ChatBU

Abstract

Information delivery in modern academic environments is frequently hindered by data fragmentation within massive documentation and the limitations of conventional lexical- based search engines. At Bahria University, these challenges are compounded by the lack of 24/7 administrative scalability and the high risk of "hallucination" when using generic AI models for critical data like fee structures and admission deadlines. This research addresses these gaps by developing ChatBU, an intelligent, production-grade assistant designed to provide grounded, real-time institutional knowledge. The first phase of this project explored the viability of Parameter-Efficient Fine-Tuning (PEFT) to embed university knowledge directly into a model's neural weights. Utilizing the FLAN-T5-Base architecture with LoRA and 4-bit QLoRA quantization, a baseline model was successfully trained in a low-resource environment. However, empirical testing revealed significant "parametric rigidity"; the model could not efficiently adapt to dynamic policy changes and suffered from numerical hallucinations, making it unsuitable for high-stakes institutional deployment. To resolve these limitations, the research pivoted to a Hybrid Retrieval-Augmented Generation (RAG) pipeline hosted on a distributed cloud architecture. By decoupling the knowledge base from the reasoning engine, we engineered a high-density vector space using intfloat/e5-base-v2 embeddings and a FAISS index. This non-parametric approach allows the system to perform semantic retrieval from official university PDFs, which are then synthesized by the Gemini-1.5-Flash API. To ensure production-grade availability and security, the system was deployed via a zero-trust architecture with automated process orchestration, ensuring the assistant is protected from external threats while maintaining a 99.9% uptime. The final contribution of this dissertation is a rigorous comparative analysis of these two approaches. The findings show that the RAG-based system achieved 98.2% accuracy, significantly outperforming the 78.4% accuracy of the fine-tuned baseline. The integration of Django analytics and Power BI dashboards also provides valuable insights into student inquiry trends. Overall, the project demonstrates that combining semantic retrieval, cloud APIs, and resilient infrastructure creates a scalable and reliable framework for future campus automation.