Chat institucional com inteligência artificial para acesso a informações oficiais do IFBA

Abstract

This work addresses the difficulty of accessing and interpreting regulatory documents on the IFBA institutional portal, characterized by a complex information architecture that creates usability barriers and risks of misinformation. It proposes the development of a virtual assistant based on Generative Artificial Intelligence and Retrieval-Augmented Generation (RAG) architecture to optimize the search and synthesis of notices, ordinances, and regulations. The methodology consists of applied and exploratory research using the Evolutionary Prototyping model for the creation of a Minimum Viable Product (MVP). The system utilizes the n8n platform for orchestration, the Supabase vector database with the pgvector extension, and OpenAI models GPT-5-Nano, gpt-4o-mini, and text-embedding-3-small. The process involves the ingestion, sanitization, and semantic fragmentation of a documentary corpus focused on IFBA. The results achieve an accuracy of 94.6% in controlled tests and a 71.43% success rate in real user queries, with 100% citation of sources in technical assistance. The system obtained an average score of 4.0 (on a scale of 1.0 to 5.0) for ease of use and maintained logical consistency in 77.78% of interactions, although 55% of users reported processing latency. The study contributes to the democratization of access to institutional information and the strengthening of administrative transparency at IFBA.