A UK professional-services firm was drowning in internal enquiries that had already been answered somewhere in 40,000+ documents. Generic chatbots hallucinated. Off-the-shelf AI search couldn't be trusted with client data. We built a Retrieval-Augmented Generation platform on Azure UK-south where every answer is traceable to its source document.
📋 To publish: replace the anonymised client with a named firm, add a screenshot of the citation UI, and ideally a short recorded demo showing a grounded answer with source links.
The problem
40,000+ client files, policy documents, briefing notes and prior-work outputs across SharePoint, network drives and email. Seniors spent an hour a day answering questions that had been answered before. Consumer AI tools would either refuse (safety) or hallucinate (unsafe). Neither was defensible.
The approach
Every answer the platform gives must cite the source document and paragraph. The model can refuse. We don't let it generate from its own training. Two-week discovery defined the evaluation harness — a gold set of 200 real questions from the client, scored on accuracy, relevance and safety before a line of production code was written.
What we built
Outcomes
Client quote
"The citation UI is what sold it internally. Every answer shows its working. We're comfortable putting it in front of partner-level staff because we can see exactly where each claim comes from."Head of Knowledge ManagementUK professional-services firm (reference available on request)
📋 To publish: replace with named client quote + role + headshot once permission confirmed.
Tech stack
Related