Desenvolvimento de um chatbot usando large language model para suporte técnico interativo a apicultores

Abstract

Beekeeping in the Brazilian semi-arid region, specifically in the Central Backlands of Ceará, faces significant challenges regarding the scarcity of on-site technical assistance and difficulty in accessing specialized information. This work presents the development of an intelligent virtual assistant integrated into the “Revise!” mobile application, designed to provide technical support to small producers. The solution employs the RAG (Retrieval-Augmented Generation) architecture to align the generative power of Large Language Models (LLMs) with the technical precision of manuals from Embrapa and SEBRAE. The methodology encompassed the structuring of a vector knowledge base, the implementation of a FastAPI backend, and field validation with beekeepers. Comparative tests demonstrated that the BERT-PT embedding model outperformed generic multilingual models, reducing hallucinations and ensuring greater terminological adherence to the regional context. The results indicate that the system is capable of diagnosing sanitary and management problems with high precision. Final qualitative validation revealed that, although there is a demand for voice interfaces, silent usage (text) is preferable during apiary management to avoid stressing the swarm. It is concluded that the tool enables democratized access to technical knowledge, mitigating the rural extension gap in the region.