ScoreGate: Adaptive Chunk Selection for Retrieval-Augmented Generation via Dual-Score Statistical Fusion

ScoreGate is a novel mechanism for adaptive retrieval-augmented generation that optimizes chunk selection during inference by leveraging bi-encoder similarity and cross-encoder reranker scores without additional model inference. It demonstrates improved retrieval efficiency, achieving a mean reciprocal rank (MRR@10) of 0.401 on the MS MARCO dataset while reducing the number of retained chunks by 35%, and maintaining high recall rates with zero false positives. This approach allows practitioners to refine retrieval processes, enhancing performance in scenarios with variable query complexity while minimizing latency and token usage.