SIEVE: Effective Filtered Vector Search with Collection of Indexes
Zhaoheng Li, Silu Huang, Wei Ding, Yongjoo Park, Jianjun Chen · 2025
Read SIEVE: Effective Filtered Vector Search with Collection of Indexes by Zhaoheng Li, Silu Huang, Wei Ding, Yongjoo Park, Jianjun Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-11907.txt.
Description
Zhaoheng Li and co-authors tackle a search problem that mixes likeness with hard rules. Many tasks, such as finding videos tagged for kids, need the items most similar to a query among only those that pass set filters. Their 2025 paper notes that graph-based search, fast without filters, slows badly once strict rules cut most items out of reach.
Rather than steer one graph around the rules, SIEVE builds many indexes, each tuned to a different form of filter. A three-part model of index size, search time, and recall guides which indexes get built for a given workload. At query time the same model picks the index that should answer fastest at the wanted recall. Tests across datasets with varied rule strictness showed stronger results than prior methods, which had to limit the rules or lose speed.
Catalog record
| CREATOR | Zhaoheng Li, Silu Huang, Wei Ding, Yongjoo Park, Jianjun Chen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 76 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-11907