A Privacy-Preserving AI and Machine Learning Based Smart Academic Support Chatbot Using Locally Hosted Large Language Models and Retrieval-Augmented Generation

Abstract: Academic institutions receive a continuous stream of repetitive student enquiries concerning admissions, fees, examination schedules, library resources, and course logistics, placing a sustained burden on administrative staff and often leaving students without timely answers. This paper presents an artificial-intelligence and machine-learning based academic support chatbot that resolves such enquiries conversationally while keeping all data and inference on institutional hardware. The proposed system combines a machine-learning intent-classification pipeline with a locally hosted large language model served through ollama and a retrieval-augmented generation (RAG) component that grounds responses in a curated campus knowledge base, thereby reducing hallucination and eliminating dependence on paid cloud APIs. A Python back end implements the natural-language and retrieval logic, a Node.js layer delivers a responsive chat interface, and a lightweight relational store persists conversations and analytics. The system was evaluated against rule- based, retrieval-only, and cloud-LLM baselines using intent accuracy, answer relevance, response latency, and user- satisfaction metrics. Experimental observations indicate that the proposed framework achieved approximately 91% intent-classification accuracy and grounded-answer relevance of 0.89, while sustaining lower latency under concurrent load than the cloud baseline and preserving data privacy. The principal contributions are a privacy-preserving on-device intelligence layer, a hybrid intent-plus-RAG pipeline that improves factual grounding, and an analytics-driven feedback loop that continuously enriches the institutional knowledge base.

Keywords: Chatbot; natural language processing; large language models; retrieval-augmented generation ollama intent classification; academic support; on-device inference.