Fanar-Sadiq: A Multi-Agent Architecture for Grounded Islamic QA

Fanar-Sadiq is a bilingual multi-agent system that addresses hallucination and source misattribution in Islamic queries by routing diverse requests to specialized modules for grounded retrieval, exact scripture lookup, and deterministic legal calculations, demonstrating high effectiveness and widespread public adoption.

The Gist

Imagine you are walking into a massive, ancient library that holds the entire history of Islamic knowledge. Inside, there are millions of books, scrolls, and tablets. You ask a question, and you need an answer that is not just "plausible-sounding," but 100% accurate, backed by the original texts, and calculated with perfect precision.

If you ask a standard AI (like a basic chatbot), it's like asking a very well-read but slightly forgetful librarian who tries to guess the answer from memory. Sometimes they get it right, but often they might mix up a quote, invent a rule that doesn't exist, or do the math wrong. In religious matters, these "hallucinations" are dangerous because they can mislead people.

Fanar-Sadiq is a new kind of librarian. Instead of one person trying to remember everything, it's a team of specialized experts working together in a high-tech command center.

Here is how this "Multi-Agent" team works, using simple analogies:

1. The Smart Receptionist (The Router)

When you walk up to the desk and ask a question, the first thing Fanar-Sadiq does is act like a super-smart receptionist.

• The Problem: If you ask, "What is the prayer time in Dubai?" and the receptionist sends you to the "Theology Professor," you'll get a long lecture instead of a time. If you ask, "How much Zakat do I owe?" and they send you to the "Poet," you'll get a beautiful poem instead of a dollar amount.
• The Solution: This receptionist instantly analyzes your question. Is it a greeting? A math problem? A request for a specific verse? A legal ruling?
• The Magic: It has a "menu" of nine different types of questions. Based on your intent, it immediately routes you to the exact specialist who handles that specific job.

2. The Specialized Experts (The Agents)

Once the receptionist sends you to the right room, you meet the specialist. They don't guess; they use tools.

• The Calculator (For Zakat & Inheritance):

  • The Job: Calculating inheritance shares or Zakat (charity) is like solving a complex math puzzle with strict rules. One wrong number changes everything.
  • The Tool: Instead of an AI "guessing" the math, Fanar-Sadiq sends this to a deterministic calculator. Think of this as a high-precision calculator that follows a strict rulebook (like a legal code). It doesn't "think" creatively; it just crunches the numbers exactly as the rules say. If the math is disputed in different schools of thought, it shows you both answers side-by-side, like a judge presenting two legal opinions.

• The Librarian (For Quran & Hadith):

  • The Job: You ask, "What does verse 2:255 say?"
  • The Tool: The AI doesn't try to recite it from memory (which is risky). Instead, it acts like a digital index. It goes straight to the database, finds the exact page, and copies the text word-for-word. It's like using a "Find" function in a book rather than trying to remember the sentence. It also checks the citation to make sure it's the right verse.

• The Astronomer (For Prayer Times & Calendar):

  • The Job: "What time is Fajr in London?" or "What is today's Hijri date?"
  • The Tool: This isn't about reading a book; it's about geometry and astronomy. The system uses a tool that calculates the position of the sun and the moon based on your GPS coordinates. It's like a digital sundial that knows exactly where you are on Earth.

• The Scholar (For Legal Rulings/Fiqh):

  • The Job: "Is it okay to do X?"
  • The Tool: This is the only time the AI "writes" a new answer. But it's a grounded writer. It searches through millions of trusted documents (books of law, history, and scripture). It reads the evidence, summarizes it, and then writes the answer. Crucially, it puts footnotes on every single claim. If it says "Scholar X says this," it links directly to the page where Scholar X said it.

3. The Quality Control (The Verification)

Before the answer ever reaches your screen, a final check happens.

• Did the calculator actually do the math right?
• Did the librarian copy the verse exactly, or did they accidentally change a word?
• Are the footnotes real?
• If the system isn't 100% sure, it admits it. It won't make up an answer. It might say, "I couldn't find a clear rule for this specific situation," which is better than giving a wrong one.

Why Does This Matter?

Think of standard AI as a storyteller who makes up stories that sound true.
Fanar-Sadiq is a forensic investigator who only reports facts found in evidence.

• For the User: You get an answer you can trust. You can see exactly where the information came from (the "source code" of the answer).
• For the Community: It prevents the spread of fake religious rules or incorrect math, which is vital for things like inheritance and charity.

The Result

This system is already being used by nearly 2 million people. It's like having a team of the world's best librarians, mathematicians, and astronomers working together 24/7, ensuring that when you ask about your faith, you get the truth, not a guess.

In short: Fanar-Sadiq stops the AI from "hallucinating" by forcing it to use the right tool for the right job, checking its work, and showing its homework every single time.

Technical Summary

Here is a detailed technical summary of the paper "Fanar-Sadiq: A Multi-Agent Architecture for Grounded Islamic QA".

1. Problem Statement

Large Language Models (LLMs) struggle with religious knowledge queries, particularly in Islamic contexts, due to three critical failure modes:

• Hallucination: Fabricating Quranic verses, Hadith, or legal rulings.
• Misattribution: Incorrectly citing sources or failing to distinguish between scholarly consensus and minority opinions.
• Lack of Determinism: Failing to perform strict arithmetic required for religious obligations (e.g., Zakat, inheritance/Faraid) or misinterpreting canonical texts.

Standard Retrieval-Augmented Generation (RAG) pipelines ("retrieve-then-generate") are insufficient because they treat heterogeneous intents uniformly. A single pipeline cannot simultaneously handle:

1. Verbatim retrieval (e.g., "What does verse 2:255 say?").
2. Rule-constrained computation (e.g., calculating Zakat or inheritance shares based on specific jurisprudential rules).
3. Jurisprudential reasoning (e.g., Fatwa-style answers requiring evidence, conditions, and school-of-thought sensitivity).

2. Methodology: Fanar-Sadiq Architecture

The authors propose Fanar-Sadiq, a bilingual (Arabic/English) multi-agent system designed to route queries to specialized modules based on intent. The architecture consists of four main components:

A. Hybrid Query Classifier (Intent Routing)

The system entry point is a hybrid classifier that predicts the query type and selects the execution route.

• Primary Classifier: An LLM (Fanar) prompted to output a JSON object containing an intent label, confidence score, rationale, and a retrieval flag.
• Fallback Mechanism: If the LLM confidence is low (<0.5) or output is malformed, a prototype-based fallback uses cosine similarity against precomputed intent embeddings (Qwen3-Embedding-4B) to route the query.
• Nine Intent Classes:
  1. Fiqh Rulings (Jurisprudence)
  2. Quran Retrieval (Specific verses)
  3. General Islamic Knowledge
  4. Greetings/Chitchat
  5. Zakat Calculation
  6. Inheritance Calculation
  7. Dua Lookup (Supplications)
  8. Islamic Calendar (Hijri dates/events)
  9. Prayer Times (Qibla and prayer schedules)

B. Specialized Tool Agents

Once routed, queries are handled by deterministic tools or retrieval agents:

• Deterministic Calculators:
  • Zakat Agent: Computes almsgiving based on asset classes, debts, and current precious metal prices using established juristic rules (Nisab thresholds).
  • Inheritance Agent: A Sunni-specific module that calculates estate distribution. Crucially, it handles Madhhab-sensitive branching: in disputed cases, it returns parallel outcomes (e.g., Hanafi vs. Jumhur) rather than a single collapsed ruling. It enforces arithmetic invariants like 'awl (proportional reduction) and radd (return of remainder).
• Symbolic/Geo-Temporal Tools:
  • Islamic Calendar: Converts Gregorian/Hijri dates and lists events using rule-based logic.
  • Prayer Times & Qibla: Computes prayer schedules and Qibla direction using astronomical formulas (great-circle bearing) and geocoding.
  • Dua Lookup: Retrieves authenticated supplications verbatim from a structured store to prevent generative rewriting.
• Retrieval-Augmented Agents:
  • Fiqh & General Knowledge: Uses a dense retriever (Qwen3-Embedding-4B) over 500k+ documents (Quran, Hadith, Fiqh texts). The agent generates answers strictly grounded in retrieved evidence with normalized citation tags (e.g., [C*]).
  • Quranic Retrieval: A dedicated pipeline for specific verses. It includes an NL2SQL module for statistical queries (e.g., "count verses in Surah X") to ensure exact numerical accuracy, avoiding LLM arithmetic errors. It supports fuzzy matching for verse references and returns verbatim text.

C. Response Assembly

A final assembler merges the outputs, attaches a standardized References block with citations and URLs, and logs execution traces for auditability. The system uses a three-tier configuration hierarchy to control generation temperatures and token budgets per module.

3. Key Contributions

1. Multi-Agent Architecture for Heterogeneous Intents: Moves beyond fixed RAG by routing queries to specialized tools (calculators, symbolic engines, and retrievers) based on intent, significantly improving faithfulness.
2. Deterministic & Verifiable Outputs:
   • Implements citation normalization and verification traces for Fiqh answers.
   • Uses exact verse lookup with quotation validation to prevent paraphrase drift.
   • Deploys deterministic calculators for Zakat and Inheritance with school-of-thought branching.
3. Comprehensive Evaluation: The system is evaluated on public benchmarks covering both generative QA (IslamicFaithQA, FatwaQA) and rule-constrained reasoning (QIAS, PalmX, IslamTrust).
4. Real-World Deployment: The system is publicly accessible via API and Web, having processed ~1.9 million queries in under a year with a 77.4% user satisfaction rate.

4. Results

The system was evaluated against strong proprietary baselines (GPT-4.1, GPT-5, Gemini-3-Pro) and open-source models (ALLaM-7B, Fanar-2-27B).

• Overall Performance: Fanar-Sadiq achieved an average accuracy of 76.5%, outperforming open-source baselines and remaining competitive with top proprietary models (Gemini-3-Pro: 79.6%, GPT-5: 77.1%).
• Faithfulness-Critical Tasks:
  • IslamicFaithQA: 65.4% (vs. 56.6% for the best proprietary baseline).
  • FatwaQA: 65.1% (closely tracking Gemini-3-Pro at 67.0%).
  • Significance: The routed architecture significantly reduced hallucinations and improved citation faithfulness in open-ended generation.
• Rule-Based Reasoning:
  • PalmX (Culture): 85.5% accuracy.
  • IslamTrust (Values): 94.2% accuracy.
  • QIAS (Inheritance): 72.2% accuracy. While lower than proprietary models (93-94%), the authors attribute this to the difficulty of mapping deterministic symbolic outputs to discrete multiple-choice options, not a failure in calculation logic.

5. Significance and Impact

• Reliability in Sensitive Domains: Fanar-Sadiq demonstrates that intent-aligned execution is superior to uniform generation for religious QA. By separating computation from generation, it mitigates the risk of fabricating sacred texts or miscalculating legal obligations.
• Transparency: The system enforces traceability through citation tags and audit logs, allowing users to verify the source of every claim.
• Handling Disagreement: Unlike many AI systems that force a single answer, Fanar-Sadiq explicitly surfaces scholarly disagreements (e.g., in inheritance cases), providing parallel outcomes rather than a potentially biased single ruling.
• Scalability: The architecture successfully scales to handle millions of queries while maintaining high accuracy and low latency, proving the viability of agentic, tool-using systems for complex, knowledge-intensive domains.

Limitations: The system does not replace qualified human scholars and cannot issue binding fatwas. Its performance is limited by the coverage of its retrieval corpus and the complexity of mapping symbolic math results to constrained MCQ formats. Future work aims to expand jurisprudential coverage across different schools of thought and improve Hadith verification.