<![CDATA[This study addresses the challenge of providing students with fast and accurate academic information by developing a document-based question-answering (Q&A) system. Many students face difficulties navigating complex campus administrative documents, from trivial inquiries to graduation policies. To solve this, we propose a novel Retrieval-Augmented Generation (RAG) system that uses a local Large Language Model (LLM) and a FAISS index for efficient information retrieval. The system's architecture ingests academic documents (PDF, DOCX), splits them into context-aware chunks, and converts them into vector representations using the all-MiniLM-L6-v2 embedding model. These vectors are then stored in topic-specific FAISS indices to enhance retrieval accuracy. Our implementation uses Qwen2.5:3B, a locally run LLM, to ensure offline functionality and minimal hardware requirements. The system is also equipped with an adaptive prompt mechanism and a domain-switching feature to handle diverse question styles. Evaluations using black-box and white-box testing confirmed the system's robustness, with an 80% success rate in automatic domain detection and an average retrieval time of 1.6 seconds. The results indicate that the primary performance bottleneck resides in the FAISS retrieval process rather than the generative model. This research demonstrates that a locally run RAG system can provide a practical, user-friendly, and effective solution for enhancing academic support services. It establishes a foundation for future developments in intelligent educational interfaces.]]>
Copyrights © 2025
