Search, Inspect, Fetch: Exploiting Boolean Retrieval for Deep-Research Agents
The paper introduces SIEVE, a search-inspect-fetch interface for deep-research agents. Instead of retrieving and reading entire web pages, SIEVE applies fielded Boolean retrieval over titles, headings, sections, and metadata, ranks the admitted candidates, exposes structure-rich result cards for inspection, and fetches only selected sections. Across three QA collections, it reportedly outperforms the best conventional Search-Visit configuration on each collection while reducing token usage by 20.7% to 50.6%. The authors also report that Boolean filtering improves every tested ranker and that the accuracy-context tradeoff persists across retrievers and agent backbones.
Why it's worth reading
Deep-research agents increasingly face context waste from whole-page retrieval. SIEVE is timely because it turns web structure into an explicit retrieval interface and reports both accuracy gains and substantial token savings across datasets and agent configurations.