The challenge
Support agents were losing minutes to scattered policy documents on every single customer question. Regulation ruled out sending customer or policy data to any external AI provider, which ruled out every hosted assistant on the market.
The obstacle was never the model. It was that nobody could show where an answer had come from.
Our solution
We designed a sovereign stack: an open-weight language model served on the organisation's own GPUs, a self-hosted vector database, and a retrieval pipeline that enforces document-level permissions per user rather than per index.
Every answer carries the exact source passage it came from, and every query lands in a log the support team cannot edit. Attribution plus an audit trail is what turned a blocked project into an approved one.
What we built
- Open-weight model served on internal infrastructure, with no external API calls
- Self-hosted vector search with role-based, document-level access
- Ingestion pipeline with version tracking, so a superseded policy stops being quoted
- Citations and confidence indicators surfaced in the interface
- Immutable audit logging and an evaluation dashboard
The results
The measures that mattered here were not speed but provenance and permission: whether every answer could be traced, and whether anything ever left the network.