MRVQ Enables Flexible Vector Search with Reduced Memory Footprint

Matryoshka Residual Vector Quantization (MRVQ) allows a single index to support multiple embedding dimensions and bit rates by dropping residual stages or coordinates. It uses 17.8–22.0x less memory than separately trained QINCo2 indices with minimal nDCG@10 performance loss.