Retrieving What Is Still True: Temporal-Aware RAG for Evolving Legal Knowledge

Laws and regulations change over time: later acts can amend, replace, or revoke earlier provisions without changing the original document. Retrieval-augmented generation (RAG) systems that search a fixed collection may therefore present outdated rules as current. We present a temporal RAG pipeline that segments resolutions by legal hierarchy, extracts validity-changing events through a lexicon- filtered language model and a deterministic admission gate, and applies them to a versioned repository. Its active view contains only the provisions in force at the selected cutoff date. On 693 resolutions from the Brazilian electricity regulator (ANEEL) and 162 questions, the updated index reaches a mean answer-correctness score of 79.0%, compared with 70.4% for a structurally identical index without updates. On the 122 questions whose governing events are extracted correctly, the scores are 83.2% and 70.1%, respectively (p = 0.003), with the largest gains on amended rules.