Lima, Jean Mario Moreira deCabral, João Pedro Freire2025-01-242025-01-242025-01-21CABRAL, João Pedro Freire. Chatbot Inteligente para acesso a regulamentos acadêmicos: um sistema de recuperação de informações baseado em RAG. 2025. 58 f. Trabalho de Conclusão de Curso (Graduação em Engenharia de Computação) - Departamento de Engenharia da Computação, Universidade Federal do Rio Grande do Norte, Natal, 2025.https://repositorio.ufrn.br/handle/123456789/61922The Brazilian chatbot market is experiencing rapid growth, marked by a notable increase in both the number of active bots and the volume of processed messages. These systems hold significant potential for modernizing services, particularly in the public sector, by streamlining the dissemination of complex regulations and enhancing accessibility for citizens and public servants. In parallel, there has been a surge in academic interest in large language models, evidenced by the growing volume of scientific publications that highlight their ability to understand and generate natural language with greater contextual accuracy. Against this backdrop, the present study aims to develop an intelligent chatbot based on large language models to facilitate access to the academic regulations of UFRN, thereby promoting improved accessibility and information accuracy. To address challenges such as the need for up-to-date data and more contextualized responses, we adopt the Retrieval-Augmented Generation approach, which combines text generation with external data retrieval. During the development phase, the structured dataset created with the RAGAS framework, alongside a well-considered semantic segmentation of the regulations, proved essential for ensuring consistent responses. Test results indicated that a traditional retrieval method enhanced by reranking yielded the most satisfactory outcomes, while techniques such as multiple queries and hypothetical document generation did not demonstrate satisfactory performance. The Rerank 1.0 model from Amazon played a key role in filtering irrelevant documents, thereby improving the precision and reliability of the responses. In a production environment, the chatbot showed robust scalability, consuming an average of 241 MB of memory and achieving a response time of approximately 4.5 seconds. Nevertheless, challenges remain, including computational costs and the need for continual adjustments. Overall, this study provides a solid and replicable approach to implementing Retrieval-Augmented Generation chatbots, with potential applications in both academic and administrative contexts that demand high precision and contextualization of information.Attribution-NonCommercial-NoDerivs 3.0 Brazilhttp://creativecommons.org/licenses/by-nc-nd/3.0/br/ChatbotNormasModelos de linguagem de grande escalaAcesso à informaçãobachelorThesis