types
Scalar Quantization (SQ)
- Converts float32 to int8/uint8
- 4x compression ratio
- Minimal accuracy loss
- Fast and simple
Binary Quantization
- Reduces to 1-bit per dimension
- 32x compression
- Hamming distance for search
- Best for high-dimensional vectors
Product Quantization (PQ)
- Subvector clustering approach
- 8-64x compression
- More complex but effective
- Widely used in production
Benefits
- Cost Reduction: Lower storage costs
- Faster Search: Less data to process
- Higher Throughput: More vectors in memory
- Scalability: Fit larger datasets
Tradeoffs
- Some recall degradation
- Quantization overhead
- Parameter tuning needed
Support
- Qdrant: All types
- Milvus: SQ, PQ, BQ
- Weaviate: PQ, BQ
- Pinecone: SQ
Pricing
Techniques, not products.
