Policy Intelligence at a European Government — LLM-Powered Analysis Across 14 Policy Domains

MindMap deployed a policy-intelligence platform composed of Enterprise Knowledge Engine (Ke) as the sovereign-knowledge-and-RAG backbone, Summarization Wizard (Sw) for the policy-summarisation-and-briefing-assembly layer, Policy Q&A Agent (Pq) for the cross-domain policy-question-answering layer, and Sovereign LLM Platform (Sl) for the on-premises model serving.

Phase one was the policy-corpus ingestion. The 240,000+ policy documents across the 14 domains were ingested into the platform's structured knowledge base, with per-domain ingestion workflows that respected each domain's document-handling conventions and access-restriction patterns. The continuous-flow ingestion handles the new material as it arrives, with material-classification routing each document to the appropriate access-restriction tier.

Phase two was the cross-domain knowledge-graph build. The platform's knowledge graph models the policy concepts, the legislative-and-regulatory references, the cross-domain policy-relationships, the historical policy-evolution and the comparable-jurisdiction reference material. The graph enables cross-domain queries that the per-domain document corpora could not support directly.

Phase three was the analyst-workflow integration. Policy analysts use the platform's conversational interface to query the policy corpus, with the platform's responses grounded in citable source-material and the cross-domain reasoning explicit. The Summarization Wizard component supports the briefing-assembly workflow — producing draft briefings with the underlying source-material citations that the analyst confirms or modifies.

Phase four was the senior-minister briefing-support layer. The platform supports the briefing-preparation workflow for senior-minister engagements — assembling the relevant cross-domain material, drafting the briefing structure, surfacing the analyst-team's recent thinking on the relevant topics — with the senior-minister briefing maintaining the analyst-team's accountability for the briefing content.

The platform runs entirely on the government's on-premises infrastructure inside the central data centre, with the access-restriction tiers enforced at the storage-and-compute layer. The sensitive material is processed only on the tiers where the appropriate access-restriction applies; the platform's user-and-role model enforces the per-tier access-control.

Sovereign LLM Platform serves Llama 3.1 70B-Instruct fine-tuned on the policy-corpus content (with appropriate handling of the access-restriction tiers — the fine-tuning corpus is drawn from the less-restricted tiers, with the restricted-tier material accessed at inference time through the access-controlled RAG layer rather than embedded in model weights).

Enterprise Knowledge Engine handles the sovereign-knowledge-and-RAG backbone. The policy corpus is indexed using a combination of classical full-text search (BM25 for the structured-term queries) and embedding-based semantic search (for the policy-concept-and-relationship queries). The knowledge graph supplements the document-based retrieval with the cross-domain policy-relationship navigation.

Policy Q&A Agent's reasoning layer combines the retrieval results with the LLM-generated reasoning to produce structured policy-question answers with explicit source-provenance. Each answer surfaces the underlying source documents (with the document title, the relevant excerpt, the publication date, the authoring body) that support the answer, allowing the analyst to validate the reasoning against the citable material.

Summarization Wizard supports the briefing-assembly workflow with policy-summary generation grounded in the underlying source-material. The generated briefings include source-citation per substantive claim, ensuring the analyst-team's accountability for the content is preserved.