Let quantization shape the graph
QuIVer asks whether binary quantization can be the metric space of the graph index itself. A training-free two-bit code combines sign and magnitude bits so Vamana edge selection, diversity pruning, and query navigation all operate in quantized space.
Compact navigation, exact reranking
Queries are encoded as binary signatures and traverse candidates with bitwise operations. Full float32 vectors are read only at the end for reranking, separating frequently accessed signatures and adjacency lists from cold full-precision data.
No learned codebook or rotation matrix is required. Because quantization participates in construction and navigation, topology, computation, and memory layout can be designed around the compact representation together.
Evaluate systems together with data geometry
Experiments on twelve million-scale datasets show strong distribution dependence: contrastive embeddings in cosine space fit the topology best, some multimodal representations follow, and native Euclidean or unstructured data perform poorly.
The work makes the tradeoff between compression, throughput, and data compatibility explicit, and provides a systems foundation for later work on quantization theory.
READ AND CITE
Continue with the original paper.
The complete derivations, experimental setup, and results are available in the public manuscript.