CAR: Confidence-Aware Reranking Boosts RAG Generation Quality

Introducing CAR (Confidence-Aware Reranking), a novel framework that enhances retrieval-augmented generation without the need for training. This approach shifts focus from traditional query-document relevance to the generator's confidence change, which is assessed through the semantic consistency of sampled responses. Documents that bolster confidence are elevated in rank, while those that diminish it are lowered, and ambiguous cases maintain their original ranking. A query-level gate ensures that confident queries remain unaffected. Testing on four BEIR datasets demonstrates that CAR surpasses current reranking techniques.