Retrieval-Augmented Generation (RAG) systems are now considered the foundation of modern Agentic AI solutions; it closes the gap between knowledge retrieval and generative reasoning. However, their total effectiveness critically depends on the quality and linguistic adaptability of the retrieval layer. Information retrieval is still considered a difficult task due to the Arabic language’s rich and deep morphology and the difficult domain-specific characteristics of the Arabic legal terminology. This study explores different retrieval models for Arabic legal texts, while focusing on the combination of morphological analysis (Farasa) and semantic embedding representations. We constructed a synthetic benchmark dataset made up of 500 legal articles and 1,000 question–answer pairs, just to support reproducible evaluation and to address the scarcity of publicly available legal datasets. We used answers generated by the GPT-4.1-mini API, to act as the ground truth for model assessment. We evaluate traditional lexical retrieval models such as BM25 against advanced embedding based models, including Ada v3, BGE M3, and GTE, with and without morphological preprocessing. The inclusion of morphological analysis substantially enhances BM25’s performance, increasing the Mean Average Precision (MAP) from 0.7178 to 0.7715. Embedding-based models also performed well, with Mistral-embed achieving the highest overall scores among them (MAP = 0.7570 and nDCG@10 = 0.7964), slightly outperforming GTE and BGE-M3. Moreover, a hybrid configuration combining BM25 (Farasa) with Ada v3 embeddings yields improved results (MAP = 0.8304; nDCG@10 = 0.8626). These findings highlight the importance of morphology-aware, semantically grounded retrieval in the construction of robust Arabic legal RAG pipelines, while the introduced synthetic dataset provides a valuable resource for future benchmarking and reproducible research in Arabic legal information retrieval and Agentic AI.
AboAsal et al. (Thu,) studied this question.