Desarrollo de un modelo de Inteligencia Artificial para la generación automática de respuestas a preguntas frecuentes de los estudiantes de la UTPL

Abstract

Abstract: In this work, an artificial intelligence model based on a Retrieval-Augmented Generation (RAG) architecture was developed to improve response to frequently asked questions from students at the Universidad Técnica Particular de Loja (UTPL). During the process, four vectorization models were evaluated (paraphrase-MiniLM-L6-v2, all-MiniLM-L12-v2, BERT-base, and RoBERTa-base), with the BERT-base model being selected for its outstanding performance in semantic similarity. For the response generation phase, the GPT-3.5-Turbo and Gemini Flash 1.5 models were evaluated. The BLEU and ROUGE metrics were used. The GPT-3.5-Turbo model outperformed Gemini Flash 1.5 by maintaining a good balance between relevance and contextualization, and was therefore the selected model. Based on this, a functional pilot was implemented in a test chatbot within the UTPL+ app, allowing for simulated queries. This pilot demonstrates that the model generates responses tailored to user needs, contributing to improving the student support experience and the quality of university life.