Legal nuggets boost retrieval accuracy by 2x on Brazilian case law

Legal retrieval over collections of court decisions is notoriously difficult because judgments are long, heterogeneous texts where the core legal thesis may span only a few sentences. In a new paper titled 'Legal Nugget Extraction for Granular Retrieval over Long Jurisprudential Texts,' researchers Lucas Pereira, Erick Brito, Roberto Lotufo, and Jayr Pereira tackle this problem by introducing the concept of 'legal nuggets' — short, self-contained legal theses extracted from each document. Their pipeline first extracts these nuggets from every source text, then indexes them with dense embeddings. At query time, it retrieves the most relevant nuggets and aggregates them back into a document-level ranking. The approach is evaluated on four Portuguese legal retrieval benchmarks from the JUA ecosystem.

The results are striking on two jurisprudence-focused datasets. On JUA-Juris, nugget retrieval increases NDCG@10 from 0.10265 to 0.20461 — nearly doubling performance. On JurisTCU, NDCG@10 rises from 0.20898 to 0.32696, a 56% improvement. However, the method underperforms standard full-document retrieval on NormasTCU and BR-TaxQA, which involve legal norms and tax law rather than court decisions. An embedding-model ablation reveals that strong domain-adapted retrievers can still remain superior in the full-document setting. The authors conclude that legal nuggets are particularly useful for jurisprudence search — especially when queries are formulated as legal theses — but may not transfer equally well to other legal retrieval scenarios. This work opens the door to more granular legal search tools, with potential applications in legal research, case law analysis, and AI-assisted document review.