URAG: Uncertainty-Guided Agentic Retrieval-Augmented Generation with Confidence-Calibrated Adaptive Reflection

We present URAG (Uncertainty-guided RAG), a lightweight agentic Retrieval-Augmented Generation system that replaces binary reflection triggers with a continuous joint confidence score, computed from cross-encoder retrieval quality signals and NLI-based generation faithfulness, to allocate an adaptive reflection budget. A central architectural contribution is the generate-then-assess paradigm: generating a candidate answer before estimating confidence ensures the system scores actual answer faithfulness rather than relying on retrieval proxies alone. Evaluated on HotPotQA and 2WikiMultihopQA (n=500 per dataset), URAG achieves F1=0.37 and hallucination rate=0.07 on HotPotQA, representing a 76% improvement in F1 over Naive RAG and a 75% reduction in hallucination rate. The complete system runs on a consumer RTX 4050 laptop GPU (6 GB VRAM) using exclusively free, open-source models. Code: https://github.com/kShaman771/URAG