What if your HNSW index stored 3-bit embeddings instead of float32? [R]

A research paper (arXiv:2601.11557) proposes replacing the dominant "HNSW + float32 + cosine similarity" vector database stack with an information-theoretic alternative that uses Maximally Informative Binarization (MIB) to convert high-precision float32 vectors into compact quantized binary codes, efficient bitwise distance metrics (EDM) in place of cosine similarity, and an Information-Theoretic Scoring (ITS) mechanism instead of HNSW graph traversal. Rather than approximate nearest-neighbor graph search, the proposed architecture performs exhaustive scanning over these compact binary representations, enabling deterministic retrieval and eliminating accuracy degradation while dramatically reducing memory and computational costs. The core argument is that all three components of the conventional stack must be replaced simultaneously, as partial substitutions (e.g., storage disaggregation or lossy quantization alone) inevitably sacrifice either performance or accuracy.

[[p-pca-before-truncation-makes-non-matryoshka-embeddings-comp|[P] PCA before truncation makes non-Matryoshka embeddings compressible: results on BGE-M3 [P]]]
Anyone have an S3-compatible store that actually saturates H100s without the AWS egress tax? [R]
[[d-60-matmul-performance-bug-in-cublas-on-rtx-5090-d|[D] 60% MatMul Performance Bug in cuBLAS on RTX 5090 [D]]]
AI Systems Performance Engineering by Chris Fregly - is it worth it? [D]

Source: research

Related