The problem
Support handled a high volume of repetitive questions that were already answered in the help centre — but nobody read the help centre, and search there was poor.
What we did
We built a retrieval assistant grounded strictly in the company's own documentation, with citations on every answer and an explicit hand-off to a human whenever confidence was low or the question involved billing. Unanswered questions were logged as a queue for the content team.
Outcome
- Repetitive questions largely deflected before reaching the inbox
- Every answer carries a source link, so trust stayed intact
- Gaps in the documentation surfaced as data instead of guesswork
Stack & channels
LLM APIsRetrieval (RAG)Node.jsWeb widgetAnalytics