Case study
AlTahqeeq
Source-grounded AI answer engine for authenticated Islamic Q&A that cites every claim to its original source.
Overview
AlTahqeeq is a source-grounded answer engine for Islamic questions. Every answer is assembled only from an authenticated, citation-linked knowledge base of hadiths, and each claim links back to the source it came from, so readers can verify rather than trust. Built end to end as CTO & Lead Developer: product direction, system design, the knowledge base, the answering pipeline, and the responsive web application.
Problem
General-purpose chatbots answer religious questions fluently but cannot be trusted: they paraphrase loosely, blend sources, and invent citations. For Islamic Q&A, where the authenticity of a narration is the whole point, an unsourced or mis-cited answer is worse than no answer.
Solution
Answers are written only from verified material. A question is first matched against the curated corpus, the response is then composed strictly from the passages found, and every statement carries a reference to its source. Nothing is written that cannot be pointed back to a citation. Validation measured a sub-3% error rate with zero fabricated citations.
Architecture
- Sources → authenticated hadith collections are cleaned, structured and stored with their references intact.
- Matching → each question is matched against the knowledge base to find the passages that actually address it, including questions asked in Arabic or transliterated forms.
- Composition → the answer is written only from the matched passages, with each claim tied to its citation.
- Verification → every citation is checked against the corpus before it is shown; questions with no supporting source are declined instead of guessed.
- Delivery → a responsive web application for asking, reading, and following citations through to the original text.
Tech stack
Key engineering decisions
- • Verify first, write second: the system may only restate what the corpus supports, which is what makes zero fabricated citations achievable.
- • Citations as a first-class part of the data: every passage carries its source reference through the whole pipeline, so the interface links each claim without guesswork.
- • Error rate, not fluency, as the product metric: an answer that reads well but mis-cites is a failure.
- • Single owner across data, answering pipeline and interface to keep the citation contract consistent end to end.
Results
- • Sub-3% error rate on validated answers.
- • Zero fabricated citations.
- • Responsive web application and citation-linked hadith knowledge base shipped end to end.
Links
What I’d improve next
- • Publish the evaluation set and grading method so the error rate is reproducible.
- • Broaden coverage to more collections and improve matching for rare terms and transliterations.
- • Add a scholar-review workflow to flag and correct answers in production.