Precedent-Aware Multi-Agent Retrieval-Augmented Generation in Case Law Analysis

Retrieval-Augmented Generation (RAG) systems promise practical legal assistance by grounding Large Language Models (LLMs) in external authority. However, standard RAG optimizes semantic similarity and often fails to respect common-law constraints such as jurisdictional bindingness, court hierarchy, temporal validity, and negative treatment. We propose Precedent- Aware Multi-Agent RAG (PA-MA-RAG), an agentic architecture that decomposes legal research and writing into specialized agents for issue framing, authority planning, retrieval, precedent ranking, conflict resolution, drafting, and citation verification. Our method introduces an authority- constrained re-ranking objective that prioritizes controlling precedents while penalizing overruled or otherwise negatively treated cases. The verifier agent enforces evidence-grounded generation by requiring each legal proposition to be supported by retrieved holdings and quotations. We describe an evaluation protocol for both precedent retrieval and citation-grounded legal analysis generation, including authority correctness, supported-claim rate, and robustness to conflicting precedent.