Arabic & Sovereign Readiness

Make your Arabic content usable by a model hosted inside Saudi Arabia

Arabic documents, correspondence, contracts and call transcripts turned into retrievable, permission-aware, evaluated content, designed to run against sovereign or privately hosted models rather than a public endpoint.

Where this sits

Foundation

Intelligence

Agents

Our point of view

What we are seeing in the market

Observations from the engagements and tenders we work on across Saudi Arabia and the wider Gulf.

Sovereign hosting changes the design

In-country and privately hosted models come with different context limits, throughput and embedding options than public endpoints. Content prepared for one is rarely optimal for the other.

Arabic breaks naive pipelines

Right-to-left layout, ligatures, diacritics, mixed script, Hijri dates and Arabic-Indic numerals routinely survive an English pipeline as garbage that looks like text.

Nobody is measuring Arabic retrieval

Retrieval quality is usually reported from English test sets. The Arabic result is assumed rather than measured, which is exactly where user-facing services fail.

The work

What this practice delivers

Arabic extraction that keeps meaning

Arabic OCR with layout, table and column recovery, diacritic and ligature handling, Hijri and Gregorian date normalisation, and a quality gate that flags unreliable pages instead of indexing them.

Dialect-aware chunking and normalisation

Segmentation that respects Arabic sentence and clause structure, transliteration variants and domain terminology, so a passage remains answerable after it is split.

Embedding and model selection on evidence

Candidate Arabic embedding models and in-country endpoints compared on your own corpus and questions, with cost, latency and residency recorded alongside accuracy.

Residency-safe retrieval architecture

Parsing, embeddings, index and inference placed according to data classification, with permissions carried into the index so retrieval respects who may read what.

How the work is structured

An Arabic-first pipeline, measured in Arabic

Extract

Arabic OCR + layout
Quality gate

Normalise

Script and numerals
Hijri dates

Chunk

Clause-aware split
Terminology kept

Embed

Arabic model chosen on evidence
In-country endpoint

Evaluate

Arabic question set
Retrieval scored

Failed Arabic questions return to the test set and drive the next parsing or chunking change

Selected work

Work we have delivered

Client identities are withheld. The situations, the build and the change afterwards are as they happened.

Arabic document corpus

A Saudi public entity

The situation: Decades of scanned Arabic correspondence were indexed, but answers cited the wrong section or missed tables entirely.
What we built: Arabic OCR with table recovery, a quality gate on unreadable pages, clause-aware chunking and a labelled Arabic question set.
What changed: Retrieval accuracy became a reported number in Arabic rather than an assumption carried over from English.

Residency-constrained retrieval

A regional insurer

The situation: Policy and claims documents could not leave Saudi Arabia, which ruled out the assistant architecture originally proposed.
What we built: In-country parsing, embedding and index with a privately hosted model endpoint, and permissions carried into retrieval.
What changed: The service cleared review without reducing the corpus it was allowed to use.

Where we are strongest

We are strongest when the content is Arabic, sensitive, and cannot leave Saudi Arabia.

Arabic handled at parse time

Most Arabic retrieval failures are created before the model is ever called, so that is where we spend the effort.

Model choice stays reversible

Parsing, chunking and embeddings are kept separable from the model, so a sovereign endpoint can be adopted or replaced without redoing the corpus.

Measured in the language of use

The evaluation set is Arabic first, including dialect and transliterated terms, because that is what users type.

Classification drives placement

Where each stage runs follows the sensitivity of the content, and the reasoning is written down for review.

Straight answers

Questions we are asked most

Can this run entirely inside Saudi Arabia?

Yes. Parsing, embeddings, the index and inference can all be placed in-country, including against privately hosted or sovereign model endpoints.

Do we have to choose a model first?

No, and we prefer not to. We prepare the corpus and the evaluation set first, then the comparison decides.

What if the documents are poor scans?

The quality gate reports what cannot be read reliably instead of indexing it silently, so the corpus is honest about its own coverage.

Clients

Organisations we work with

Across banking, healthcare, government, retail, energy and telecom.

FAQ

Frequently asked questions

How do I make Arabic documents AI-ready?

Extract with Arabic-capable OCR that preserves layout and tables, normalise script, diacritics, numerals and dates, chunk along Arabic clause boundaries, embed with a model evaluated on your own Arabic questions, and carry access permissions into the index.

Is an Arabic-specific embedding model necessary?

Often it helps, but not always, the honest answer comes from comparing candidates on your corpus and your question set rather than from a leaderboard.

How does this relate to a sovereign model programme?

It is the prerequisite. A sovereign model still answers from whatever content it can retrieve, so the quality of the Arabic corpus sets the ceiling on the result.