A RAG-powered knowledge assistant that cut ops Q&A time by 70%.

Orbit's ops team answered the same 80 questions in different shapes every day. Sources were scattered across Notion, Google Drive, Linear, Confluence, and four internal dashboards. The team's NPS had cratered and the cost of every new hire's ramp was climbing.

We refused to ship a chatbot nobody would open. The product had to live in Slack, where the questions were already being asked, and it had to cite the source document in the same message as the answer. Anything else would not be trusted.

Pinecone holds a vector index refreshed every 15 minutes from the 14 connected sources. A Cloudflare Worker runs the retrieval and orchestrates a small agent loop: classify the question, fetch the right slice, draft an answer, cite the source, send it back to Slack. Permissions are enforced at retrieval — you only see what you could already see.

The team got 30% of their week back. The bot resolves 78% of questions without human escalation, median answer latency sits at 1.2 seconds, and the internal NPS for the tool is +62 — the highest of any internal system the company has ever shipped.