Abstract

This paper introduces a Graph Retrieval-Augmented Generation (GraphRAG) pipeline tailored for Question Answering (Q A) within Portuguese legal documents. Applied to a corpus of 203 normative resolutions from Companhia Energética de Minas Gerais (CEMIG), the proposed approach addresses the structural complexity of legal texts, such as hierarchical dependencies and temporal modifications. By explicitly modeling documents as knowledge graphs with nodes representing structural units (Articles, Paragraphs, Items) and edges denoting normative relationships, the system preserves context and traceability. The retrieval mechanism reconstructs evidence paths from root to leaf, performing semantic re-ranking before generation. Evaluation using the RAGAS framework yielded a mean answer accuracy of 0.81, with a median of 1.00. Results indicate that the system performs robustly on short, focused queries, while intermediate-length questions present challenges related to semantic dispersion. The findings suggest that structurally aware retrieval significantly enhances the interpretability and precision of legal Q A systems.