Knowledge assistant
Cited answers from your own documents and systems.
A knowledge assistant lets staff ask questions in plain Arabic or English and get a direct answer drawn from your documents, intranet, e-mail archives and business systems. Every answer shows its sources, and it draws only on what the person asking is already allowed to open. Most of the engineering goes into retrieval: finding the right passage, in the right language, from the right version.
The problem
Where knowledge gets stuck
- 01Policies, procedures and precedents sit across SharePoint sites, file shares, e-mail and older systems, with no single search that covers them.
- 02Older Arabic circulars and letters are often scanned PDFs, which keyword search cannot see at all.
- 03Enterprise search returns pages of links, when staff want the answer and the paragraph it came from.
- 04Generic chat tools answer confidently without sources, and cannot respect who may see what.
How it is built
How the assistant is built
Retrieval does most of the work. Content is ingested with its permissions, cleaned and indexed for both keyword and semantic search. At question time, only passages the user may open are retrieved, re-ranked and passed to the model, which must cite them.
Layer 1
People and channels
- Web assistant
- Microsoft Teams
- Intranet portal widget
Layer 2
Answering
- Query understanding, Arabic and English
- Permission-filtered retrieval
- Re-ranking
- Answer drafting with citations
Layer 3
Ingestion and index
- Connectors with permission sync
- OCR and Arabic text normalisation
- Chunking and metadata
- Hybrid keyword and vector index
Layer 4
Infrastructure
- Private model serving
- Index and document stores
- Identity provider
- Logs and analytics
Capabilities
Knowledge assistant
Permission-aware by design
Access lists are synced from each source and applied at retrieval time, not after the answer is written. If a user cannot open a document in SharePoint, the assistant cannot quote it to them.
Arabic that search can see
Scanned Arabic PDFs go through OCR, then text normalisation — alef and hamza forms, taa marbuta and haa, alef maqsura and yaa, diacritics and tatweel — so different spellings of the same word still match.
Hybrid search and re-ranking
Keyword search catches exact terms such as circular numbers and article references; semantic search catches the same idea in different words. A re-ranking model then orders the combined results before anything reaches the language model.
Answers with citations
Each answer links to the passages it used, with document title, version and page. When the sources do not support an answer, the assistant says so rather than guessing.
A feedback loop
Users rate answers and flag wrong ones. Flags go to content owners, and recurring problems join the evaluation set that every change is tested against.
Content-gap reporting
Reports show the questions no document answers well, and the documents that are out of date or contradict each other. Content owners get a clear list of what to write or fix.
How it works
From question to cited answer
- Question in Arabic or English
- Permitted sources searched
- Passages re-ranked
- Answer drafted with citations
- User checks and rates
Connects to
Works with
- Microsoft SharePoint and OneDrive
- Microsoft Exchange and Outlook archives
- Microsoft Teams
- Network file shares
- Confluence and other wikis
- OpenText and other document management systems
- Google Workspace
- ServiceNow knowledge bases
- SAP and Oracle reference data
- Microsoft Entra ID and Active Directory
Where it runs
On-premises
Models and data run on hardware in your own facilities. Nothing leaves your network unless you allow it.
Private cloud in the Kingdom
Dedicated, isolated capacity in a Saudi data-centre region of a provider you select — the elasticity of cloud without a shared public AI service.
Hybrid
Sensitive workloads stay on-premises; less sensitive ones use private cloud capacity. You set the policy for what goes where.
What we measure
Measures we agree with you
We agree these measures with you before we build and report against them once the assistant is in use. They are targets we track together, not results we promise in advance.
01
Answer accuracy
Share of answers judged correct and fully supported by their citations, on an agreed Arabic and English test set and on sampled live questions.
02
Retrieval quality
How often the right passage appears among the top results for questions with known answers — the early indicator of answer quality.
03
Time to answer
Time staff spend finding the answer to common questions, measured before launch and again in daily use.
04
Adoption
Weekly active users, and the volume of routine questions still reaching experts and help desks.
Indicative delivery path
Indicative delivery path
- 01
2–3 weeks
Scope and sources
We choose the first audience and content sources, review permissions and content quality, and build a test set of real questions with answers your experts agree on.
- 02
4–6 weeks
Ingest and tune
Connectors, OCR and indexing go in. We tune chunking, search and re-ranking against the test set until retrieval quality reaches the agreed level.
- 03
3–4 weeks
Pilot
A pilot group uses the assistant in daily work. We review feedback, close gaps and confirm that permissions hold before a wider launch.
- 04
Ongoing
Roll out and operate
Sources and departments are added in stages, and content-gap reports go to their owners on a regular cycle.
Durations are indicative and depend on scope, data and integrations.
Questions
Knowledge assistant
How do you stop people seeing documents they should not?
Permissions are copied from each source system along with the content and kept in sync. Retrieval filters by the user’s identity and groups before anything reaches the model, so a restricted document never becomes part of the answer. We test this explicitly with accounts in different roles before launch.
Can it read our scanned Arabic documents?
Yes, with limits worth knowing. OCR quality depends on the scan, the fonts and any handwriting: clean printed Arabic reads well, while faded or handwritten pages may not. We measure OCR accuracy on a sample of your documents during scoping and tell you which collections need re-scanning or manual review.
What happens when the answer is not in the documents?
The assistant says it could not find a supported answer and shows the closest sources it did find. It is instructed not to answer policy questions from general knowledge. Those questions are logged as content gaps for the owners to address.
Does it replace our intranet or document management system?
No. It sits on top of the systems you already have and reads from them through connectors. Documents stay where they are, owned and maintained by the same teams; the assistant only makes them easier to find and use.
Other solutions
All solutionsPrivate AI platform
Open-weight and commercial models served inside your boundary, behind one governed gateway.
View solutionDocument intelligence
Invoices, contracts and applications read, checked against your records and posted — people review the rest.
View solutionCustomer service agent
Arabic and English service on web chat, WhatsApp and voice — approved answers, lookups, hand-over to people.
View solutionAgentic operations
AI agents that handle procurement, invoice, HR and IT requests end to end — within limits you set.
View solution
Have a process that should run itself?
Tell us about it. We will help you work out whether AI is the right answer, what it would take, and what it should return.

