Product quantization (PQ) compresses dense vectors by splitting each vector into subvectors and replacing each subvector with the index of its nearest entry in a learned codebook. This lossy representation reduces memory substantially while preserving most retrieval quality. section 6.4.2
PQ solves a production constraint rather than replacing an index. Search strategy and vector storage are separate decisions: IVF or HNSW narrows the candidates, while PQ shrinks the stored vectors. IVF-PQ combines clustering with compression so large indexes fit in memory. section 2.2.2
Use PQ when the deployment trade-off favors a smaller index:
- Expect roughly 8–16× lower memory use in large-scale FAISS indexes, with search quality retained well enough for billion-scale deployments. section 2.2.2
- Accept lossy compression: PQ introduces a small retrieval-quality drop in exchange for a substantially smaller footprint. section 6.4.2
- Treat compression as part of deployment planning when a Wikipedia-scale embedding index would otherwise exceed single-GPU capacity. section 6.4.2 section 6.5
Compressed retrieval for a large RAG index
A Wikipedia-scale RAG system can apply PQ to reduce its embedding index by roughly 10×, allowing the retriever and reader to coexist on one high-end GPU with only a small downstream-performance loss. Pairing PQ with IVF avoids scanning every compressed vector. section 6.4.2 section 6.5 section 2.2.2
Table 2.2 FAISS Indexing Strategies
|
Index Type
|
Best For
|
Key Trade-off
|
Seminal Paper / Origin
|
|
Flat Index (Exact Search)
|
The highest accuracy is required
on smaller datasets (< 100K vectors). Critical for medical/legal
diagnosis.
|
Slow, linear scaling time, making
it unusable for big data.
|
Classical Baseline (Brute-force linear scan)
|
|
IVF (Inverted File)
|
Balance of speed and accuracy for
mid-sized datasets (100K–10M vectors). Enterprise document search.
|
Parameter Sensitivity. Accuracy
depends heavily on the number of clusters (nlist) and probes (nprobe).
|
"Video Google" (Sivic & Zisserman, 2003)
adapted inverted indices for visual vectors.
|
|
HNSW (Hierarchical Navigable Small World)
|
Ultra-fast search is the top
priority, especially for large datasets (> 1M vectors). Real-time
recommendations.
|
High Memory Cost. ~90-95% accuracy
with log-linear search time, but storing the graph structure consumes
significant RAM.
|
Malkov and Yashunin (2016) introduced the hierarchical
graph structure
|
|
Product Quantization (PQ)
|
Low Memory (mobile/embedded).
Often combined with IVF (IVF-PQ) for billion-scale.
|
Lossy. Up to 10x memory reduction
in exchange for a small drop in accuracy.
|
Jégou et al. (2011) defined subspace quantization for
nearest neighbor search.
|
Compare flat, IVF, and HNSW search strategies alongside the separate choice to store vectors fully or in compressed form. section 2.2.2



