Scene-Aware Multi-agents on Collaborative Retrieval Augmented Generation Model (CRAG)

The RAG question answering of large model LLMs is currently a hot research direction, widely used in artificial intelligence question answering, such as ChatGPT. RAG Q&A is the process of decomposing and collaborating RAG tasks through multi-agents. However, due to the poor correlation between the retrieved knowledge sources and the poor adaptability of the generated answers to the scene, it brings a problem of semantic illusion when generating semantics in RAG, that is, answers that are not the question asked. In response to this issue, this article proposes a scene-aware multi-agents on collaborative retrieval augmented generation (CRAG) model. CRAG is a master-slave network model, where the main network and the slave network learn semantic information in the scene through the hint module. Specifically, multi-agents achieve consensus among all agents by deploying the same slave network to generate a shared perception of scene information. All agents, with the assistance of scene semantic information, think around the same question when answering RAG Q&A. Finally, comparative experiments were conducted on the TruthfulQA dataset to demonstrate the effectiveness of the proposed method.