Joint Optimization of Dual Retrieval and Adaptive Re-Ranking for Retrieval-Augmented Generation

Abstract

This paper proposes a Retrieval-Augmented Generation framework based on sparse–dense dual retrieval fusion combined with an adaptive re-ranking algorithm to improve ranking precision and generation consistency. Sparse retrieval ensures high coverage, dense retrieval provides semantic matching ability, and dynamic weight allocation is introduced during re-ranking to balance the importance of evidence from different sources. An end-to-end optimization objective is constructed that jointly constrains ranking relevance and generation consistency, enabling retrieval and generation to be improved in coordination. The framework is systematically evaluated across hyperparameters, environment, and data sensitivities, covering candidate size, caching strategy, domain transfer, label sparsity, and feedback noise. Results show that the proposed method outperforms existing approaches across multiple metrics, maintaining stable advantages in candidate relevance, recall rate, and semantic consistency of generation. This study demonstrates that dual-path fusion and adaptive re-ranking effectively mitigate the conflict between coverage and precision in complex retrieval environments and provide new ideas and methods for building robust RAG systems.