Design and Development of a Contextual Search System for Qur’anic Text Based on Large Language Models (LLMs) or Artificial Intelligence (AI)

Abstract

This study addresses the challenge of providing reliable, safe, and contextually accurate access to the Holy Qur’an texts using Large Language Models (LLMs). While traditional approaches often rely on Retrieval-Augmented Generation (RAG) for factual grounding, this research proposes and evaluates a novel, non RAG architectural approach centered on advanced, multi-layered GuardRail 
Prompting to manage the inherent risks of LLM stochasticity and hallucination in sensitive religious domains. The system integrates Input Guardrails for prompt injection mitigation, and critical Output Guardrails utilizing an LLM-as-a-Judge framework and a re-generation loop to validate responses against semantic relevance, Shariah compliance, and structural integrity (JSON Schema). The research adopts a Research and Development (R&D) methodology combined with a computational experimental approach to evaluate system performance, moderation effectiveness, and the functional role of rule-based control in regulating generative model outputs. Results demonstrate that a well-engineered GuardRail architecture can effectively constrain LLM behavior, achieving high 
faithfulness and relevance, with low PGR and acceptable FPR across adversarial and benign query datasets. This research establishes GuardRail Prompting as a viable and robust alternative for contextual grounding in sensitive, knowledge intensive applications where RAG deployment may be restricted or structurally undesirable.