Reproduce “Efficiency Optimizations for Superblock-based Sparse Retrieval” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Parker Carlson, Wentai Xie, Rohil Shah, Tao Yang

A new Lightweight Superblock Pruning (LSP) method enhances sparse retrieval efficiency and robustness by guaranteeing a fixed number of top-scoring superblocks and integrating specialized index compression. This approach achieved up to 8.1 times faster query processing compared to baseline methods while significantly reducing memory usage across various in-domain and zero-shot retrieval benchmarks.

Key insights

  • Guaranteed inclusion of a top-gamma set of superblocks effectively mitigates erroneous pruning, enabling more aggressive pruning without compromising the availability of top-k results.
  • Removing the average-score-bound check from superblock pruning logic simplifies the algorithm and reduces overhead, as its utility for maintaining relevance is diminished when top-superblock inclusion is in place.
  • Systematic index compression, combining 4-bit quantization for maximum weights and optimized inverted/forward indexes, drastically reduces memory usage and can decrease retrieval latency by improving data locality.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.