How we built an AI assistant for the MINDS Alumni Network that cites its sources and politely declines everything else.
MINDS wanted alumni to get quick answers about the institute and the network without anyone making things up. A general-purpose chatbot was out: it had to answer only from MINDS's own material.
We used retrieval-augmented generation. Site content from Sanity, documents the team uploads, and selected pages from minds-africa.org are split into chunks, embedded with Google Gemini and stored in Upstash Vector.
For each question, the assistant finds the closest chunks and answers from those alone, with links to the sources it used. If nothing relevant comes back, it says so instead of guessing, and it declines off-topic and personal-data questions.
Content changes, so the index is rebuilt on a schedule and on demand. Editors keep writing in Sanity as usual, and the assistant stays in step with the site.