Software that reads Arabic the way people write it, not the way textbooks do

Arabic words are built from roots with prefixes, suffixes and attached pronouns, so one written token can carry the meaning of a whole English phrase. Short vowels are usually left out, which makes many words ambiguous until context resolves them. Letters change shape by position, and writers vary alef forms and the ending letters of words. A pipeline that ignores this will count the same word as five different words.

Dialect adds another layer. Modern Standard Arabic is used in formal documents, but customers write in Gulf, Najdi, Hijazi and other varieties, often mixed with English and sometimes typed in Latin letters. We do not pick one model and hope. We label a sample of your own text, run several candidate approaches against it, and choose on measured results: precision, recall and error types, reported separately for formal text, dialect and mixed text.

For narrow, high-volume tasks such as routing a complaint to the right department, a small fine-tuned model is often faster, cheaper and more predictable than a large language model. For open-ended tasks, such as summarizing a long contract or extracting a clause that is worded differently every time, a large model usually does better with a few well-chosen examples. Many real systems use both: a small classifier first, and a larger model only for what is left.

We show the trade-off in numbers: accuracy on your test set, cost per thousand documents, speed, and what it costs to retrain when your categories change. Labeling effort is the real cost driver. Five hundred to two thousand labeled examples are typically enough for a first useful classifier, and your own staff can label them in a few days with a tool we provide.

Arabic text from customers, employees or patients is personal data. The Personal Data Protection Law governs how it is collected, kept and transferred, and SDAIA publishes the law and the AI ethics principles that sit around it. We keep the labeled sample inside your environment, remove names and identifiers before sending text to any external service, and document what was processed and where.

Where the text is sensitive, small and mid-sized models can be hosted on infrastructure inside the Kingdom, so nothing leaves the country. The pipeline is delivered with the test set, the evaluation script and the trained model files, so you can retrain without us. A first pipeline takes four to eight weeks. A multi-source program with search, extraction and dashboards takes eight to sixteen.