Data Foundation

Your AI is only as good as the Arabic it can retrieve

Arabic and English documents, images and audio turned into governed, retrievable knowledge: ingestion, chunking, embeddings, vector search and a retrieval quality score, prepared where the data already lives in Saudi Arabia.

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.

AI-readiness replaced the modern data stack as the mandate

Requests that used to ask for warehouses and dashboards now ask for retrieval architecture. Vector search and RAG appear in enterprise requirements that had never named them before.

Most of the value is in the unstructured pile

Contracts, policies, tickets, reports, scans and call recordings hold the answers people actually ask for, and almost none of it is in the warehouse.

Retrieval quality, not the model, decides the answer

Teams change models hoping for better answers when the real fault is chunking, metadata, permissions or a stale index. Retrieval is the part worth measuring first.

The work

What this practice delivers

Unstructured ingestion

PDFs, scans, spreadsheets, images and audio parsed with layout and table awareness, Arabic OCR and speech, with the source and version retained on every extract.

Chunking, embeddings and vector store

Chunking strategies chosen per document type, bilingual embeddings, hybrid keyword-plus-vector search and a store selected for your scale, cost and residency rules.

Permission-aware retrieval

Row and document-level access carried into the index so an assistant can never surface a document the person asking is not allowed to open.

Retrieval evaluation and lineage

Recall and precision measured against a labelled question set, freshness monitoring, and every AI answer traceable to the document, page and version behind it.

How the work is structured

From an unstructured pile to retrievable, governed knowledge

Sources

Documents and scans
Tickets, mail, audio
Warehouse tables

Parse

Layout and tables
Arabic OCR and speech

Prepare

Chunking strategy
Metadata and permissions
Bilingual embeddings

Serve

Hybrid search
Re-ranking
Grounded answers

Prove

Recall and precision
Freshness and lineage

Missed answers become labelled cases that improve chunking, metadata and ranking

Selected work

Work we have delivered

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

Policy and standards retrieval

A Saudi utility

The situation: Engineers searched thousands of standards documents by filename, and an early assistant answered from superseded revisions.
What we built: Layout-aware parsing, revision metadata, hybrid retrieval with re-ranking and a labelled question set used to tune chunking.
What changed: Answers started citing the current revision with the page behind them.

Permission-aware knowledge

A financial services firm

The situation: A knowledge assistant was blocked at review because the index ignored who was allowed to read each document.
What we built: Access rules carried into the index, per-user filtered retrieval and an audit trail of which documents supported each answer.
What changed: The assistant cleared review without narrowing the corpus it could use.

Where we are strongest

We are strongest when the answer exists somewhere in the organisation and nobody can retrieve it reliably.

Arabic documents treated properly

Right-to-left layout, mixed-script terms, scanned quality and Arabic numerals are handled at parse time, where the damage otherwise happens.

Retrieval is measured, not assumed

A labelled question set exists before the assistant does, so tuning is evidence-led from the first week.

Governance travels with the chunk

Classification, retention and access follow the content into the index rather than stopping at the warehouse boundary.

Portable by design

Parsing, chunking and embeddings are kept separable from the store and the model, so neither choice becomes permanent.

Straight answers

Questions we are asked most

Do we need a vector database?

Often the store you already run is enough at your volume. We size it against your corpus, latency and residency rules before recommending a new component.

Can this stay inside Saudi Arabia?

Yes. Parsing, embeddings and the index can all run in-country, including with private model endpoints where required.

How does this relate to our data platform?

It extends it. The same governance, lineage and access model covers the unstructured side instead of a second, ungoverned copy appearing beside it.

Clients

Organisations we work with

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

FAQ

Frequently asked questions

What does AI-ready data actually mean?

Content that can be retrieved correctly for a question: parsed, chunked, embedded, permission-aware, current, and traceable back to its source document and version.

How do you evaluate a RAG system?

Against a labelled question set: did retrieval return the right passage, did the answer use it, was it current, and how often did it miss, reported per release.

Do you handle Arabic scanned documents?

Yes, Arabic OCR with layout and table recovery, plus a quality check that flags pages too poor to trust rather than silently indexing them.