DivinityAI – Islamic Grounded RAG: Hallucination-Free Quran & Hadith Search

In the realm of religious scholarship, textual accuracy isn’t just a preference—it’s a sacred obligation. General-purpose large language models (LLMs) frequently hallucinate Islamic religious texts, fabricating Quranic verses, misattributing Hadith narrations, or generating incorrect jurisprudential rulings. DivinityAI – Islamic Grounded RAG solves this critical problem through a Retrieval-Augmented Generation (RAG) system with an uncompromising corpus-lock policy.

The Challenge: Why DivinityAI Was Built

General-purpose LLMs exhibit three critical failures when handling Islamic texts:

• Verse Fabrication: Inventing non-existent Quranic surah/ayah combinations
• Hadith Corruption: Merging narrations or misattributing chains of transmission
• Jurisprudential Errors: Generating fatwas unsupported by classical scholarship

DivinityAI addresses these through architectural enforcement of:

1. Source Lock - Only King Fahd Uthmani Quran and six canonical Hadith collections
2. Verification Chains - Every citation validated before response generation
3. Boundary Policies - Automatic rejection of unanswerable or off-domain queries

Core Architecture & Technical Stack

Pipeline Design

The system implements a twelve-stage processing pipeline:

1. Intent classification via Gemini 2.5 Flash
2. Scope enforcement with confidence thresholding
3. Query rewriting (HyDE + sub-query decomposition)
4. Parallel BM25 (sparse) and BGE-M3 (dense) retrieval
5. Reciprocal Rank Fusion to merge result sets
6. Four-tier citation verification
7. Evidence sufficiency evaluation
8. Grounded generation with context-only prompting
9. Fatwa boundary detection
10. Hallucination verification
11. Disclaimer injection
12. Structured response assembly

Component Stack

Layer	Technology	Purpose
Frontend	React 19 + Vite + Tailwind CSS v4	RTL-optimized Arabic interface
Backend	Django ASGI + DRF	Query orchestration
Vector DB	ChromaDB	Quran/Hadith embeddings
Sparse Search	BM25	Exact token matching
Dense Search	BGE-M3	Semantic retrieval
LLM Routing	OpenRouter (Gemini 2.5 Flash)	Primary generation
Validation	Groq (Llama 3.3 70B)	Semantic verification

Key Features Breakdown

Deterministic Citation Verification

Implements a cascading validation chain:

1. Exact string match against source texts
2. Normalized matching (diacritic-insensitive)
3. Levenshtein distance threshold (85%+)
4. Semantic equivalence confirmation via Llama 3.3

Hybrid Retrieval System

Combines strengths of two search methodologies:

• BM25 Sparse: Precise matching on normalized Arabic text
• BGE-M3 Dense: Cross-lingual semantic understanding
• RRF Fusion: Blends rankings from both approaches

Fatwa Boundary Detection

Rule-based triggers append disclaimers when detecting:

- Inheritance calculations (Mirath)
- Medical ethics questions
- Marriage/divorce rulings
- Any potentially evolving jurisprudential matter

Real-World Use Cases

• Scholarly Research: Authenticated reference discovery without contamination risk
• Comparative Analysis: Cross-source verification across Hadith collections
• Architectural Pattern: Reference design for high-stakes RAG implementations (legal/medical)

How It Works: Step-by-Step

Example Query: “What did the Prophet say about kindness to parents?”

1. Intent router classifies as hadith with 92% confidence
2. HyDE generates hypothetical authentic response pattern
3. BM25 finds matches for “kindness parents” in Arabic text
4. BGE-M3 retrieves semantically similar narrations
5. RRF merges and re-ranks top 10 candidates
6. Verifier confirms Sahih Bukhari 5971 matches exactly
7. Llama 3.3 confirms evidence sufficiency
8. Gemini generates response strictly from verified chunks
9. Safety layer adds “consult scholar” disclaimer

Comparison: DivinityAI vs Traditional Approaches

Criteria	General LLM	DivinityAI
Citation Accuracy	≈40-60%	>95%
Hallucination Rate	15-25%	<1%
Jurisprudential Safety	None	Boundary checks
Query Rejection	Rare	Confidence-gated

Frequently Asked Questions (FAQ)

1. What makes DivinityAI different from ChatGPT for Islamic questions?

DivinityAI implements architectural enforcement of zero hallucination through its corpus-lock policy and deterministic verification chains, whereas ChatGPT has no mechanisms to prevent fabrication of religious texts.

2. How does the system handle differences between Madhahib (schools of thought)?

It surfaces authentic source texts with clear attribution but avoids synthesizing new rulings. For fiqh questions, it retrieves and presents relevant positions from primary sources with appropriate contextual disclaimers.

3. Can I contribute additional Hadith collections to the corpus?

The current implementation intentionally limits sources to the most universally authenticated collections to maintain verifiability standards. Expansion would require rigorous scholarly validation.

4. What languages does the system support?

Query input accepts Arabic, English, and Malay, while responses maintain original Arabic script with parallel translations when available.

Conclusion & Next Steps

DivinityAI represents a significant advancement in domain-specific RAG architectures, demonstrating how to achieve verifiable accuracy in high-stakes knowledge domains. The system’s strict verification chains and boundary enforcement provide a template for applications in legal, medical, and other precision-critical fields.

Experience the system firsthand at: https://muslim.nevatal.tech

For technical teams, the architectural patterns demonstrated—particularly the hybrid retrieval system and deterministic verification approaches—offer immediately applicable insights for building reliable AI systems.