Use message ratings¶
People can rate an agent's answers, and the ratings are readable two ways: on the turn itself, and in aggregate for whoever administers the deployment.
Rating messages¶
Like/dislike¶
Each AI assistant message displays two buttons:
- Like (👍) — Click to indicate the response was helpful
- Dislike (👎) — Click to indicate the response had issues
Toggle behavior¶
- Clicking the same button again removes your rating
- Clicking the opposite button changes your rating (like → dislike or vice versa)
- Only assistant messages can be rated (not your own messages)
Adding feedback¶
When you dislike a response, a dialog appears asking "What went wrong?"
You can optionally provide a comment (up to 2000 characters) explaining the issue. This feedback is valuable for understanding why responses weren't helpful.
Common reasons to dislike:
- Incorrect or hallucinated information
- Response didn't address the question
- Too verbose or too brief
- Poor formatting or structure
Rating counts¶
Each message shows the total number of likes and dislikes from all users. Your own rating is highlighted (green for like, red for dislike).
For administrators¶
Ratings dashboard¶
Navigate to Admin → Response Ratings (or /admin/ratings) to access the analytics dashboard.
Summary statistics¶
- Total ratings — All ratings across the system
- Likes — Count of positive ratings
- Dislikes — Count of negative ratings
- Average — Overall satisfaction score (-1.0 to 1.0)
Ratings chart¶
A bar chart shows ratings over the last 30 days. Green bars represent likes, red bars represent dislikes.
This page's window is fixed at 30 days
To read the same numbers over a period you choose, use the dashboard's Answer quality, deployment-wide card, which follows the period filter at the top of the page.
Filtering ratings¶
Use the filter dropdowns to narrow down results:
| Filter | Options |
|---|---|
| Rating type | All / Likes Only / Dislikes Only |
| Comments | All / With comments only |
Ratings table¶
The table shows individual ratings with:
- Date — When the rating was submitted
- Rating — 👍 Like or 👎 Dislike
- Comment — Feedback text (if provided)
- Message — Preview of the rated response
- User — Who submitted the rating
- Actions — Link to view the full conversation
Exporting data¶
Export ratings for external analysis:
- JSON — Full structured data, suitable for scripts and analysis tools
- CSV — Spreadsheet-compatible format for Excel or Google Sheets
Exports respect the current filters
Narrow to dislikes with comments, then export, and that is what you get - the filters are part of the query, not of the view.
Viewing conversations¶
Click "View conversation" on any rating to open the chat interface with that conversation loaded. This is useful for understanding the context of a rating.
Admin conversations page¶
Navigate to Admin → All Conversations (or /admin/conversations) to view all user conversations.
This page provides:
- Searchable list of all conversations
- Filter by user email or username
- Filter by date range (preset or custom)
- Quick links to view conversation details
Direct conversation links¶
You can share a direct link to a specific conversation by adding the id parameter to the chat URL:
This is useful for:
- Sharing conversation context with team members
- Bookmarking important conversations
- Linking from external tools or documentation
Click "View conversation" from any rating or the admin conversations list to open a direct link.
API access¶
For programmatic access to ratings data, use the admin API endpoints:
| Endpoint | Method | Description |
|---|---|---|
/admin/ratings |
GET | List ratings with pagination and filters |
/admin/ratings/summary |
GET | Aggregate statistics over from/to (inclusive UTC dates, defaulting to the last 30 days) |
/admin/ratings/export |
GET | Export ratings (JSON/CSV) |
/admin/conversations |
GET | List all conversations |
Every /admin endpoint is the deployment superadmin, not an org role
The gate is CurrentAppAdmin — a signed-in user whose users.is_app_admin is
true. There is no users.role column, and no organization role reaches these
routes. See
permissions.