Skip to content

Government

A public sector organisation makes its intranet answerable without anything leaving the network

Internal guidance was scattered across departments on a Drupal intranet. Staff asked colleagues instead of searching. Everything now runs on their own servers, model included.

Problem

Twelve departments published their own procedures, guidance, and forms on a shared Drupal intranet, each with its own structure and vocabulary. A staff member with a question had to guess which department owned the answer before they could search for it, and most gave up and asked a colleague. That worked until the colleague was on leave or had retired. Every hosted AI tool the organisation looked at was rejected during the security review, because internal guidance was not allowed to leave the network under any circumstances.

Solution

The full stack was installed on the organisation's own servers: the module, the vector database, and an open language model, so no request reaches an external service. Access follows the existing Group configuration, so staff only get answers from the departments they already have access to. Because departmental vocabulary differed so much, the index stores the department and the last revision date with every passage, and answers show both, which lets staff tell immediately whether they are reading current guidance.

Result

Staff stopped needing to know the organisational chart in order to find an answer. Finding applicable guidance went from a chain of questions to a single one. The security review passed without conditions, because the assessment came down to a system running on infrastructure the organisation already operated and audited.

Impact

Knowledge that used to sit with long serving staff became available to everyone, which mattered most for new hires and for departments facing a wave of retirements. The intranet stopped being a place things were filed and became a place questions get answered.

Could this be your knowledge base?

We will show you what it looks like on your own content.