Agentes conversacionales basados en LLMs para dispositivos móviles: construcción de un piloto para los estudiantes de tecnologías de la información

Abstract

Abstract: This project addresses the development of a conversational agent based on state-of-the-art language models. To achieve this, a complete workflow was designed, beginning with the collection and preprocessing of data related to academic processes within a university program. The data were then transformed into semantic vectors, enabling effective information retrieval in conversational contexts. The system s primary objective is not only to provide answers but to do so in a coherent, relevant, and contextualized manner, thereby ensuring its usefulness for the end user.To meet this goal, various tools were employed, with the RAG (Retrieval-Augmented Generation) architecture guiding the language model in generating context-based responses. Among the most notable tools used were Haystack, a framework for building modular pipelines; PyPDF, for extracting content from PDF files; and Trafilatura, for cleaning and structuring information from HTML pages by removing noise. Additionally, embeddings were applied through Sentence Transformers, converting text into numerical representations of semantic meaning. This approach allows for more intelligent searches, based not only on keywords but also on the meaning of phrases.The Mistral model was incorporated through an API to generate natural language responses from retrieved contexts. After implementation and a series of tests, results showed that the system can accurately retrieve information and produce clear, coherent, and contextually relevant answers in educational settings. In this way, students can resolve academic inquiries and receive timely and direct support. Finally, the study acknowledges its limitations, particularly the relatively small dataset used. Future work should therefore expand the information corpus, train new language models, and test the system with real users to maximize its impact and usability.