SIEVE deep-research agent cuts token use by 50% with Boolean retrieval
New SIEVE interface filters web pages by fields, slashing tokens 20.7-50.6% while boosting accuracy across three QA sets.
A new paper introduces SIEVE, a search-inspect-fetch interface driven by fielded Boolean retrieval (BQL) for deep-research agents. Instead of retrieving whole pages, SIEVE filters candidates over document fields, ranks the admitted set, presents structure-rich result cards for inspection, and fetches only selected sections. Across three QA collections, SIEVE achieves higher accuracy than the most accurate conventional search-visit configuration on each collection while using 20.7–50.6% fewer tokens. Further analyses show that BQL filtering improves all tested rankers and that the accuracy-context advantage persists across retriever choices and agent backbones.
- SIEVE uses fielded Boolean retrieval (BQL) to filter documents by title, heading, section, and metadata before fetching
- Outperforms conventional search-visit agents on 3 QA collections while using 20.7–50.6% fewer tokens
- BQL filtering improves all tested rankers; gains persist across retrievers and agent backbones
- Code and data are publicly available
Why It Matters
SIEVE cuts cost and latency for AI research agents while improving answer quality—a key breakthrough for scalable deep-research systems.