Uncertainty-Aware Hybrid Retrieval for Long-Document RAG

The article introduces Uncertainty-aware Multi-Granularity RAG (UMG-RAG), a hybrid retrieval framework for Retrieval Augmented Generation (RAG) that optimizes the granularity of retrieved evidence based on query-specific reliability estimation without requiring additional training. UMG-RAG leverages both dense and sparse retrievers to create an evidence distribution from which it estimates reliability using distribution entropy, enhancing the quality of generated responses. Additionally, the proposed UMGP-RAG variant improves local coherence by using fine-grained hits to identify relevant evidence while returning broader parent chunks, demonstrating significant improvements in generation quality on question answering benchmarks.