RQ-RAG: Learning to Refine Queries for Retrieval-Augmented Generation This paper introduces RQ-RAG, a framework that enhances large language models (LLMs) by training them on a carefully curated dataset. The model is trained to refine queries through rewriting, decomposition, and disambiguation. The key idea is to enable the model to dynamically refine search queries based on the context retrieved from external sources, thereby improving the accuracy and relevance of generated responses. The paper addresses the limitations of LLMs, which are prone to generating inaccurate or hallucinatory responses due to their reliance on pre-trained data. RQ-RAG integrates retrieval functionalities into generative models, allowing them to access relevant information from external databases. This approach enriches standard parametric language models with non-parametric retrieval elements, enabling them to access up-to-date information. The proposed method, RQ-RAG, is trained on a dataset that includes sequences of actions with special tokens, specifying the type of refinement (rewrite, decompose, disambiguate) and the refined query. The model is trained to generate responses based on the retrieved information, with the ability to dynamically refine queries through rewriting, decomposition, and disambiguation. The paper evaluates RQ-RAG on three single-hop QA tasks and three multi-hop QA tasks. The results show that RQ-RAG outperforms previous state-of-the-art methods by an average of 1.9% on single-hop tasks and demonstrates superior performance on multi-hop tasks. The method also shows high potential for system performance, with an upper bound of 76.8% on ARC_C, 65.6% on POPQA, 84.0% on OBQA, 80.5% on HOTPOTQA, 60.6% on 2WIKI, and 54.5% on MUSIQUE. The paper also highlights the effectiveness of regenerating answers based on search results during data construction, which proves more effective than previous methods. Additionally, the system is shown to be resilient to different data sources, with minimal impact on performance when using different retrieval sources. The paper concludes that RQ-RAG is an effective approach for enhancing LLMs by enabling them to refine queries through rewriting, decomposition, and disambiguation, leading to improved accuracy and relevance of generated responses.RQ-RAG: Learning to Refine Queries for Retrieval-Augmented Generation This paper introduces RQ-RAG, a framework that enhances large language models (LLMs) by training them on a carefully curated dataset. The model is trained to refine queries through rewriting, decomposition, and disambiguation. The key idea is to enable the model to dynamically refine search queries based on the context retrieved from external sources, thereby improving the accuracy and relevance of generated responses. The paper addresses the limitations of LLMs, which are prone to generating inaccurate or hallucinatory responses due to their reliance on pre-trained data. RQ-RAG integrates retrieval functionalities into generative models, allowing them to access relevant information from external databases. This approach enriches standard parametric language models with non-parametric retrieval elements, enabling them to access up-to-date information. The proposed method, RQ-RAG, is trained on a dataset that includes sequences of actions with special tokens, specifying the type of refinement (rewrite, decompose, disambiguate) and the refined query. The model is trained to generate responses based on the retrieved information, with the ability to dynamically refine queries through rewriting, decomposition, and disambiguation. The paper evaluates RQ-RAG on three single-hop QA tasks and three multi-hop QA tasks. The results show that RQ-RAG outperforms previous state-of-the-art methods by an average of 1.9% on single-hop tasks and demonstrates superior performance on multi-hop tasks. The method also shows high potential for system performance, with an upper bound of 76.8% on ARC_C, 65.6% on POPQA, 84.0% on OBQA, 80.5% on HOTPOTQA, 60.6% on 2WIKI, and 54.5% on MUSIQUE. The paper also highlights the effectiveness of regenerating answers based on search results during data construction, which proves more effective than previous methods. Additionally, the system is shown to be resilient to different data sources, with minimal impact on performance when using different retrieval sources. The paper concludes that RQ-RAG is an effective approach for enhancing LLMs by enabling them to refine queries through rewriting, decomposition, and disambiguation, leading to improved accuracy and relevance of generated responses.