Orbis Juris AI: Intent-Aware Retrieval and Precedent Assessment for Brazilian Case Law

When a client presents a legal dispute, the practitioner must find controlling precedents, assess their normative force, and determine how strongly each applies. Current legal search systems typically rank similar decisions, but they do not explicitly model the user's research objective or the relative authority of retrieved precedents. We present Orbis Juris AI, a system for retrieving and assessing judicial precedents from pre-litigation narratives, including both client statements and practitioner intake memos. The system has three main components: (1) a query-understanding component that reformulates the narrative, identifies the user's research intent, and adjusts subsequent retrieval accordingly; (2) a retrieval component that combines hybrid search, cross-encoder reranking, authority signals, procedural filtering, and intent-specific scoring; and (3) an assessment component that identifies operative holdings, validates extracted evidence, assigns confidence labels, and supports constrained synthesis using rule-based safeguards. In a stage-level evaluation over 100 stratified queries and 1,138 annotated pairs from a corpus of 576,000 Brazilian appellate decisions, the retrieval stages achieve nDCG@10 of 0.862 ($+$24.9~pp over baseline, $p < 0.001$) and MRR of 0.963. The assessment stage's confidence labels show monotonic calibration against graded relevance annotations ($\rho = 0.61$, $p < 0.001$), with 88.5\% of high-confidence recommendations rated highly relevant compared with 22.3\% for low-confidence items. These results suggest that the system can reduce manual effort involved in precedent search and preliminary case assessment, while improving retrieval quality and providing interpretable confidence signals.