Add-Rag: Agent-Driven Dynamic Rag with Adaptive Retrieval Strategies and Multi-Retriever Collaboration for Enhanced Generation

Currently, Retrieval-Augmented Generation (RAG) is employed to mitigate the hallucination issue in Large Language Models (LLMs) during knowledge retrieval, by integrating external knowledge bases. However, it also exposes several problems, such as relatively fixed retrieval strategies and overly simplistic feedback mechanisms. To address these issues, we propose a highly lightweight agent capable of dynamically selecting appropriate retrievers and adjusting retrieval parameters in response to changes in the retrieved content, thereby substantially improving retrieval efficiency. Moreover, the framework allows for the customization of task-specific retrieval strategy spaces to meet diverse application requirements. During training, we introduce a multi-dimensional reward mechanism—comprising retrieval strategy rewards, re-retrieval advantage rewards, and policy rewards—into the agent to enhance its ability to determine optimal retriever selection and configure retrieval parameters. For queries of varying difficulty, we have developed a cross-retrieval approach that utilizes multiple cooperating retrievers. For re-retrieval, we will also re-adjust the retrieval strategy to adapt to the new retrieved content. We evaluated the framework on multiple datasets and conducted ablation studies to verify the effectiveness of our method.