Methodology

Source corpus

• Quran: all 6,236 ayat indexed with Arabic + multiple English translations.
• Hadith: the full Kutub al-Sitta (Bukhari, Muslim, Abu Dawud, Tirmidhi, Nasa'i, Ibn Majah), Muwatta Malik, Musnad Ahmad — 46,000+ narrations, 37,000+ of them with grades from al-Albani, Shu'ayb al-Arna'ut, and Darussalam; disputed gradings are flagged as disputed.
• Tafsir: Tabari, Ibn Kathir, Qurtubi, Razi, Sa'di, Jalalayn, Ibn 'Ashur, Alusi, and more.
• Fiqh: classical works from all four madhahib, including the commentaries and glosses (shuruh and hawashi) where madhhab positions are actually argued, not only the base texts — Hanafi (Sarakhsi, Kasani, Marghinani, Ibn Abidin), Maliki (Khalil, Dusuqi, Ibn Rushd), Shafi'i (Nawawi, Ramli, Haytami), Hanbali (Ibn Qudamah, Buhuti).
• Rijal & biographies: Tahdhib al-Kamal (Mizzi), Siyar A'lam al-Nubala (Dhahabi), Tahdhib al-Tahdhib (Ibn Hajar).
• Seerah: Ibn Hisham, Zad al-Ma'ad (Ibn al-Qayyim).

Total: 440,000+ embedded passages from 229 distinct works (about 1,500 volumes). That is a curated selection, not a ceiling: the full turath.io library of 8,583 works is held locally, and extending the corpus is an embedding decision, not an acquisition one.

Retrieval pipeline

1. Query understanding: an LLM decomposes complex questions into focused sub-queries with source-type filters (Quran, hadith, fiqh, tafsir).
2. Bilingual vector search: Qwen3-Embedding-8B vectors retrieve semantically related passages in both Arabic and English.
3. Keyword search: in parallel, exact-match lookups for narrator names, hadith numbers, and technical terms.
4. Knowledge graph expansion: 58K nodes covering hadith chains, scholar networks, and book metadata add neighbors the embedding wouldn't find.
5. Reranking: Qwen3-Reranker-8B reorders the top candidates by claim-relevance.
6. Evidence assessment: a second LLM pass identifies remaining gaps and issues follow-up searches.

Source trust hierarchy

• Tier 1: Quran, Sahih Bukhari, Sahih Muslim.
• Tier 2: rest of Kutub al-Sitta, Muwatta, Musnad Ahmad.
• Tier 3: tafsir, classical fiqh, aqeedah, biographical works.
• Tier 4: contemporary scholarly works, fatwa collections.

Higher-tier sources are surfaced first and cited preferentially.

What the tiers do not do.They rank authenticity of transmission and set display order; they do not decide juristic weight. A madhhab's position often rests on a hasan narration in the Sunan, an athar, or qiyas rather than on a report in the Sahihayn, and ranking by hadith grade is not ranking by relevance to the mas'ala. So in fiqh questions the model is instructed to present the evidence each school actually relies on, with its grade noted, rather than replacing it with a higher-tier narration on a different point or dropping the position.

When graders disagree.Hadith carry verdicts from several scholars where available. The badge shows one primary grade, chosen in a fixed order (al-Albani, then Shu'ayb al-Arna'ut, then the collection's own editor), and when the graders differ the badge is marked as disputed and expands to show every verdict with its scholar, so the reader sees the disagreement rather than a silently chosen side.

Anti-hallucination guardrails

• Strict source grounding: every factual claim must include a [Source N] citation tied to a retrieved passage.
• Citation verification, layer 1 (existence): post-generation pass that fails loud on any [Source N] that doesn't resolve to a real retrieved chunk. This catches fabricated references.
• Citation verification, layer 2 (support): each claim is then checked against the passage it cites — first by lexical grounding over the full passage, then by scoring the claim against the best-matching window of that passage with the reranker. A claim that fails both is listed in a visible notice under the answer. This targets the harder failure, a real passage cited for a sentence it does not support. It is not a formal entailment model, and it is skipped when the claim and the passage are in different languages (an English sentence citing an Arabic source), so today it is mostly an Arabic-answer guarantee. A cross-lingual entailment check is the planned fix.
• Quran text canonicalization: quoted ayat are matched against a complete mushaf shipped with the service (all 6,236 ayat) using high-threshold sequence similarity; a quote that differs is replaced by the canonical text and the correction is flagged to the reader. A quote that cannot be matched at all is marked unverified rather than passed through.
• Hadith grade preservation: weak (da'if) and fabricated (mawdu') hadith are always flagged with a visible warning.

Evaluation

Measured on IslamicFaithQA (QCRI) under the Fanar SimpleQA protocol: a 500-question sample per language, GPT-4.1 as judge, binary correct / incorrect / not attempted. Run July 2026.

• English: 74.5% correct (372 of 499 graded). Arabic: 73.9% (369 of 499). 95% confidence interval about ±3.8 points.
• Same model with retrieval disabled, identical 500 questions and judge: 55.1%. The 19.4-point gap is what the retrieval, terminology and verification layers add.
• Weakest category: inheritance law at 56.5%, which is about a quarter of the benchmark. Qur'anic studies, creed and jurisprudence score 80–93%.

Read this precisely. Fanar-Sadiq's published score on this same benchmark is 65.4%, but its published average across its five benchmarks is 76.5%, higher than the number above. ShAIkhy has been evaluated on only one of those five, so the suites are not comparable; on the one benchmark both have run, ShAIkhy is ahead.

Not yet measured. The benchmark grades final answers, not citations. Whether each cited passage supports the exact sentence attached to it has not been evaluated as a separate metric. A citation-fidelity set (contested questions, every citation opened and checked against its sentence) is the next evaluation planned, and its results will be published here.

Theological scope

ShAIkhy operates within Ahl as-Sunnah wa al-Jamaa (Sunni Islam) across the four madhahib (Hanafi, Maliki, Shafi'i, Hanbali) and the three Sunni aqeedah schools (Athari, Ash'ari, Maturidi). Non-Sunni positions surfaced from the corpus are labeled explicitly.

ShAIkhy does NOT issue takfir, declare individuals or sects deviant, or rule on personal life situations. For those, consult a qualified scholar who knows your full context.