How this is built

You do not need Arabic to check a claim on this site. Here is one walked end to end, in English, using only links.

1. Suppose you read that “mercy” is central to the Quran and want to see for yourself. Start from the English side: the reverse lookup for “mercy” shows every Arabic word that translators render with that English word — and that they are not one word but several, from different roots.
2. Follow the most common one to its root page, ر ح م. Everything there is computed from the corpus: how often the root occurs, in which forms, in which surahs — with every occurrence linked to its verse.
3. Reproduce the count yourself: run the search root:ر ح م. The result states the corpus version and parameters it was computed with; the number on the root page and the number here must agree, and you can watch them do so.
4. See how translators disagree: compare the translations of 1:1, where two words from this one root are rendered differently by every edition on file.

Every number on this site can be interrogated this way. If a page shows a figure you cannot reproduce by clicking through to the search that generated it, that is a bug — report it.

The Arabic is the Tanzil Quran Text (Uthmani) (Uthmani 1.1), published by the Tanzil Project under CC-BY-3.0. It is the immutable source of truth. We never modify Quranic text: normalised, tashkeel-stripped and segmented forms are stored as separate labelled fields, and the source text is always attributed to Tanzil.

Its sha256 is recorded in the manifest, so you can confirm the exact bytes we ingested:

Root, lemma, part-of-speech and sub-word segmentation come from the Quranic Arabic Corpus — Morphology (mustafa0x fork of Leeds QAC v0.4), licensed GPL-2.0-or-later. This is an external annotation. It never alters a token’s id, position or text, and it carries its own copyleft licence — which is why the tokenised tokens.jsonl is distributed under the GPL while the plain text stays CC-BY.

Each token records how the Leeds word aligned onto it, so a questionable annotation is traceable rather than anonymous.

Search runs over a normalised form derived from the Uthmani text by a fixed, disclosed set of rules. Normalisation only ever produces a new labelled field — it never rewrites the source.

1. strip-tashkeel

   Remove Arabic diacritical marks.

2. remove-tatweel

   Remove tatweel (kashida) elongation.

3. unify-alef

   Unify alef variants to bare alef.

4. normalise-ta-marbuta

   Map teh marbuta to heh.

5. normalise-alif-maqsura

   Map alef maksura to yeh.

6. to-simple

   Plain-orthography form: remove Quranic recitation/pause signs and unify alef wasla to bare alef, keeping harakat. Applied to derive token text_simple, which cannot be sourced from the Simple edition because the editions do not tokenise word-for-word.

The basmala is handled as separated: in most surahs it is a separated opening, stored as its own token group that is not a counted verse row. Surah 1 counts it as a verse; surah 9 has none. Because it is a parameter, a computation discloses whether it counts the basmala (basmala=1 or 0) rather than baking one choice in.

Verse numbering is a recorded parameter, not a fact baked into identifiers. The active scheme is kufan (6236 counted verses). Each scheme is a data file; the active scheme’s counted segments are the rows of verses.jsonl. An identifier is scheme-qualified, so it means the same thing regardless of which scheme you view under.

Computed: every count, frequency, form list, root distribution and search result. These are derived deterministically from the loaded corpus and carry the exact parameters that produced them. Re-running the same query on the same version gives the same number.

Asserted from an external source: the morphology (Leeds QAC) and the human translation editions. We attribute these, never present them as our own computation, and never present them as Quranic text.

Asserted by us (editorial): investigations and notes. These are labelled editorial and are the only layer that is our opinion.

Every number on the site links to the query and version that produced it. To reproduce one independently:

1. Download the corpus version from /data and verify its sha256.
2. Re-run the query against the API — for example /api/v1/search?q=root:z-k-w. The response echoes the parsed query, the corpus version and the computation parameters.
3. Compare the total to what the page shows. They must match.

This page is worth more than any feature, so it is blunt:

• Verse slot labels remain Kufan-derived. The counted verse rows follow the Kufan tradition. Other counting traditions exist and are modelled as alternative schemes, but the default slots are Kufan and should not be read as tradition-neutral.
• Translation alignment is verse-level only. The 4 translation editions align to the Arabic by verse identifier, never word-for-word. Word-level alignment data does not exist for these editions and is never fabricated — we do not claim which English word renders which Arabic word.
• Morphology is a third-party annotation. The Leeds corpus is a scholarly reconstruction with its own editorial choices; disagreements with it are legitimate and traceable to the aligned source.
• Token text_simple is computed, not sourced. It is derived from the Uthmani text because the Simple edition does not tokenise word-for-word — see the to-simple normalisation rule above.