Resident product-quantization verification screen

Develop and evaluate a resident product-quantization screen that precedes exact verification in the misi metric inverted sample index, with the purpose of reducing random verification reads and potentially improving memory-limited serving.

Background

The misi query pipeline ranks candidates using shared-neighbor evidence and then performs exact distance verification on the top C candidates. In memory-constrained serving, this verification step causes substantial random I/O because database vectors remain disk-resident. The paper identifies resident product-quantization codes as a possible inexpensive screening layer before exact distance computation, but reports no implementation or measurements for it.