FARSIQA: Faithful & Advanced RAG System for Islamic Question Answering

ACL ARR 2026 January Submission6701 Authors

Keywords: Retrieval-Augmented Generation, Question Answering, Persian NLP, Low-Resource Domains, Iterative Refinement, Hallucination Mitigation, Islamic Domain

Abstract: While Large Language Models (LLMs) have revolutionized NLP, their application in high-stakes domains like religious question answering remains challenged by hallucinations and lack of faithfulness, particularly for the Persian-speaking Muslim community. Existing Retrieval-Augmented Generation (RAG) systems often struggle with complex, multi-hop queries requiring rigorous evidence aggregation. To address this, we introduce FARSIQA, a novel end-to-end system for the Persian Islamic domain based on the FAIR-RAG architecture. Unlike conventional single-pass pipelines, FAIR-RAG employs a faithful, adaptive, and iterative refinement framework that dynamically decomposes queries and self-corrects retrieval gaps to ensure comprehensive context generation. Leveraging a curated knowledge base of over one million authoritative documents, FARSIQA achieves state-of-the-art performance on the IslamicPCQA benchmark. Notably, it attains a 97.0% Negative Rejection rate-a 40-point improvement over baselines-demonstrating exceptional safety in handling out-of-scope queries, alongside a high Answer Correctness score of 74.3%. Our work establishes a new standard for Persian Islamic QA, validating the critical role of adaptive RAG architectures in building reliable AI for sensitive domains.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: Question Answering, Information Retrieval, Low-resource Methods

Contribution Types: NLP engineering experiment, Approaches to low-resource settings, Data resources

Languages Studied: Persian

Submission Number: 6701