UFCHAT: uma proposta de assistente de inteligência artificial para apoio informacional na UFCA

Abstract

In the academic environment, the dispersion and volume of information hinder quick access to essential data by students, staff, and professors, such as institutional contacts, links, and documents. The absence of a centralized channel compromises communication efficiency and overloads university agents with recurring questions. This work proposes the use of Large Language Models (LLMs) as conversational assistants capable of acting as virtual receptionists, offering fast and contextualized responses to the institution’s informational demands. To evaluate the receptivity of the academic community, a questionnaire was applied regarding the prior use of Artificial Intelligence (AI) tools and the functionalities that would make a virtual assistant useful for UFCA activities. Based on these indications, UFChat is presented, an LLM model based on GPT-4.1 mini, which uses a Semantic Router mechanism to classify and route user requests, followed by integration with a Vector Database through Retrieval Augmented Generation (RAG) technology, responsible for information retrieval and response improvement. The system also incorporates Web Scraping techniques for data collection from official sources, allowing the generation of more precise and institutionally aligned responses. Tests involving the UFCA academic community indicate that UFChat produces more effective and preferable responses compared to other commercial assistants, such as ChatGPT and Gemini, in addition to presenting the flexibility of operation via WhatsApp, contributing to the improvement of institutional communication.