This paper presents the Structure-Aware Temporal Graph RAG (SAT-Graph RAG) framework to address the Limitations challenge of Retrieval-Augmented Generation (RAG) systems in the legal domain. The simple text retrieval approach of existing RAG systems fails to account for the hierarchical, temporal, and causal structures of law, resulting in outdated and unreliable answers. SAT-Graph RAG addresses these challenges by explicitly modeling the formal structure and temporal characteristics of legal norms. Based on a formal knowledge graph inspired by the LRMoo model, it distinguishes between abstract legal documents and their versioned representations, efficiently aggregates temporal states, and reuses versioned representations of unchanged components. Furthermore, it embodies legislative events as primary action nodes, making causal relationships explicit and queryable. This structural foundation allows for the deterministic resolution of complex requests, such as (i) point-in-time retrieval, (ii) hierarchical impact analysis, and (iii) auditable provenance reconstruction, using a planner-guided, integrated query strategy. A case study of the Brazilian Constitution demonstrates that this approach provides a verifiable and time-accurate LLM foundation, enabling high-level analytical capabilities and significantly reducing the risk of factual errors. Consequently, it provides a practical framework for building more reliable and explainable legal AI systems.
