The challenge
A large organisation operating in a highly regulated environment had valuable knowledge scattered across internal documents and systems. Employees needed a faster way to find answers — without exposing sensitive data, and without an assistant that could confidently give incorrect information.
What we built
We built and deployed a secure generative-AI knowledge assistant using Retrieval-Augmented Generation (RAG). It retrieves relevant information from approved internal sources before generating an answer, and was designed around the organisation's strict privacy and infrastructure requirements.
Crucially, we added automated evaluation across groundedness, relevance and completeness — so answer quality could be measured and systematically improved, rather than spot-checked by hand.
The outcome
Through improvements to retrieval and prompting, evaluated answer accuracy rose from roughly 78% to 92%. Employees gained a faster way to access organisational knowledge while sensitive information stayed within approved infrastructure — and the evaluation framework gives the team a measurable way to keep the system reliable over time.