Improving Reliability of Retrieval-Augmented Generation by Detecting and Correcting Irrelevant Retrieved Context

Retrieval-Augmented Generation (RAG) lets a language model answer from external documents instead of relying on its parameters alone, yet the answer can be no more dependable than the passages placed in its prompt. Retrievers rank by similarity, and similarity is not the same as usefulness: passages that share vocabulary or subject matter with a query frequently lack the evidence it needs, and a generator may still follow them. This paper reviews three recent lines of work that attack RAG reliability from different directions—corrective retrieval (CRAG), a word-level hallucination corpus (RAGTruth) and counterfactual risk control—and compares what each does before, during and after generation. Using their published results, we show that judging retrieval quality ahead of generation brings large gains, whereas detection and abstention methods only guard the output once the context is already fixed. Building on this analysis we outline a pre-generation framework that scores every retrieved chunk, separates relevant, partially relevant, irrelevant and insufficient evidence, refines or re-retrieves within a bounded budget, and abstains when evidence stays inadequate. A latency model and an evaluation protocol for testing the framework are provided. The paper closes with open problems in evaluator accuracy, cost and cross-domain robustness.