Sebastian Bruch, Cosimo Rulli, and Rossano Venturini present SEiSMiC, a novel approximate retrieval algorithm for learned sparse representations. The algorithm organizes inverted lists into geometrically cohesive blocks, each equipped with a summary vector. During query processing, it quickly determines if a block must be evaluated using the summaries. SEiSMiC achieves sub-millisecond per-query latency on various sparse embeddings of the Ms Marco dataset while maintaining high recall. It outperforms state-of-the-art inverted index-based solutions and the winning submissions to the BigANN Challenge by a significant margin. The algorithm leverages the concentration of importance property of learned sparse embeddings, where a small subset of coordinates holds most of the importance. This allows for efficient pruning and skipping of blocks during retrieval. SEiSMiC uses a forward index for inner product computation and an inverted index to pinpoint the subset of documents that must be evaluated. The algorithm's design enables efficient retrieval over learned sparse representations, with significant gains in query latency and performance. The results show that SEiSMiC is one to two orders of magnitude faster than existing methods, with a notable improvement in accuracy and efficiency. The algorithm's effectiveness is demonstrated through extensive experiments on public datasets, including the Ms Marco and Natural Questions datasets. The results indicate that SEiSMiC outperforms other retrieval methods in terms of accuracy, latency, and space usage. The algorithm's design is based on the concentration of importance property, which allows for efficient pruning and skipping of blocks during retrieval. The algorithm's performance is further enhanced by the use of quantization and pruning techniques to reduce the size of summaries. The results show that SEiSMiC achieves high accuracy with minimal latency, making it a promising solution for efficient retrieval over learned sparse representations.Sebastian Bruch, Cosimo Rulli, and Rossano Venturini present SEiSMiC, a novel approximate retrieval algorithm for learned sparse representations. The algorithm organizes inverted lists into geometrically cohesive blocks, each equipped with a summary vector. During query processing, it quickly determines if a block must be evaluated using the summaries. SEiSMiC achieves sub-millisecond per-query latency on various sparse embeddings of the Ms Marco dataset while maintaining high recall. It outperforms state-of-the-art inverted index-based solutions and the winning submissions to the BigANN Challenge by a significant margin. The algorithm leverages the concentration of importance property of learned sparse embeddings, where a small subset of coordinates holds most of the importance. This allows for efficient pruning and skipping of blocks during retrieval. SEiSMiC uses a forward index for inner product computation and an inverted index to pinpoint the subset of documents that must be evaluated. The algorithm's design enables efficient retrieval over learned sparse representations, with significant gains in query latency and performance. The results show that SEiSMiC is one to two orders of magnitude faster than existing methods, with a notable improvement in accuracy and efficiency. The algorithm's effectiveness is demonstrated through extensive experiments on public datasets, including the Ms Marco and Natural Questions datasets. The results indicate that SEiSMiC outperforms other retrieval methods in terms of accuracy, latency, and space usage. The algorithm's design is based on the concentration of importance property, which allows for efficient pruning and skipping of blocks during retrieval. The algorithm's performance is further enhanced by the use of quantization and pruning techniques to reduce the size of summaries. The results show that SEiSMiC achieves high accuracy with minimal latency, making it a promising solution for efficient retrieval over learned sparse representations.