History and Preview
Chat History
History holds every conversation the assistant has had, with the full transcript. It is the fastest way to find out what customers actually ask and where the assistant falls short.
The list shows one row per conversation:
| Column | Description |
|---|---|
| Visitor | The visitor, identified as a customer where the storefront knew who they were. |
| Date | When the conversation started. Sortable, newest first by default. |
| Location | The country the visitor connected from. |
| Pages | How many storefront pages the visitor had open during the conversation. |
| Messages | Messages exchanged. |
| Tokens | Tokens the conversation consumed - in effect, what it cost you. |
| Duration | How long the conversation lasted. |
Filters
Conversations can be narrowed by:
- Created from and Created to - a date range.
- Pages at least / at most.
- Messages at least / at most.
- Tokens at least / at most.
- Location - one or more countries.
The token filters are the useful ones for cost work. Filtering on Tokens at least a high value pulls up the handful of long, expensive conversations, and reading those transcripts usually shows a concrete cause: a question the knowledge base cannot answer, so the assistant keeps searching, or a loop that a Guidance rule should have ended with an escalation. See Costs and billing.
Chat details
Opening a conversation shows the full transcript, including any files the visitor attached, which can be downloaded.
Correcting an AI answer
Reading past conversations is where you find the assistant's mistakes, so it is also where you can fix them. Every AI message in Chat details carries a Correct action. Visitor, agent and system messages do not - there is nothing to correct in what the customer said.
Correct opens a form over the conversation, with the customer's own question already filled in. You choose what kind of mistake it was:
- A fact - the assistant said something untrue. Saved as a FAQ entry: a question and the answer it should have given. This is the default, because a wrong answer is almost always a wrong fact.
- How it behaves - the assistant should have acted differently. Saved as a Guidance rule, with the same Name, Trigger, Instructions and Status fields as the Guidance screen.
Switching between the two keeps whatever you have already typed, so changing your mind costs nothing. Save stays unavailable until the fields that matter have text, and Cancel saves nothing.
The marker
Once saved, the message shows what came out of it - FAQ item added or Guidance rule added - with a link to the record in its own editor. The marker is read from the record itself, so it is there after a reload, it is there for your colleagues, and correcting the same message twice leaves two markers rather than replacing the first.
Deleting the record removes its marker, because the marker was only ever the record's shadow.
A correction changes what the assistant says next time. It does not rewrite the conversation the customer already had, and the customer is not notified.
Preview
Preview runs the real chat widget against your real workspace, in the backend, without touching the storefront. It uses your live settings, your live knowledge base, and your live provider chain.
Use it to:
- Check a Guidance rule fires on the phrasing you expect.
- Read how a change to Instructions or Tone of Voice actually sounds.
- Confirm a newly synced product is answerable before customers ask.
- Try a provider or model change before it reaches customers.
Preview conversations are real conversations: they consume tokens on your provider account and they appear in History.