Enhancing Legal Document Understanding and Analysis Using Retriever-Augmented Generation (RAG)

This paper presents a Retriever-Augmented Generation (RAG) framework for analyzing Arabic legal texts, focusing on Egypt's 2024 Unified Insurance Law. We apply article-level semantic chunking and evaluate three embedding models—AraBERT, multilingual-E5, and Cohere's embed-multilingual-v3.0—using Pinecone for retrieval and a local LLaMA 3.2 model for generation. Experiments show Cohere's model achieved the best performance (MRR@K 0.843, Precision@K 0.356). Our modular pipeline addresses challenges in Arabic legal NLP for low-resource languages.