LLM agent with a Naive RAG architecture applied to the SUS Ombudsman's Office: an intelligent automation proposal for screening and analyzing citizen submissions

Abstract

The ombudsman services of the Brazilian Unified Health System (SUS) receive an increasing volume of citizen complaints (complaints, denunciations, suggestions and compliments), the manual analysis of which poses operational challenges to public health management. This work presents the implementation of an agent based on the Large Language Model (LLM), using Naive Retrieval-Augmented Generation (RAG) architecture, to automate the triage, sentiment classification, and response generation of complaints registered with the SUS Ombudsman. The solution employs the Llama 3 model through the Ollama API, multilingual vectorization with the paraphrase-multilingual-MiniLM-L12-v2 model, information retrieval by cosine similarity, and an interactive interface developed in Gradio, running in a Google Colab environment. The study describes the components of the implemented architecture and discusses possibilities for evolution towards an Advanced RAG approach, incorporating orchestration via LangChain, vector persistence in PostgreSQL with pgvector, and reranking by Cross-Encoder. This initiative aligns with recent efforts to incorporate artificial intelligence into the SUS (Brazilian Unified Health System) ombudsman services, reinforces the potential of these technologies to increase the efficiency of listening and service processes in public health.