A private assistant that answers from the site's own documentation, with the page it came from.

SITUATION

The answer was in the documentation, and the documentation was in ten places.

Technical documentation, procedures, drawings and supplier files, spread across systems and shared drives, and the fastest way to an answer was finding the person who remembered it. Every cloud assistant that could have read them asked for the documents to leave the organization first.

WHAT WAS BUILT

A local language model with retrieval over the organization's own documents, running on one GPU node on site. It answers a question from the documentation and shows the page the answer came from, so the person checking it reads the source, not a summary. The same system runs at a robotics and automation integrator, at a wood industry group and inside Global Vision.

HOW IT RUNS

On the organization's own hardware, on-premise or fully air-gapped, on an open-weight model under a permissive license. No document, question or answer leaves the site.

WHAT CHANGED

An answer with its source in seconds, from the documents the organization already had.

THE STACK

MODEL
Open-weight language model, local

RETRIEVAL
Over the organization's own documents, page returned with the answer

RUNS ON
1 GPU node on site

CONNECTIVITY
On-premise or air-gapped