RAG for Question-Answering for Vocal Training Based on Domain Knowledge Base
Chun-hung Jonas Leung, Yicheng Yi, Le Kuai, Zongxi Li, Siu-Kei Au Yeung, Kwok-wah John Lee, Ka-him Kelvin Ho, Kevin Hung · 2024
Although Large language models (LLMs) are well-known due to their superior capacity for text generation and logical inference, they are found to be inaccurate in domain-specific question-answering tasks. The powerful generator still tends to generate content even when the LLM does not have sufficient knowledge at all, which is known as the hallucination problem. We find there is a research void in applying LLMs in the vocal training industry, which requires intensive expert knowledge in any chatbot or intelligent tutor services. This paper details employing Retrieval-Augmented Generation (RAG) technology to develop a domain-specific language model, addressing inherent challenges such as hallucination, where large models generate plausible but inaccurate content, and lack of domain specificity. By segmenting the knowledge base and establishing semantic similarities between user queries and knowledge data, the project lays a solid foundation for integrating RAG, significantly improving response accuracy and contextual relevance. The report highlights the successful implementation of RAG, enhancing system intelligence and personalization for user-specific needs, discusses challenges and solutions during the implementation process, and outlines future directions to expand RAG capabilities and improve user experiences.