Skip to content
Menu

Arabic search guide

How SeekGist understands Arabic, Arabizi, dialects, units and full requests.

Shoppers in Arabic-speaking markets type in many ways: formal Arabic, dialect, Arabic written in Latin letters (Arabizi), English, or a mix. They make typos and write full requests. SeekGist handles all of this out of the box. You do not need to set anything up.

The examples on this page are real test queries on our public demo catalog (made-up fashion and electronics products). You can try them in the search box on the home page.

1. Spelling forms are made the same

The same rules run on your products and on every query, so different spellings of one word meet in the middle.

WhatExample
Alef forms (أ إ آ ٱ) become اأسود = اسود
ة becomes ه, ى becomes ي, ؤ becomes و, ئ becomes يعباية = عبايه, حقيبة = حقيبه
Diacritics (tashkeel) and tatweel are removedفســـتان = فستان
Arabic-Indic digits become 0-9٤٢ = 42
Persian letters become Arabic letters (ڤ گ پ چ)ڤيفو = فيفو

2. Sizes and units

Sizes are matched as they are written on your products, in Arabic and English. Numbers and their units are joined in one form:

Shopper typesSeekGist reads
حذاء رياضي مقاس 42, white sneakers size 42sneakers in size 42
فستان مقاس M, black abaya size sa dress in size M, a black abaya in size S
جوال 256 جيجا, phone 256gba phone with 256gb of storage

3. Word endings and prefixes

Light stemming removes common prefixes (such as ال, لل, بال) and suffixes (such as ات, ين), but only while at least 3 letters stay. So لحفلة (for a party) is also read as حفلة. Brand names and broken plurals come from a dictionary instead, for example جوال and جوالات.

4. Dialect, Arabizi and typos

A built-in dictionary knows common dialect words and Arabizi spellings, and a typo-tolerant engine catches the rest:

KindQueries from the test set
Arabic short formsسماعه, شاشه
English typossneekers, jaket, evning dress, adiddas hoodie, hedphones, samsng phone
Arabizi (Arabic in Latin letters)shanta nike, sma3at sony, sama3at, shasha
Arabic and English mixedحذاء nike, فستان layali, تيشيرت oversize, جوال samsung
Arabic brand namesنايكي, اديداس, سامسونج, شاومي

Words like ابغى or عايز ("I want") are ignored, so ابغى فستان searches for a dress.

5. Synonyms

SeekGist starts with a shared starter list of synonyms in Arabic and English (for example جوال = موبايل = smartphone). You add the words of your own store in the dashboard under Synonyms, import or export them, and approve synonyms that SeekGist suggests from searches that found nothing.

  • Two-way synonyms: عباية = abaya; جاكيت = jacket = جاكيتات.
  • Phrases count as one thing: فستان سهرة means an evening dress, not every dress.
  • One-way rules: sneakers can also find running shoes, but not when the query has a word such as socks, so a socks search stays a socks search.

6. Full requests (intent)

SeekGist reads the intent in a full request. Example:

فستان سهرة تحت 100 ("an evening dress under 100")

PartWhat SeekGist does
فستان سهرةproduct: evening dress
تحت 100price limit: at most 100

It also understands:

  • Price words in Arabic and English: تحت, اقل من, بحدود, ميزانيتي, under, max, over, and k for thousands. Example: حذاء رياضي ابيض تحت 80 (white sneakers, at most 80).
  • "Cheap" (رخيص, رخيصة, ارخص, على قد الايد, cheap): in سماعات رخيصة the best matches are re-ordered by price. SeekGist never sorts the whole list by price, so the results stay relevant.
  • Use cases for computers and phones: programming, video editing and design ask for strong machines. Business, home and study only change the order.

On plans with AI rewrite, SeekGist can also ask an AI model to rewrite a hard request, such as فستان لحفلة زفاف ("a dress for a wedding party"). By default (ai=auto) the AI runs only when it helps: price or use-case requests, long requests, unknown words, or a search with no results. The answer is cached, so the same request does not call the AI again. If the AI is slow or fails, the rule-based reading above is used, so search always answers.

7. Measured quality

These numbers are read from eval/report.md when this site is built (run of 2026-10-14 (commit d772f92)). nDCG@10 is a score from 0 to 1 for how good the top 10 results are; 1 is perfect.

The public demo catalog (1981 made-up products, the queries on this page):

TestResult
All test queries, without AI: nDCG@100.927
Queries with no results, without AI2.5%
Arabizi group, without AI: nDCG@100.625

The Arabizi group of the demo catalog is the weakest part today: three Arabizi fashion queries (fostan, jakeet, 7itha2 riyadi) find nothing yet, because these spellings are not in the dictionary. The method is in eval/README.md.