From Retrieval Recall to Claim-Level Provenance: Reliability Design for Retrieval-Augmented Question Answering

Retrieval-augmented generation (RAG) permits language models to consult external evidence but does not make generated statements automatically faithful. This review analyzes reliability across document selection, retrieval, evidence composition, answer construction, citation alignment, and uncertainty-aware refusal. It distinguishes a relevant passage from a passage that actually entails a specific claim, and shows why retrieval recall, answer correctness, and evidence faithfulness should be evaluated separately. Established work on dense passage retrieval, retrieval-augmented generation, factuality assessment, and automatic RAG evaluation is synthesized into a claim-level auditing framework. A practical test protocol includes dated sources, contradictory documents, unsupported questions, and malicious or misleading retrieved passages. The findings are conceptual: the major gap is not merely retrieving more text, but preserving traceable support for individual assertions while controlling abstention and latency. No new benchmark scores or model evaluations are reported.