Internal Enquiry Desk — RAG Chatbot on Swiss Infrastructure
A Swiss research institution runs a specialist enquiry desk. The knowledge behind it has accumulated over years in internal note collections. We were asked to quote for a pilot making that knowledge accessible through natural-language questions — with source attribution, exclusively on Swiss infrastructure. The quote is on the table and the decision is pending; this entry describes the proposed solution, not delivered results.
Starting point
About ten people answer specialist enquiries, researching in internal note collections that grew organically over years. Newcomers need a long time to learn where things are. What was wanted was not a replacement for the experts, but a research path that finds the existing answer faster and shows where it came from.
Our proposed solution
A RAG architecture built on our proven production stack rather than a build from scratch: the knowledge base as a vector database in a Swiss data centre, a language model operated in Switzerland or the EU or self-hosted from open weights, containerised and therefore movable to their own infrastructure later. Every answer points to its source. Staff ratings and corrections are captured so quality improves verifiably rather than by impression.
Acceptance criteria instead of promises
The client set at least 90 percent correct answers as the success criterion. In the quote we stated openly that this threshold is not reliably reachable after the first proof of function but only after the build-out stage — and that it depends on the jointly chosen model family that can be operated in Switzerland. Measurement runs against a catalogue of twenty to forty test questions with model answers supplied in advance, in a moderated acceptance session, with one rework cycle included. What «correct» means is defined jointly before the project starts.
What the pilot deliberately excludes
Replacing the existing note collection with a central knowledge base, automatic write-back of corrections, provision to external users, an extended guard layer against prompt injection, high availability with an SLA, single sign-on, continuous synchronisation of new content, and additional languages. These are accounted for architecturally but are not part of the pilot. A pilot that contains everything is no longer a pilot — it is a project without an exit point.
Status
Quote delivered, decision pending. Three separately orderable steps are planned: requirements and architecture workshop, proof of function, build-out with measured acceptance. The source code and the knowledge base created during the project transfer to the institution.