Why it matters

PQ enables billion-scale ANN. Understanding shapes vector DB.

Advertisement

The architecture

Split d-dim vector into m subvectors.

Quantize each with k centroids.

Store bytes-per-subvector.

Query via lookup tables.

PQ structureSplit into m subvectorsblock subdivisionQuantize eachk centroids per blockLook-up distancetable-basedJegou et al. 2011 PQ; Faiss + Milvus + others; billion-scale enabler
PQ math.
Advertisement

How it works end to end

Jegou et al. 2011.

m = 8-32 typical.

k = 256 (1 byte).

Combined with IVF for scale.