Product quantization

Compress cached vectors up to 32× with asymmetric distance.

S3 Vectors stores float32 and manages its own layout, so PQ does not change what it stores. PQ compresses the vectors dynavec caches — the in-memory hot tier and local candidate caches — turning a dim × 4 byte vector into m bytes.

from dynavec import ProductQuantizer

pq = ProductQuantizer(m=96, nbits=8).fit(training_vectors)   # 768-d -> 96 bytes (32x)
codes = pq.encode(vectors)          # uint8 codes
dists = pq.asymmetric_distances(query, codes)   # ADC, fast at scale
print(pq.reconstruction_error(vectors))

# Persist and reload codebooks across restarts
pq.save("pq_model.npz")
loaded_pq = ProductQuantizer.load("pq_model.npz")