Abstract

The analysis of unstructured civil petitions is often hindered by procedural noise and verbose argumentation. To address this, we propose a pipeline composed of LLM-based fact extraction followed by legal-domain embeddings of texts for unsupervised density clustering. We employ Large Language Models to isolate factual narratives from raw texts, which are then encoded using domain-specific representations (Legal-BERT) and grouped via UMAP dimensionality reduction and the HDBSCAN algorithm. Comparative experiments on a Brazilian judicial corpus reveal that clustering based solely on extracted yields significantly more cohesive and semantically well-defined groups than, which suffer from fragmentation due to content variability. Results indicate that the proposed method is a promising approach for thematic organization, procedural triage support, and large-scale discovery of legal patterns.