Evidence Aware Retrieval Augmented Generation for Reliable Large Language Model Question Answering

Abstract

Large language models offer a flexible interface to technical knowledge, yet fluent answers can conceal unsupported statements. Retrieval augmented generation improves access to external evidence, but retrieval success alone does not establish answer reliability. This paper develops an evidence aware framework that treats retrieval, claim verification, and abstention as distinct decisions. The framework maintains source provenance, selects passages under a token budget, decomposes candidate answers into verifiable claims, and releases answers only when their support satisfies an explicit decision policy. A probabilistic analysis connects evidence sufficiency, generation fidelity, verifier behavior, answer coverage, and residual unsupported output. Deterministic sensitivity calculations evaluate these analytic relationships. Under one stated configuration, a specified support probability of 0.72 combined with verifier acceptance rates of 0.90 for supported answers and 0.15 for unsupported answers yields coverage of 0.690 and conditional unsupported risk of 0.061. The paper also specifies a reproducible evaluation protocol with paired baselines, source version controls, calibration analysis, and targeted robustness tests. The central conclusion is that reliable knowledge assistance requires explicit control over both evidence use and selective answering, supported by separate measurements of correctness, citation support, completeness, and operating cost.