LiRe: Efficient Query Rewriting for Retrieval Augmented Generation Systems

With the continuous advancement of information retrieval technologies, query rewriting has become a key technique for enhancing retrieval effectiveness. In the context of the widespread application of large language models and retrieval-augmented generation technologies, high-quality query rewriting plays a crucial role in tasks like open-domain question answering. Conventional Conversational Query Rewriting (CQR) methods often prioritize human-friendly query formulation, which may not consistently produce optimal retriever-friendly results. To address this limitation, we present LiRe, a lightweight and training-free query rewriting system. LiRe employs a retrieval-oriented strategy to refine queries, optimizing them for retrievers. This approach allows for seamless integration into diverse Retrieval-Augmented Generation (RAG) systems. Our approach involves problem categorization, keyword augmentation from contextual information, and integration with the filtering process. LiRe-enhanced queries demonstrate superior performance compared to existing CQR methods, as evidenced by experimental outcomes.