Support (conversations & tickets)
The Support pages — monitoring every AI conversation, taking over as a human, replying live, escalating to a ticket, and working the ticket queue. Includes how the AI first-response is grounded.
Support (conversations & tickets)
Support in god mode is two pages: Conversations (/support/chat) and Tickets
(/support/tickets). Both require the support.manage privilege; without it each shows
a locked state. See the god-mode overview for how the gate works. Both
stores are dual-backend — Postgres when DATABASE_URL is set, a seeded in-memory mock
otherwise — so on the demo backend the rows are seed data, not live customers.
The AI first-response (how a conversation starts)
Before an operator ever sees it, a customer conversation is handled by the AI. The
public chat endpoint (/api/chat) is what powers the customer-facing widget:
- Grounded on the docs. Each answer is grounded on a static content index built from the docs and blog. A lightweight keyword-overlap retriever (with heading/title boosts) finds the most relevant chunks and cites the doc or blog the answer came from — the same shape a Postgres/pgvector query would return later.
- Claude when configured, graceful fallback otherwise. When
ANTHROPIC_API_KEYis set, the endpoint streams a real conversational answer from Claude (theSystem Checkpage reports this as "Fleet Copilot"). When it isn't, it streams a doc-grounded fallback so the experience still works in any environment. - Silent under human control. When a human has taken over (mode
human_active), the AI stays silent — the customer's message is simply delivered to the operator, who replies live. - Hardened. The endpoint is public and unauthenticated, so it's rate-limited per IP with a global circuit-breaker, rejects cross-site origins, and redacts anything that looks like a key or secret out of the message before it's stored or sent.
Conversations
The Conversations page is where you watch every AI conversation across the product, take over as a human, reply, and escalate. A KPI row counts:
| Stat | Meaning |
|---|---|
| Conversations | Total in the store |
| Awaiting human | Mode human_requested — a visitor asked for a person (accented) |
| You're handling | Mode human_active — an operator is driving |
| Resolved | Status resolved |
A conversation moves through these modes: ai (the assistant is answering),
human_requested (the visitor asked for a person), and human_active (an operator has
taken over). The workflow actions on a conversation:
| Action | What it does |
|---|---|
| Join / take over | Sets the conversation to human_active under your name, and posts a system note that you joined. The AI goes silent. |
| Reply | Sends a live message as the admin while you're in control. |
| Escalate to ticket | Creates a support ticket and copies the transcript, then posts a system note linking the new ticket. |
| Resolve | Posts a "resolved" system note and closes the conversation (handing it back). |
Every conversation carries its category (e.g. sales, bug), the contact if the visitor left one, the assigned admin, and its message history — cited sources included on AI messages.
Tickets
The Tickets page is the inbound support queue — including chats escalated from Conversations, with the transcript attached. A KPI row counts Open (accented), In progress, High / urgent (by priority, accented), and Resolved / closed.
Working a ticket:
| Action | What it does |
|---|---|
| Reply | Posts a message on the ticket. It can be a customer-facing reply or an internal note (staff-only). |
| Change status | Moves the ticket through open → in progress → resolved / closed. |
Tickets carry a subject, description, priority, category, requester, the assigned admin, and — when escalated from a chat — a link back to the source conversation.
Related
- God mode (admin overview) — the RBAC gate and demo-vs-persisted rules
- Fleet Copilot — the operator-side grounded assistant (distinct from the customer chat)
- Settings & system — where the Claude API key and chat wiring show up in the health checks