Researchers have developed a method that makes document search faster while preserving retrieval quality in the tasks they tested.
Called Sparcent, the technique targets multi-vector retrieval, which represents text using multiple numerical vectors. These representations help match queries with relevant material, but searching them can require computation.
The study, published in Knowledge and Information Systems on October 12, focuses on a filtering step. Retrieval systems group vectors into clusters, then compare a query with cluster representatives, known as centroids, before performing detailed scoring.
Sparcent creates sparse representations of those centroids, allowing a different indexing approach to accelerate the matching process.
Integrated into two retrieval systems, it delivered speedups of up to 4.6 times over PLAID and 1.3 times over WARP. The researchers reported preserved retrieval effectiveness on both in-domain and out-of-domain tasks, with virtually no additional memory.
Those results concern the tested retrieval systems, rather than a universal acceleration of AI. They nevertheless point toward more efficient document search, an important component of applications that need to find relevant information before producing an answer. Production performance would require separate evaluation.