Expert Multi-Agent Conversational System Using Retrieval-Augmented Generation and Dynamic Text-to-SQL for Government Transparency

Flores N.; Ramirez C.; Mauricio D. · IEEE Access (2025) · DOI: 10.1109/access.2025.3635530

Resumen

Access to structured and actionable public information remains a major challenge for government transparency in Peru, where official data are often fragmented, inconsistently published, and presented in low-quality formats. This study introduces a multi-agent conversational system designed under the Design Science Research paradigm, which is instantiated through three artifacts. First artifact is a data processing system that integrates procurement and legislative records through classification, region detection, and OCR. A second artifact is a set of specialized agents, combining Retrieval-Augmented Generation for attendance and voting with a Text-to-SQL agent for procurement. The third artifact is a multi-agent orchestration system that unifies these components within a microservices architecture. Together, they transform raw governmental records into structured, auditable, and citizen-accessible responses. The pipeline achieved strong results: classifiers reached accuracies up to 0.989, YOLOv11 obtained mAP@50–95 of 0.965 for attendance and 0.886 for voting, and Tesseract delivered a character error rate of 0.031 and word error rate of 0.078. The RAG agents achieved MAP of 0.950 for attendance and 0.926 for voting, confirming robust retrieval even under noisy conditions. For procurement queries, the Text-to-SQL agent outperformed baselines with FLEX of 0.909 and EX of 0.801. End-to-end evaluation confirmed reliability with Faithfulness of 0.942, Answer Relevancy of 0.907, Answer Similarity of 0.984, and Answer Correctness of 0.891, complemented by BERTScore, BLEU, and ROUGE metrics. These results validate the feasibility of large language models and microservice-based architectures to enhance transparency and accountability in low-data-quality environments.

Instituciones: Universidad Nacional Mayor de San Marcos; Universidad Nacional de Ingeniería