Answers are grounded in this paper and cite their sources.
Existing retrieval-augmented generation (RAG) methods improve access to external knowledge but do not adequately coordinate three critical requirements: determining whether retrieved evidence is permissible, acquiring sufficient evidence without unnecessary retrieval, and establishing whether that evidence materially supports generated claims. Fixed-depth retrieval can waste resources or miss necessary information, adaptive retrieval does not necessarily establish evidential sufficiency, and relevant citations may neither sustain a claim nor satisfy licensing, privacy, freshness, or jurisdictional constraints. To address these limitations, we propose BEL–RAG, a unified framework that treats retrieval as controlled evidence acquisition rather than context collection. BEL–RAG integrates Policy-Aware Retrieval, Bayesian Evidence Leasing, Counterfactual Auditing, and Calibrated Generation and Reporting within a closed inference loop. Policy constraints restrict evidence before inference; admissible passages are acquired according to their posterior contribution and retrieval cost; and minimum-support identification, counterfactual necessity, and sufficiency tests determine whether the acquired evidence sustains individual claims. Inadequate support triggers further retrieval, qualification, or abstention. The central novelty lies in coupling these mechanisms so that admissibility, evidence sufficiency, and claim-level support jointly govern retrieval and generation. Experiments across open-domain, multi-hop, knowledge-graph, implicit-reasoning, and policy-sensitive settings show that BEL–RAG achieves the highest reported answer-quality scores among the evaluated methods, including 74.9% F1 on HotpotQA and 90.1% accuracy on MetaQA, while using 3.18 mean passages and attaining 0.895 citation F1 and a 4.1% unsupported-claim rate. In matched policy-control comparisons, BEL–RAG achieves a 0.5% policy-violation rate, reducing violations by 37.5% relative to deny-by-default access control and 77.3% relative to budget-matched filtered dense RAG, while also reducing latency by 19.5% and 7.8%, respectively. These findings support the value of integrating policy enforcement, evidence-efficient retrieval, and counterfactual claim auditing to improve the reliability, efficiency, and traceability of RAG. The BEL–RAG reproducibility package is publicly available and includes source code, experimental configurations, evaluation scripts, datasets, benchmark results, and supporting documentation for implementing and evaluating the proposed framework.
No takes yet. Share an insight, caveat, or question.
Kamal Taha (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: