Support, customer success, and operations were all answering the same questions a hundred times a week. Confluence search was unreliable; Slack history was fragmented across years of channels.

Earlier 'just chatbot it' attempts produced confident-sounding but wrong answers, with no source visibility — so leadership pulled the plug.

We unified the corpus first: Confluence pages, Slack archives (curated), Drive docs, and a hand-picked set of authoritative SOPs. Each chunk carries its source URL and a freshness score.

Retrieval runs through a re-ranker tuned to operational queries; the model is forced to cite the chunks it used or it returns 'not enough context'. No source means no answer.

Every answer ships with an embedded feedback control that writes into a dataset we use to retrain the re-ranker weekly.

Median answer time dropped from minutes (and several Slack hops) to under five seconds with a citation.

Acceptance rate sits at 92% across the three teams that adopted it; the unaccepted queries are now the dataset that drives improvements.

Two human-written runbooks were retired because the system answered better and stayed fresh.