This study explores creating and evaluating chatbots for college websites, focusing on their performance with limited data. Chatbots can aid visitors, teachers, and students, but data scarcity often leads to incorrect responses from neural network-based chatbots. The research examines various models, including CNNs, RNNs, and advanced Language models like BERT, T5, and Rasa, and explains why building an LLM from scratch is impractical. Despite challenges like emotional understanding, gender bias, and contextual comprehension, the study optimises LLMs for university-specific tasks. Findings show that fine-tuned LLMs outperform standard models in sparse data scenarios, revamping the experience of users by providing precise and faultless timely information on college websites and highlighting broader implications for intelligent agents in data-scarce environments.