Sparcent Speeds Up Multi-Vector Search by Turning Centroids into Sparse Indices

Search engines face a stubborn trade-off at the heart of modern information retrieval: the models that find the best answers are also the models that cost the most to run. Multi-vector dense retrieval methods, which represent every passage as a large collection of contextualized token embeddings rather than a single compressed vector, currently deliver state-of-the-art ranking quality on benchmark tasks. Yet their computational appetite has kept them largely out of large-scale production search, where billions of documents must be scanned under strict latency and throughput budgets. A new study published in Knowledge and Information Systems by Erman Yafay and Ismail Sengor Altingovde of Middle East Technical University in Ankara proposes a way out of this bind, and its central idea is deceptively simple: use one family of neural retrieval techniques to accelerate another.

The method, called Sparcent, targets the most expensive stage of multi-vector dense retrieval, the stage in which a query must be matched against candidate documents before expensive fine-grained scoring begins. Systems such as PLAID and WARP, both of which are well-known multi-vector retrieval engines already heavily optimized for single-core CPU execution, organize document embeddings into clusters and use cluster centroids as a coarse filtering mechanism. The query is first compared against these centroids, and only clusters whose centroids look promising are expanded into full token-level scoring. The quality of the final ranking depends on this coarse stage being both fast and accurate: prune too aggressively and relevant documents slip through the net, prune too conservatively and the speed advantage evaporates.

Yafay and Altingovde observed that this query-centroid matching stage could be reframed as a problem that learned sparse retrieval, or LSR, already solves remarkably well. Learned sparse retrieval models, exemplified by systems such as SPLADE, produce high-dimensional but mostly zero-valued representations of text, where each dimension corresponds to a vocabulary term and the nonzero weights reflect learned, context-aware term importance. These sparse representations can be indexed with inverted indexes, the same data structure that has powered classical keyword search engines for decades, and traversed with mature dynamic pruning techniques such as WAND-style algorithms. In other words, the sparse retrieval ecosystem offers extremely fast, memory-efficient query evaluation machinery that dense retrieval pipelines have not traditionally been able to exploit.

To bridge the two worlds, Sparcent derives sparse centroid representations for the term-embedding clusters that underlie the multi-vector index. Each cluster of token embeddings is summarized not only by its dense centroid vector but also by a sparse profile that learned sparse retrieval systems can index and score. The authors then apply pruning to these sparse centroids, trimming away the least informative entries to produce compact centroid vectors that remain both efficient and effective when indexed via LSR methods. This pruning step matters because raw sparse expansions can be long, and an unpruned centroid representation would erode the very efficiency gains the approach is designed to deliver. The result is a hybrid index in which the coarse pruning stage of dense retrieval is executed by battle-tested inverted-index machinery rather than by brute-force vector comparisons.

The integration was carried out with two established multi-vector systems. PLAID, built around the ColBERT family of models, pioneered cluster-centric pruning for late-interaction retrieval, while WARP represents a more recent generation of cluster-based multi-vector engines with aggressive CPU-side optimizations. By plugging sparse centroid indices into both systems, the researchers could measure how much of the remaining bottleneck their approach removes on top of already highly tuned baselines. The evaluation covered both in-domain and out-of-domain retrieval tasks, using the widely adopted MS MARCO passage ranking dataset and the LoTTE benchmark, both of which are publicly available, which strengthens the reproducibility of the comparison.

The headline numbers are striking. Sparcent delivers up to 4.6 times speed-up over the original PLAID and 1.3 times speed-up over WARP, all while maintaining retrieval effectiveness on both in-domain and out-of-domain tasks. Equally important for practical deployment, the approach requires virtually zero additional memory. In search engine engineering, memory overhead is often the hidden tax that kills otherwise attractive acceleration techniques, since every extra structure must be kept resident alongside the main index. A method that speeds up query processing without inflating the index footprint is therefore unusually well suited to production environments, where hardware budgets are fixed and index sizes are already measured in hundreds of gigabytes.

The experimental design reflects a realistic deployment assumption. Following common practice in web search engines, which typically prioritize inter-query throughput over intra-query latency, the authors assume a maximum allocation of eight threads per individual query. This framing matters because it shifts the evaluation question from how quickly a single query can be answered with unlimited resources to how many queries a fixed-capacity cluster can absorb per second. Under that lens, a 4.6-fold reduction in per-query work translates directly into a proportional increase in sustainable query load, or equivalently into a substantial reduction in the number of machines needed to serve a given traffic level. For operators of large-scale search services, that is the difference between an interesting research result and an actionable cost saving.

The work also sits within a broader and rapidly evolving research conversation about efficiency in neural information retrieval. The reference list of the paper traces the lineage from ColBERT and ColBERTv2, which introduced contextualized late interaction, through dense retrieval models trained with approximate nearest neighbor techniques, to the SPLADE line of learned sparse models and the recent wave of efficiency-focused systems such as SLIM, guided traversal methods, and block-max pruning for learned sparse representations. Hybrid approaches that fuse sparse and dense signals have been gaining momentum, and Sparcent contributes a distinctive twist: rather than combining sparse and dense scores at ranking time, it uses sparse retrieval as the acceleration substrate for a dense pipeline. The authors’ own prior work on dynamic pruning, document reordering, and score caching in inverted indexes informs this design, grounding it in decades of classical query processing wisdom.

Why does the centroid stage dominate the cost profile in the first place? In cluster-based multi-vector retrieval, the number of centroids is far smaller than the number of token embeddings, but every query must still be scored against every centroid candidate before pruning decisions can be made, and these comparisons involve full-dimensional dense vectors. As corpus sizes grow, centroid counts grow too, and the coarse stage becomes a first-class bottleneck rather than a cheap prelude. By converting centroid matching into sparse inverted-index traversal, Sparcent benefits from the extreme selectivity of learned sparse representations: only the handful of postings that actually contain the query’s high-weight expansion terms need to be visited, and well-established threshold estimation techniques can terminate traversal early once the top candidates are safely identified.

The implications extend beyond academic benchmarks. As retrieval-augmented language models increasingly depend on fast, high-quality search over massive text collections, the efficiency of the retrieval layer directly shapes the latency, cost, and energy footprint of AI systems that answer questions, summarize documents, and power enterprise search. Techniques like Sparcent suggest that the supposed rivalry between sparse and dense retrieval is better understood as a partnership, in which each paradigm covers the other’s weaknesses. The authors report no competing interests, and the work was partially funded by Middle East Technical University under grant ADEP-312-2024-11484. The code is maintained in a GitHub repository that will be made publicly available upon publication, and all datasets used in the study are openly downloadable, so practitioners and researchers can verify the speed-ups and explore the sparse centroid construction for themselves. If the reported gains hold at production scale, the era in which the most accurate neural retrievers were too expensive to deploy may be drawing to a close.

Subject of Research: Accelerating multi-vector dense retrieval using learned sparse centroid indices

Article Title: Sparcent: faster multi-vector dense retrieval with sparse centroid indices

Article References: Yafay, E., & Altingovde, I. S. (2026). Sparcent: faster multi-vector dense retrieval with sparse centroid indices. Knowledge and Information Systems, 68(1), Article 280. https://doi.org/10.1007/s10115-026-02907-x

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02907-x

Keywords: multi-vector dense retrieval, learned sparse retrieval, Sparcent, PLAID, WARP, inverted indexes, centroid pruning, neural information retrieval, query efficiency, ColBERT, SPLADE, search engines