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 electronics and fashion 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.
| What | Example |
|---|---|
| Alef forms (أ إ آ ٱ) become ا | أسود = اسود |
| ة becomes ه, ى becomes ي, ؤ becomes و, ئ becomes ي | شاشة = شاشه, سماعة = سماعه |
| Diacritics (tashkeel) and tatweel are removed | لابـــتوب = لابتوب |
| Arabic-Indic digits become 0-9 | ٢٥٦ = 256 |
| Persian letters become Arabic letters (ڤ گ پ چ) | ڤيفو = فيفو |
2. Units are understood
Numbers and their units are joined in one form, in Arabic and English:
| Shopper types | SeekGist reads |
|---|---|
لابتوب 16 جيجا رام, laptop 16gb ram | a laptop with 16gb of RAM |
جوال 256 جيجا, phone 256gb | a phone with 256gb of storage |
شاشة 27 بوصة, monitor 27 inch | a 27 inch monitor |
monitor 144hz | a 144hz monitor |
laptop 1tb ssd | a laptop with a 1tb SSD |
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 للبرمجة (for programming) is also read as برمجة. 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:
| Kind | Queries from the test set |
|---|---|
| Arabic typos and short forms | لابتب, سماعه, شاشه |
| English typos | labtop, lenvo laptop, hedphones, moniter, keybord, samsng phone |
| Arabizi (Arabic in Latin letters) | labtob, jawal samsung, sama3at, shasha, sa3a thakiya |
| Arabic and English mixed | لابتوب lenovo, جوال samsung, ماوس wireless, حذاء nike |
| Arabic brand names | لينوفو, سامسونج, شاومي, اديداس |
Words like ابغى or عايز ("I want") are ignored, so ابغى لابتوب searches for a laptop.
5. Synonyms
SeekGist starts with a shared list of electronics synonyms in Arabic and English. You can add your own in the dashboard under Synonyms, import or export them, and approve synonyms that SeekGist suggests from searches that found nothing.
- Two-way synonyms:
جوال=موبايل=هاتف ذكي=smartphone;كيبورد=لوحة مفاتيح=keyboard. - Phrases count as one thing:
حامل شاشةmeans a monitor stand, not every kind of stand. - One-way rules:
laptopalso finds product lines such as ThinkPad and IdeaPad, but not when the query has an accessory word such as "bag", so a laptop bag search stays a bag search.
6. Full requests (intent)
SeekGist reads the intent in a full request. Example:
لابتوب للبرمجة تحت 800("a laptop for programming under 800")
| Part | What SeekGist does |
|---|---|
لابتوب | product type: laptop |
تحت 800 | price limit: at most 800 |
للبرمجة | heavy use case: at least 16 GB RAM, and stronger machines first |
It also understands:
- Price words in Arabic and English:
تحت,اقل من,بحدود,ميزانيتي,under,max,over, andkfor thousands. Example:لابتوب العاب تحت 1000(a gaming laptop, at most 1000). - "Cheap" (
رخيص,رخيصة,ارخص,على قد الايد,cheap): inسماعات رخيصةthe best matches are re-ordered by price. SeekGist never sorts the whole list by price, so the results stay relevant. - Use cases: programming, engineering, 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 هدية لابني يحب الالعاب ("a gift for my son who loves games"). 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.
A real store catalog (9,428 electronics products; the store stays private and is not named):
| Test | Result |
|---|---|
| 228 real shopper queries, without AI: nDCG@10 | 0.905 |
| Same queries, without AI: queries with no results | 0.0% |
| Same queries, typo group / Arabizi group, without AI: nDCG@10 | 0.902 / 1.000 |
| Same queries, with AI rewrite (agent-recorded stand-in, not live Haiku 5.5): nDCG@10 | 0.906 |
| Second test set, 387 new queries written by a different person than the synonym author, without AI: nDCG@10 | 0.804 |
| Same 387 queries, with AI rewrite (agent-recorded stand-in, not live Haiku 5.5): nDCG@10 | 0.851 |
| Suggestions while typing: the right product in the top 5 for 2-4 letter prefixes | 0.935 |
The public demo catalog (1981 made-up products, the queries on this page):
| Test | Result |
|---|---|
| All test queries, without AI: nDCG@10 | 0.927 |
| Queries with no results, without AI | 2.5% |
| Arabizi group, without AI: nDCG@10 | 0.625 |
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 AI numbers were not made with a live model: an agent wrote the AI answers with the production prompt, so treat them as a stand-in until a live run. The method is in eval/README.md.