Context-Aware Question Answering for Vietnamese University Admissions via Multi-LLM Architecture

The increasing demand for transparent and accessible communication in university admissions highlights the need for intelligent systems to support students, especially in complex contexts like Vietnam. Traditional QA approaches, primarily based on rule-based systems and basic retrieval models, face scalability and contextual understanding challenges, and current Vietnamese benchmarks do not fully address admissions-specific needs. To overcome these gaps, we introduce a contextual chunking framework and a modular multi-LLM architecture to improve question answering performance over long, complex admission documents. Our framework addresses key limitations in semantic preservation and task modularity, enabling scalable educational chatbot systems. Experiments conducted on a newly constructed multimodal Vietnamese admissions dataset validate the effectiveness of our methods, achieving substantial gains in retrieval accuracy and answer generation quality compared to baseline approaches. These results demonstrate the promise of our approach for advancing intelligent educational QA systems in low-resource, policy-intensive domains.