Arabic Retrieval-Augmented Generation (RAG) systems face significant challenges, especially beyond hundreds of millions of words. Persistent issues include low retrieval and generation accuracy for queries involving person names, dates, religious discourse, and Arabic poetry. Additionally, high redundancy from near-duplicate content reduces accuracy and slows retrieval. This paper introduces an Arabic-centric RAG pipeline designed to address these limitations. Evaluated on a set of 575 questions covering diverse Arabic query types, the proposed system demonstrates strong performance. The best-performing retriever, multilingual-e5-large, achieved an Acc@1 of 58.7% and an MRR of 0.715. Meanwhile, the top large language model (LLM), Gemini-2-Flash, reached an answer relevance of 0.83, a context utilization rate of 0.84, and a faithfulness score of 0.94.
Alothman et al. (Thu,) studied this question.