Question Answering System Based on Retrieval-augmented Generation

Abstract

With the advancement of deep learning, large-scale neural networks have been widely applied in the field of natural language processing. However, question-answering systems based on large models suffer from issues such as hallucinations and outdated information, and struggle to capture complex relationships between entities, leading to biased results. To address these challenges, this paper aims to design and implement a digital human question-answering system that integrates a knowledge graph with vector retrieval. The system utilizes Unreal Engine to construct a high-fidelity 3D virtual museum scene, providing users with an immersive visual foundation.The system comprises four modules: the Digital Human Driver Module, the Digital Human Avatar Module, the Knowledge Graph Module, and the UE Presentation Module. The Driver and Avatar Modules facilitate interaction with large language models for natural language understanding and dialogue generation. The Knowledge Graph Module employs Enhanced Retrieval-Augmented Generation technology to retrieve relevant information from the knowledge graph before the large language model generates responses, thereby enhancing answer accuracy, reliability, and knowledge depth. The Presentation Module realizes high-fidelity exhibition scenes.These four modules collectively implement a novel architecture for an intelligent museum guide system. This architecture integrates UE's immersive environment,,Fay-driven digital human interaction, and LightRAG-empowered KG-RAG knowledge services. The goal is to deliver a more engaging and intelligent user experience.