Comparison and Dynamic Fusion of RAG One-Stage Retrieval Methods for Heterogeneous Queries
Abstract
Retrieval-Augmented Generation (RAG) strengthens the factuality of large language model outputs by grounding the generation process in retrieved external knowledge. Given that the efficacy of one-stage retrieval strictly defines the evidence available for generation, it is imperative to analyze the applicable boundaries of various retrieval methods under heterogeneous queries. This study compares BM25, a DPR-style dense retrieval baseline, and several fusion retrieval strategies. Queries are systematically categorized into terminology, semantic, and mixed types, utilizing the ‘enhanced mid’ dataset as the primary experimental corpus. Performance is evaluated using Top-1 Accuracy, Recall@3, and MRR@3. Experimental results demonstrate that the proposed Dynamic v2b method achieves the highest Top-1 Accuracy and MRR@3, at 82.50% and 0.8708, respectively. However, grid search analysis reveals that a fixed weight of 0.2 (BM25) + 0.8 (Dense) slightly outperforms it in Recall@3 and MRR@3, indicating that while the Dense-biased design of Dynamic v2b is effective, its dynamic weighting strategy requires further refinement. Finally, generation-side auxiliary scoring indicates that although retrieval improvements significantly enhance local sample quality, the overall impact on generation performance remains relatively limited.