How it works

• Widget: the visitor asks in mango.js, optionally pointing at part of the page.
• Gateway: the backend checks the project's public key against its domain allowlist, rate limits, and the daily token budget.
• Retrieval: the question is embedded and matched against published facts, project layer first, then company layer.
• Prompt: grounding rules, then the project persona, then the retrieved facts with their sources and verified dates.
• Model: Gemini Flash-Lite through a router with fallbacks generates the reply and any tool calls.
• Actions: tool calls are validated, then hand off, capture a lead, show a CTA, log a gap or link the source page.
• Delivery: the reply streams back over WebSocket, and the inbox sees it live.

Retrieval scores every published fact against the question. Facts above the confidence threshold go to the model. If none clear it, the agent says it doesn't know and logs the gap. Drag the threshold to see the effect.

Handoffs, lead capture and CTAs happen through validated tool calls instead of regex rules.

The conversation switches to waiting. Every online teammate with access gets a push, and the first to claim it answers live. Hand back to AI returns it to the agent.

Unanswered questions, thumbs-down replies and unclaimed handoffs land in the review queue. A teammate's correction becomes a knowledge draft, gets published, and the agent answers it next time. The weekly report shows whether the answer rate is rising.