Module cold

Source
Expand description

Cold storage tier with Product Quantization (PQ) compression.

Vectors are encoded into compact PQ codes for massive storage reduction. A codebook of centroids is trained via k-means on representative data, then each vector is encoded as a sequence of centroid indices.

§Compression

With M=8 subspaces and K=256 centroids per subspace:

  • Original D=768 vector: 768 × 4 bytes = 3,072 bytes
  • PQ code: 8 × 1 byte = 8 bytes
  • Compression ratio: 384×

§Asymmetric Distance Computation (ADC)

Query-to-code distance is computed without decoding: precompute query-to-centroid distances, then sum lookup table entries.

Structs§

ColdEntry 🔒
A single entry in cold storage.
ColdStore
Cold store using PQ-encoded vectors.
PqCodebook
Product Quantization codebook.

Functions§

find_closest_centroid 🔒