A Chat Bot for Enrollment of Xi 'an Jiaotong-Liverpool University Based on RAG*

Recent advancements in large language models (LLMs) have established pre-training on extensive textual cor-pora as a foundational methodology. However, in specialised applications such as admissions systems, the focus shifts from general knowledge-based reasoning to ensuring accuracy and relevance in domain-specific responses. This study presents the development of an automated admissions system for Xi'an Jiaotong-Liverpool University, leveraging GLM-4 in conjunction with Retrieval-Augmented Generation (RAG) to handle targeted queries. The implementation of RAG mitigates the occurrence of hallucinations often seen in LLM outputs, thereby enhancing the reliability and alignment of generated responses with real-world data, which is critical for prospective students and their parents. This paper details the construction of the RAG corpus and cue word methodology, and provides an empirical comparison of the efficacy of various major language models with and without RAG integration. The results demonstrate the potential of RAG to significantly improveresponse accuracy in domain-specific tasks, and suggest directions for future research in optimising LLMs for admissions processes.