E-Prosiding Persidangan Serantau Manasik Haji 2025 (MANASIK2025)

MANASIK2025 Putrajaya, Malaysia, 7-8 Oktober 2025 633 timeliness — as current studies suggest — RAG can surface outdated or inapplicable texts. Hence, technical advances must be paired with content governance (shared schemas, publication SLAs, and release notes) and audit mechanisms (Raji et al., 2020). Thematic synthesis of recent studies (Malaysia & beyond) 1) Standardization and authority alignment (2012 – 2021) Scholars document cross-state variability and its social costs (confusion, forum-shopping), arguing for harmonized dissemination while respecting state sovereignty (Zaini et al., 2012) (Mahyudin et al., 2018). Administrative sources reiterate that national decisions are advisory unless adopted by states-an essential constraint for any automated system. Critical take. The literature specifies who is authoritative and where , but it stops short of data- engineering prescriptions (e.g., authority tags, versioning, jurisdiction filters). This is a gap a technical framework can fill. 2) Digitization of Hajj guidance and recurrent scenarios (2010s – present) TH’s accumulated Muzakarah resolutions constitute a scenario -rich dataset (e.g., illness-based concessions, sequencing errors), ideal for structured retrieval and case-based reasoning (Lembaga Tabung Haji, n.d.). Saudi digitization via Nusuk shows how integrated platforms can reduce friction for pilgrims and centralize official content (Saudi Ministry of Hajj and Umrah, n.d.). Critical take. The materials are “narrative - heavy” PDFs/webpages with limited machine -readable structure. Without upstream curation (IDs, headings, controlled vocabularies), downstream AI systems will struggle to ensure coverage and currency. 3) AI reliability, audits, and guardrails (2020 – 2024) NLP research describes hallucination taxonomies and mitigation strategies; RAG is a leading approach to ground responses in cited documents (Ji et al., 2022, 2024; Lewis et al., 2020). Governance work outlines audit documentation across the lifecycle — useful blueprints for fatwa institutions that wish to certify automated outputs (Raji et al., 2020). also warns that even grounded systems face prompt- injection and other adversarial risks when connected to live web sources (Financial Times, 2025). Critical take. These works are strong on generic techniques but silent on jurisprudential semantics (e.g., how to encode rukhsah conditions; how to represent conflict between a national advisory and a state- gazetted ruling). Identified gaps the present article addresses 1. Provenance-first fatwa automation. There is no comprehensive framework that binds every answer to a specific, citable sighah (or TH resolution) and displays its authority level (advisory vs. gazetted; state vs. national). The article proposes a RAG pipeline with mandatory citation and an interface that foregrounds jurisdiction, date, and status — closing a recurrent ambiguity highlighted in governance studies (Zaini et al., 2012)(Jabatan Mufti Selangor, n.d.). 2. Jurisdiction-aware personalization. Existing chatbots seldom segment by madhhab or state authority. For Malaysian Hajj/ʿUmrah, answers must default to Shāfiʿī positions as gazetted in the user’s state , or flag divergences and escalate. This operationalizes the devolved structure documented in official sources (JAKIM i-FIQH 2.0, n.d.). 3. Safety cases and audits for religious QA. Few deployments publish audit trails. The article defines measurable safeguards (e.g., “no - answer” thresholds, automatic escalation, incident