Search less. Read the right section.

SIEVE changes a deep-research agent's retrieval contract: it searches document fields, shows a structure-rich result card, and fetches only a named section instead of defaulting to the whole page. The v2 paper reports higher accuracy with 20.7–50.6% fewer tokens than the strongest conventional Search–Visit setup. 1

That targets a concrete product cost: whole-page visits flood context with material the agent did not need. In controlled tests on HotpotQA, MuSiQue, and BCP-S, SIEVE beat the strongest Search–Visit setup by 1.6, 3.1, and 0.7 accuracy points while using 30.4%, 50.6%, and 20.7% fewer tokens, respectively. 1

First build: preserve metadata and section boundaries in your index; add Boolean admission with field filters before your current BM25 or dense ranker; return headings plus a short query-focused snippet; then fetch one section. The paper's public pages link the SkimSearchAgent implementation. 2 Shadow it against your current full-page path and log task success, token use, calls, latency, and citation coverage. Keep the full-page fallback: the study did not establish performance on your live corpus, agent, or web. 1