Running summarizing tools on a new article

Abstract: Large Language Models (LLMs) offer significant potential for educational chatbots, but their tendency to generate hallucinations and fluent yet factually unsupported responses limits their reliability in high-stakes academic settings. This study presents the design and evaluation of a university-focused virtual assistant, built on a Retrieval-Augmented Generation (RAG) architecture, to support academic and administrative inquiries at Greenwich Vietnam. The system integrates a structured institutional knowledge base derived from official documents, internal portals, regulations, and websites, along with hybrid retrieval mechanisms, including lexical and semantic search, query routing, and reranking, to improve evidence selection and contextual grounding. We evaluated the system on a benchmark comprising nearly 700 real-world and expert-constructed question-answer pairs with associated reference sources. Experimental results demonstrate strong retrieval and generation capabilities, achieving a Context Relevance score of 0.92 and a Semantic Answer Similarity score of 0.87, while also receiving high human evaluation scores for accuracy, completeness, freshness, faithfulness, and naturalness. The findings indicate that the proposed RAG-based approach substantially improves the trustworthiness of educational chatbot responses while remaining practical for institutional deployment. At the same time, the study highlights ongoing challenges in handling complex queries and maintaining long-term knowledge freshness, offering useful guidance for future domain-specific RAG systems in higher education.