SIRA: Superintelligent Retrieval Agent

简体中文

L1 DF gate and one-pass weighted retrieval. No author LLM or 25.58-million-document index.

Paper information

Field	Details
Paper	arXiv
Institution	Meta Superintelligence Labs / Rice University (first author; work done at Meta).
First public release	2026-05-07 (arXiv v1)
Author code	facebookresearch/sira.
Local adapter or method key	sira
Local reproduction code	src/auto_research/agent_research/official_meta_backfill_20261004.py

Paper method and evidence

A frozen LLM proposes evidence terms missing from the original query. Filter absent or overly common terms by inverted-index document frequency; combine original and verified expansion in one weighted BM25 pass.

flowchart LR
  Q[Original query] --> E[Propose missing evidence terms]
  E --> F[Filter using document frequency]
  Q --> B[Single weighted BM25]
  F --> B

Original paper figure

The original figure shows expansion followed by DF filtering and a single retrieval. From the original paper; copyright remains with the authors.

Core equation

\(score(d)=BM25(q_{orig},d)+w\,BM25(q_{exp},d)\) with \(0<DF(t)\le\tau|C|\) for expansion terms.

Reported results

The paper reports better mean retrieval across ten BEIR tasks than compared dense, learned sparse, and agent retrieval baselines. BrowseComp-Wikipedia Recall@1/10/100: 9.70%/15.27%/36.14%. No attributable production A/B; author offline metrics are not local.

Local reproduction

Compare disabled/default with three-seed mechanism diagnostics. diagnostic_only=true, excluded from capability ranking. No paper LLM, full index, or online experiment.