Sieve: A New Search-Inspect-Fetch Strategy for Deep-Research Agents
A recent study published on arXiv (2608.02751) presents Sieve, a novel search-inspect-fetch methodology designed for deep-research agents. In contrast to traditional Search-Visit processes that download entire webpages, Sieve utilizes a Boolean Query Language (BQL) to refine webpage fields, prioritize candidates, and retrieve only specific sections. In evaluations across three QA collections, Sieve surpasses the most effective standard configuration while consuming 20.7–50.6% fewer tokens. The use of Boolean filtering enhances every ranker tested, and the accuracy-context benefit remains consistent across different retrievers and agent architectures. Authored by a team of researchers, this cross-type submission offers an implementation, although the abstract is incomplete. This research aims to mitigate the inefficiency of incorporating irrelevant information into agent context by leveraging webpage structure.
Key facts
- Paper arXiv:2608.02751 introduces Sieve, a search-inspect-fetch strategy.
- Sieve uses a Boolean Query Language (BQL) to filter webpage fields.
- Sieve outperforms conventional Search-Visit configurations on three QA collections.
- Sieve uses 20.7–50.6% fewer tokens than conventional methods.
- Boolean filtering improves every tested ranker.
- The accuracy-context advantage persists across retriever choices and agent backbones.
- The paper is a cross-type submission on arXiv.
- The implementation is mentioned but details are cut off.
Entities
Institutions
- arXiv