8
ministries onboarded via repeatable playbook
Answer time reduced from 30+ minutes to seconds
Tens of thousands of hours projected annual savings at full rollout
Full data sovereignty on private infrastructure
RAG-powered knowledge assistants deployed across eight European government ministries reduced policy question answer times from over 30 minutes to seconds, with every answer traceably cited to source documentation. Built on sovereign infrastructure with open-source LLMs, the platform used hybrid retrieval and a repeatable onboarding playbook to scale across ministries in weeks.
The Problem
A European government IT services provider needed to deploy knowledge assistants across eight ministries. Civil servants spent 30 minutes to several hours finding answers to routine policy questions, navigating fragmented SharePoint sites, shared drives, and internal portals with keyword search that was essentially useless for nuanced queries. Institutional knowledge lived in people's heads, and when those people changed roles, it left with them.
The Solution
Working alongside a platform partner, we designed and deployed Retrieval-Augmented Generation (RAG) assistants for multiple ministries within weeks. The architecture combined open-source large language models on private infrastructure with a retrieval layer that returned answers grounded in specific source passages, with traceable citations back to the original document and section.
The retrieval layer used hybrid search, combining semantic vector similarity with BM25 keyword matching, to handle the full range of government queries: from conceptual policy questions to precise regulatory lookups where exact wording matters. Every answer included citations linking back to the source passage, and a verification step confirmed that each citation accurately represented the source content. Civil servants could click through to the original document; auditors could trace from question to answer to source.
Scaling from one ministry to eight required a repeatable onboarding playbook covering data readiness assessment, document ingestion, access model configuration, evaluation, and operational handover. Each ministry had different documentation, sensitivity levels, and access requirements. Tenancy was enforced at the retrieval layer: a civil servant in one ministry could not receive answers grounded in another ministry's restricted documentation. The playbook compressed onboarding from months to weeks.
Results and Impact
| Metric | Outcome |
|---|---|
| Time to answer | Reduced from 30+ minutes to seconds for typical policy questions |
| Projected time savings (initial ministries) | Thousands of hours annually across the first ministries onboarded |
| Projected time savings (full rollout) | Tens of thousands of hours annually once scaled to all eight ministries |
| Deployment timeline | Functional prototypes with real ministry documentation delivered within weeks |
| Security and sovereignty | All data processed within controlled sovereign infrastructure; no external data transfer |
| Vendor independence | Architecture built on open-source LLMs; no proprietary lock-in |
| Onboarding cycle | Standardised playbook enabling weeks-per-ministry scaling |
User testing was particularly telling. Civil servants who had spent years navigating fragmented repositories responded positively not just to the speed, but to the citation model. Being able to see exactly where an answer came from addressed the trust barrier that had undermined previous AI adoption attempts in government.
Key Takeaways
-
RAG quality is bounded by document quality. The majority of engineering effort went into ingestion: parsing, cleaning, chunking, and metadata tagging. If the source documentation is poorly structured or inconsistent, no amount of model sophistication will produce reliable answers.
-
Hybrid retrieval outperformed embeddings alone on real government documents. Semantic search handled conceptual queries well but struggled with precise regulatory language and directive references. Adding BM25 keyword matching produced measurably better results across the mixed query types civil servants actually ask.
-
A reusable onboarding playbook made scaling realistic. Without a standardised process for data readiness, ingestion, access configuration, and evaluation, each new ministry would have been a standalone project. The playbook gave the government confidence that scaling from two ministries to eight was a scheduling problem, not an engineering one.