In the evolving landscape of higher education admissions, the deployment of robust chatbot systems is of paramount importance. This study compares two methodologies within the Qwen 2.5 7B Instruct model: QLoRA, which employs 4-bit quantization and low-rank adapters for fine-tuning, and RAG, which integrates an external retrieval pipeline for dynamic content updates. Using a domain-specific corpus derived from the official student handbooks and frequently asked questions (FAQs) published by Ho Chi Minh City Open University, we conducted evaluations through automated metrics (ROUGE-L, BLEU, BERT F1, and cosine similarity) as well as human and GPT-4o assessments. The results indicate that while QLoRA achieves superior text coherence, fluency, and semantic alignment, RAG provides greater update efficiency with lower computational overhead. To quantify these trade-offs, we propose the Update Efficiency Ratio (UER), which offers guidance for selecting chatbot architectures based on resource constraints and update frequency. In the future work, we will explore hybrid models that combine quantized fine-tuning with retrieval mechanisms under hardware limitations.