Study Shows Complementary Query Rewriting Boosts Strong RAG Baselines

A scientific preprint published on arXiv examines the effectiveness of query rewriting in Retrieval-Augmented Generation (RAG) when applied on top of an already strong retrieval pipeline featuring BGE dense retrieval, cross-encoder reranking, and MMR diversification.

The authors found that single rewriting methods barely outperform a strong baseline on their own. However, uniting complementary approaches (S1+S3+S4+HyDE) produced significant improvements, raising HIT@10 on the 512K-document EnterpriseRAG-Bench from 39.22 to 51.70 (+12.5 points), while a five-method union reached 52.98.

To mitigate computational costs, the researchers evaluated a confidence-gated router that triggers rewriting only when baseline confidence is low. This mechanism activated on under 40% of queries, capturing around half the enterprise merge gains and improving downstream answer quality F1 by +1.92.