Exécution de notre outil de synthèse sur un nouvel article

Résumé : The rapid evolution of generative artificial intelligence has introduced new possibilities for legal document understanding, yet Arabic legal contexts remain largely unaddressed. Existing legal AI systems are developed exclusively for English-language corpora, leaving the unique morphological complexity, domain-specific terminology, and right-to-left processing challenges of Arabic law unresolved. This paper presents Lawsuit AraRAG, a Retrieval-Augmented Generation system specifically designed for Arabic legal document analysis using Saudi Arabian court documents. The system integrates dense vector retrieval via the Multilingual-E5-base embedding model with a locally hosted Qwen2.5-14B generator, incorporating a query rewriting module and a deliberate Arabic-aware chunking strategy to preserve semantic integrity across morphologically complex legal clauses. The system is evaluated on 40 manually constructed legal queries spanning civil procedure, tax violations, and document authenticity disputes. Experimental results show that Lawsuit AraRAG achieves a Precision@3 of 68% and an F1-Score of 70%, reducing hallucination rates from 31% to 12% compared to a baseline LLM without retrieval. Qualitative case studies further reveal system robustness against out-of-distribution queries, while exposing cross-domain semantic disambiguation as the primary failure mode for Arabic legal RAG systems. The paper concludes by proposing hybrid retrieval and metadata filtering strategies, and outlines a path toward an agentic legal reasoning architecture for future deployment.