These docs describe the current Dante AI platform. On the earlier platform? Your API reference is api-v2.dante-ai.com/openapi.json, and your key goes in the x-api-key header.
Chat history
Every conversation your AI agent has is saved here so you can review what visitors ask, spot gaps in your knowledge, and improve your answers over time.
Open your dashboard, then pick an agent and go to Monitor → Chat history →
What you can see on each plan
Conversations from tests you run inside Dante AI are always visible; they appear with the Sandbox channel. Conversations from your live, published agent require a paid plan.
| Plan | What shows in Chat history |
|---|---|
| Free | Preview only. Just the chats you ran inside Dante AI, shown as Sandbox, with their full transcripts readable. Export is locked. |
| Starter and up | Every real visitor conversation from your published agent across all channels, plus export. |
On Free the page shows a Preview label. You can always open and read your own Sandbox test chats, and the widget preview picks up where you left off. Upgrade to Starter to unlock real visitor chats and exporting.
Example conversations on the Free plan
On Free you will also see a handful of example conversations in the list, sitting alongside your own Sandbox chats. They are there to show what a live agent's inbox looks like once real visitors start talking to it: the kinds of questions people actually ask, and how your agent answers them.
They are not your conversations and they are not from your agent. Every example is marked so you can tell at a glance:
- An Example badge on the row.
- A greyed row with a dashed outline instead of the coloured circle real visitors get.
- The visitor is called Example visitor, never a person's name.
- Opening one says "Example conversation. Not from your agent." at the top.
They disappear as soon as you upgrade, because from then on the list fills with the real thing. Nothing you do to an example affects your account, and they never count toward your credits.
One example carries a thumbs-down reply so the Needs review filter and the Revise answer button have something to show. Clicking Revise answer on it shows a short Example conversation notice instead of opening the revise box, because there is nothing of yours to retrain. The button works as normal on your own conversations.
The conversation list
The left panel lists one row per conversation. Each row shows:
- Visitor, the name they gave through lead capture when available, otherwise a stable label like User 0001 or Visitor a1b2c3d4.
- Title, a short summary of the conversation, or Untitled conversation if none has been generated.
- Status, one of Active, Idle, Closed, or Live.
- Channel and message count, the source of the chat and how many messages it holds.
- Last activity, when the most recent message arrived.
- Satisfaction, an emoji rating if the visitor rated the chat.
- To revise, an amber badge counting thumbs-down replies you haven't revised yet.
Channels are Widget, Sandbox, API, WhatsApp, Slack, and Email. Use the Search box at the top of the list to filter by visitor, title, or channel, and the Needs review pill next to it to show only conversations with thumbs-down replies waiting to be revised. Click any row to open its full transcript on the right.
Reading a transcript
The transcript shows the visitor's messages and your agent's replies in order. When the agent looked something up, a Knowledge search step appears in line. Open its Details to see exactly which sources matched, each with a match score and a link to the original page or document, useful for spotting where the agent got something wrong and which source needs an update.
Individual replies can also carry:
- Credits charged, a small badge showing how many credits the reply used.
- Thumbs up or down, feedback left on the reply (read-only here).
- Live agent replies, messages a human teammate sent when they took over the chat.
A reply your agent did not write
Very rarely, a transcript will show a reply your agent did not produce and that you cannot configure. It looks like this: the visitor is told plainly that the assistant is automated and that this is not something it can help with, and nothing further is shown.
This happens when a visitor's message plainly signals self-harm or suicidal intent. In that case we do not ask the AI anything at all. A fixed reply is sent instead, the same on every agent on Dante, and the conversation is not continued or escalated. Your instructions, your knowledge and your chosen model are not involved.
The reply declines and stops there. It does not link to a helpline, name an emergency number, or offer to pass the visitor to a person. That is deliberate: showing a resource is a way of engaging with the subject, and an offer to reach a human is one we cannot answer at every hour of the day. A reply nobody comes to would be worse than a clear no.
It is deliberately narrow, so it only catches unambiguous phrasings and will miss others. It is not a moderation system and should not be relied on as one. It cannot be turned off, on any plan.
We record only that it happened. The message itself is never written to our logs. It is still stored in your transcript, because that is your conversation.
Revising it will not change it. You can leave feedback on the reply like any other, but the check runs before your agent looks anything up, so a Q&A pair cannot replace it.
Revising a bad reply
When a reply has a thumbs-down, a Revise answer button appears next to it. Click it, and a box opens with the visitor's question and the reply the AI gave, both editable. Edit the answer into what the AI should have said (and add an optional internal note), then save. That adds the Q&A pair to the agent's knowledge and retrains it so it answers correctly next time, and the reply shows a Revised badge. On an example conversation (Free plan) the button shows a short notice instead of opening the box, because the reply is not from your agent.
Exporting
- One conversation, the Export button in the transcript header offers JSON, CSV, or emailing a formatted copy of the transcript to any address.
- The whole list, the Export all button offers JSON, CSV, Markdown, or HTML (HTML prints cleanly to PDF). Export all covers the list rows, not full transcripts, so use the per-conversation Export when you need the messages.
Exporting is a paid feature. On Free, clicking Export opens an upgrade prompt instead.
Deleting a conversation
Open a transcript, click the More menu in the header, and choose Delete. This permanently removes the conversation and its full transcript and cannot be undone, so you are asked to confirm first.
Still stuck? We'll help.
Every plan, including Free, comes with human support. If something isn't behaving the way you expect after following the steps above, send us a note and we'll get back to you promptly.