A Multi-Stage AI Framework for Legal Case Retrieval (PGDip Defence)

In common-law and mixed legal systems, earlier judicial decisions form the backbone of legal reasoning. The doctrine of stare decisis obliges courts to follow higher-court rulings on the same point of law, while lower-court or foreign judgments may be cited as persuasive authority. A persuasive brief therefore does three things: it highlights factual similarities, shows that the governing legal issue was previously resolved the same way, and demonstrates that the earlier outcome supports the advocate’s position. Finding and ranking useful precedents is harder than it first appears. Courts sit at multiple hierarchical levels—magistrate, trial, appellate and supreme—and issue decisions across a century or more. These opinions are written in different styles, use changing terminology, and may cite statutes that have since been amended. Additionally, modern legal practice crosses borders: counsel in a Sri-Lankan negligence case may cite English or Australian authorities, while a Chinese commercial dispute can involve contract concepts borrowed from German civil law. The result is a sprawling multilingual archive in which the “right” ruling might reside thousands of pages—and several jurisdictions—away from the obvious keywords. Missing a controlling precedent can undermine an argument, delay proceedings, and in extreme cases constitute professional negligence. Conversely, uncovering a closely aligned case early in the research process saves hours of analysis and can even prompt settlement. Manual methods cannot keep pace with the volume of material: major databases now index well over ten million judgments, with thousands more added each month. As a result, legal professionals increasingly rely on automated retrieval systems to sift, rank and explain potential precedents. An effective system must navigate court hierarchies, recognise evolving legal terminology, and surface the passages that truly make a prior decision relevant—tasks that go beyond straightforward keyword matching. The challenge of meeting those needs forms the starting point for the research presented in this thesis.