D2R-RAG Diagnoses and Repairs Factual Errors in RAG Systems

Retrieval-Augmented Generation (RAG) systems, while improving LLM factuality, often suffer from fragility due to missing or irrelevant evidence, or generations that don't faithfully reflect the retrieved context. Existing solutions often require fine-tuning or privileged model access, limiting their practical use in black-box or resource-constrained environments. To address these limitations, D2R-RAG (Diagnose-to-Repair RAG) has been developed as a model-agnostic and resource-aware framework. It identifies failure signatures from observable signals in the query, retrieved evidence, and generated response, then selects from a small set of corrective actions while adhering to explicit latency and VRAM budgets. Experiments on benchmarks like FEVER and HotpotQA show that D2R-RAG significantly enhances reliability and achieves superior accuracy-efficiency trade-offs compared to other methods.