Seismic: Efficient and Effective Retrieval over Learned Sparse Representation
- Day - Time: 08 April 2026, h.11:00
- Place: Area della Ricerca CNR di Pisa - Room: C-29
Speakers
Referent
Abstract
Learned sparse embeddings offer a compelling mix of effectiveness,interpretability, and compatibility with classical IR systems, and are nowadaysa highly effective alternative to standard dense embedding. In this talk, wepresent Seismic, a state-of-the-art approach that rethinks the inverted indexby organizing posting lists into coherent blocks and using compact summaries toaggressively prune the search space. Grounded in a strong theoreticalinsight—the concentration of importance in sparse representations—Seismicenables efficient approximation of inner products with minimal loss inaccuracy. The result is sub-millisecond query latency at scale, supportingretrieval over collections of hundreds of millions of vectors and consistentlyoutperforming both optimized inverted indexes and graph-based ANN methods.