Feedback

End-user thumbs up / down feedback, dislike categories, and the most recent comments.

Where the analysis layer measures quality with an LLM-as-judge, the Feedback page brings in the other half of the picture: what users actually thought of your agents’ replies, captured as a like/dislike vote and any optional comment they left.

It exists in two scopes:

• Organization Feedback (sidebar entry at the org level): every agent in the org.
• Agent Feedback (sidebar entry at the agent level, simply labeled “Feedback”): only the selected agent.

The two views share the same layout. The agent-scoped variant hides the Agent column.

Period selector

The date filter in the header scopes the data on the page: presets of 7 days, 30 days (the default) or 90 days, or a custom range of UTC days. The organization Feedback page and an agent’s Feedback tab share the same window. A refresh button forces a re-fetch.

Distribution cards

Four KPI cards:

Card	Meaning
Total	All feedback items submitted in the period.
Likes	Thumbs-up count.
Dislikes	Thumbs-down count.
Like Rate	Likes ÷ Total, as a percentage.

Dislike categories

When users dislike a reply, they can optionally tag the reason. The categories card shows progress bars for each:

Category	What it means
Incorrect / Incomplete	The reply was wrong or missed key information.
Not what I asked	The agent answered a different question.
Slow / Buggy	Response took too long, errored, or felt broken.
Style / Tone	Phrasing felt off: too formal, too casual, robotic, etc.
Safety / Legal	The reply raised a safety or compliance concern.
Other	Reasons that don’t fit the categories above.

The card is hidden when there are no dislikes in the period.

Recent feedback table

The most recent feedback items, paginated with a See more button. Columns:

Column	Notes
Vote	Thumbs-up / thumbs-down icon.
Agent	Hidden in agent scope.
Comment	The user’s free-text comment, or “No comment” if empty.
Category	The dislike category badge (likes have none).
Date	A relative time-ago label.

On mobile the table collapses to a card layout with the same data.

When this page is empty

If no feedback has been collected, the page shows: “No feedback collected yet.” You can:

• Make sure the agent’s UI exposes thumbs-up / thumbs-down; Chat and the Builder UI components do this by default.
• Encourage users with prompts (“Was this helpful?”) inside the agent’s reply.

What to use this for

• Quality triage. A surge in the Style/Tone category often points at a system prompt change. The Safety/Legal category should be reviewed daily; every entry is a candidate audit case.
• Compare against LLM-as-judge. If your evaluation scores are high but the like rate is low, your criteria don’t align with what users actually value.
• Per-agent diffing. Compare a known-good agent against a struggling one in the dislike category mix.

Where to go next

Read the per-conversation analysis

Cross-reference disliked conversations with the LLM-as-judge result.

Adoption

See like rate aggregated as part of the broader quality picture.