Government AI Assistant with On-Premise RAG

What problem does this solve?

Government entities hold the answers their own staff need, but the answers are buried. A procurement officer looks for the clause that governs a variation order; a licensing officer needs the circular that superseded last year's rule; a new engineer needs to know which specification applies to a road category. The material exists in a document management system, a shared drive, an email archive and a set of scanned PDFs, and finding it depends on knowing which folder to open. Search returns file names, not answers, and scanned Arabic documents are often not searchable at all.

The obvious shortcut is a public AI chat service, and it is the one thing information-security policy cannot allow. Uploading internal correspondence, HR files, contracts or draft legislation to a service outside the entity's control breaches data-classification rules and, in several GCC entities, the terms of the hosting approval itself. Staff do it anyway on personal devices, which turns a productivity problem into a data-leakage problem.

The second failure mode is a pilot that impresses in a demo and collapses in production. A model that answers fluently but invents a clause number is worse than no assistant, because officers act on it. And an assistant that ignores document permissions will happily quote an HR investigation file to whoever asks. Both problems are architectural, not cosmetic: they are decided by how retrieval, permissions and evaluation are built, long before anyone chooses a model.

How the solution works

Retrieval-augmented generation (RAG) separates knowledge from language. The documents stay in your systems and are indexed inside your network; when a question arrives, the system retrieves the passages that user is allowed to read and gives them to a private language model with an instruction to answer only from that text and cite it. The model contributes wording and reasoning, not facts. When nothing relevant is retrieved, the correct behaviour is to say so and offer the closest documents, not to improvise.

Swedish Technology builds this as an owned system rather than a subscription: open-weight models running on your GPU servers or in a private UAE-region cloud, an index you control, connectors to your document management and correspondence systems, and role mapping that inherits existing permissions. Delivery includes an evaluation set of real questions with agreed correct answers, so accuracy is a measured number that can be re-tested after every change, and an audit trail that records which sources produced which answer.

1. 1
   Input Agree the document scope: which repositories, which classifications, which languages, and who is allowed to see what. Scope is a security decision, not a technical one.
2. 2
   Capture Connectors pull documents from the DMS, correspondence system, shared drives or database exports. Scanned files pass through OCR tuned for Arabic and English; layout is preserved so tables and headings survive.
3. 3
   Processing Text is split into passages with metadata (document ID, section, issue date, version, classification, source ACL), embedded into vectors and written to a hybrid index that combines semantic and keyword search.
4. 4
   Retrieval A question is rewritten if needed, filtered by the user's identity and the document ACLs, then matched against the index. Candidate passages are re-ranked so the strongest evidence reaches the model.
5. 5
   Generation The private model receives the question and the retrieved passages with a strict instruction: answer from this text, cite each statement, and state clearly when the answer is not present.
6. 6
   Action The officer reads the answer with its citations, opens the source document in one click, and uses it to draft a reply, complete a form or route a case. Sensitive workflows require an explicit human approval step.
7. 7
   Reporting Every question, retrieved source, answer and feedback rating is logged. Dashboards show usage by department, unanswered questions, and accuracy against the evaluation set over time.

Reference architecture

The design goal is that the assistant can be switched off, re-pointed at a different model, or audited end to end without touching the source systems. Five layers, each replaceable.

Deployment options: Runs on-premise on your own GPU servers, in a UAE-region private cloud, or fully air-gapped with model weights and container images transferred through a controlled media process. Nothing in the design requires an outbound internet connection at run time; updates are delivered as signed offline bundles.

Key capabilities

Integrations

The assistant reads from systems of record and never becomes one. Write-back, where it exists, is limited to drafts and is always subject to human approval.

Industry use cases

Implementation approach

1. 1
   Discovery (1-2 weeks) Identify the user group, the document set, the classification levels involved and twenty to fifty real questions staff actually ask. This becomes both the scope and the test.
2. 2
   Security architecture Agree deployment mode (on-premise, private cloud, air-gapped), identity model, permission inheritance, logging and retention with the information-security team before any build.
3. 3
   Ingestion and OCR baseline Connect one repository, run OCR on a representative sample of scanned and native documents, and measure extraction quality. Poor OCR is the most common silent cause of poor answers.
4. 4
   Proof of concept (4-6 weeks) Build the retrieval pipeline and interface on the agreed scope, run the evaluation question set, and report measured accuracy, refusal behaviour and latency.
5. 5
   Pilot with a real department Twenty to fifty users for four to eight weeks, with a feedback button on every answer and a weekly review of unanswered or wrong answers.
6. 6
   Hardening and integration Add remaining repositories, tighten guardrails, integrate workflow and reporting, complete penetration testing and load testing at realistic concurrency.
7. 7
   Roll-out and training Bilingual user guidance on what the assistant may and may not be used for, administrator training on re-indexing and evaluation, and a published escalation path.
8. 8
   Operate and re-evaluate Scheduled re-indexing, quarterly evaluation runs, model upgrade assessments, and a documented rollback to the previous index and model version.

Security & deployment

The assistant runs entirely inside the entity's boundary: model weights, index, application and logs on your infrastructure, with no outbound calls at inference time. Users authenticate through the existing identity provider and retrieval is filtered by their group membership and the source document's access control list, so the assistant can never widen access. Prompts and outputs pass through guardrails for prompt injection, sensitive-data patterns and out-of-scope requests, and every interaction is logged with the retrieved source identifiers for audit. Administrative actions such as re-indexing, changing prompts or replacing a model are privileged, versioned and reversible. For classified environments the entire stack is delivered air-gapped, with signed offline update bundles and a documented media-transfer procedure.

Limitations & prerequisites

• The assistant is only as good as the documents indexed. If policy is held in people's heads, in email threads that are not ingested, or in unreadable scans, the answer will be incomplete regardless of the model.
• OCR quality on old scanned Arabic documents, faxes and stamped pages is the single biggest constraint; some archives need re-scanning or manual correction before they are usable.
• Grounded generation reduces fabrication but does not eliminate it. Answers on legal, financial or safety matters must be verified against the cited source before action, and the interface should make that easy rather than optional.
• Numerical reasoning across many documents (totals, comparisons across hundreds of contracts) is unreliable in a language model and should be handled by a query against structured data instead.
• Running open-weight models on-premise requires GPU capacity, power and cooling, plus staff who can operate it; concurrency targets drive hardware cost more than model choice.
• Permission inheritance depends on the source system having correct permissions today. Existing over-sharing in a document repository becomes visible quickly once search works well.
• Model and framework versions move fast; budget for a periodic upgrade assessment and re-evaluation rather than assuming a fixed system for five years.

On-premise RAG vs public AI service vs fine-tuned model

The three approaches solve different problems and have very different governance profiles.

For most government use cases retrieval is the right default, with fine-tuning considered later and only for style, terminology or a narrow classification task.

FAQ

Related topics & solutions

• Arabic Government AI: Bilingual Assistants & Document AI
• AI Document Intelligence: OCR, Classification & Extraction
• AI Tender Analysis & Bid Evaluation Support
• Document Management & Correspondence System
• AI Products Portfolio
• Primavera P6 – Project Management
• SAIF ÔÇô Cybersecurity
• Business Intelligence