Reproduce “ColBERTSaR: Sparsified ColBERT Index via Product Quantization” 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

Eugene Yang, Andrew Yates, Dawn Lawrie, James Mayfield, Saron Samuel, Rohan Jha

Researchers at Johns Hopkins University developed ColBERTSaR, a method that sparsifies ColBERT-style indexes via product quantization by eliminating residual vectors. This approach reduced index sizes by 53% to 77% compared to existing 1-bit compressed PLAID indexes while largely maintaining retrieval effectiveness across various monolingual, cross-language, and multilingual tasks.

Key insights

  • Eliminating residual vectors in ColBERT's MaxSim approximation, when combined with query-aware or unsupervised anchor optimization, enables significant index size reduction with competitive retrieval effectiveness.
  • ColBERT with embedding quantization can be theoretically viewed as a learned sparse retrieval model, providing a bridge between dense and sparse retrieval paradigms.
  • Using in-batch document examples as pseudo-queries offers a robust unsupervised method for anchor optimization, removing the dependency on external training queries.

Summary by alphaXiv.

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