Assistente universitário baseado em modelos de linguagem e geração aumentada por recuperação

Abstract

Quick and clear access to institutional information represents a challenge for higher education students, as regulations and academic documents are often dispersed and written in technical language. Aiming to mitigate this problem, this work details the development of an intelligent university assistant prototype, based on Large Language Models (LLMs) and the Retrieval- Augmented Generation (RAG) technique. Contextualizing itself with institutional academic documents, the solution was implemented in Python, using the Transformers library to orches- trate the data flow, the FAISS library for creating a vector database of official documents, and a compact language model (Llama-3.2-1B-Instruct) for execution in a local environment. The user interface was built with Streamlit. In functional tests, although the RAG architecture was successfully implemented, the quality of the responses generated by the assistant proved unsatisfactory, with generic, evasive, and disconnected information. It was concluded that the results were inconclusive due to severe hardware limitations, which forced the use of a language model with low context interpretation capabilities. The study validates the proposed architecture but highlights that the effectiveness of systems like this is directly dependent on the computational capacity to support more robust LLMs.