From RAG to Agentic RAG for Faithful Islamic Question Answering

The article introduces IslamicFaithQA, a new generative benchmark for Islamic question answering consisting of 3,810 bilingual items designed to measure hallucination and abstention in responses. It presents an agentic Quran-grounding framework (agentic RAG) that incorporates structured tool calls for iterative evidence seeking, demonstrating significant performance improvements over standard RAG, particularly with the Qwen3 4B model. This work is crucial for practitioners as it provides a robust evaluation framework and resources that enhance the reliability of LLMs in sensitive applications like religious question answering.