Skip to main content
Every agent answer carries a Was this helpful? prompt with a thumbs up and a thumbs down. It’s the cheapest signal you will get from your users, and it lands directly next to the automated evaluations in Chat History.
The action row beneath an agent answer, with a Was this helpful? prompt and thumbs up and thumbs down buttons

​
Thumbs up

Clicking thumbs up records the rating immediately — no dialog, no extra step. Clicking it again clears the rating, so a mis-click costs nothing.

​
Thumbs down

Thumbs down opens a short Provide feedback dialog asking “What was wrong with this response?”. The user can pick one of five common causes:
  • Wrong chart type
  • Incorrect SQL query
  • Wrong data selected
  • Doesn’t answer my question
  • Inaccurate analysis
…or type their own explanation in the free-text box, which takes precedence over the selected option. Either way the rating and the comment are submitted together.
The Provide feedback dialog asking what was wrong with the response, with five preset causes and a free-text box
The five preset options aren’t cosmetic — they map onto the stages of the agent’s pipeline, so a cluster of “Incorrect SQL query” reports points you at System Prompts and Golden Queries, while a cluster of “Wrong chart type” points at Golden Charts.

​
Feedback is per answer, not per conversation

The rating attaches to the specific exchange the user was looking at. A conversation where the third answer went wrong keeps that precision:
  • In the Traces view of Chat History, the thumbs down sits on the row for that exchange.
  • In the Sessions view, the session’s Feedback column reflects the feedback given in that conversation, and the detail sidebar shows the rating and comment.

​
Where it shows up for you

Chat History — the Feedback column in both the Sessions and Traces views. Hover a badge to read the written comment. You can filter the table on feedback to pull up every negative answer in a time range.
The Chat History table filtered to thumbs-down answers, with a feedback badge hovered to show the user's written comment
The detail sidebar — when you open a conversation, the metadata card carries Feedback and the user’s Comment verbatim. In the thread itself — a rated answer renders a read-only card beneath it saying “User found this helpful” or “User found this unhelpful”, along with their comment, so you see the rating in the context of the answer that earned it. Alongside evaluations — a thumbs down on an answer that every automated evaluation passed is the most interesting row in the table. It means the answer was defensible and the user still didn’t get what they needed, which is usually a context problem rather than a correctness one.

​
Turning feedback into a fix

The workflow that works:
  1. Filter Chat History’s Traces view to thumbs-down answers over the last week.
  2. Open each one and read the user’s comment, then switch to the Observability tab.
  3. Find the step that went wrong — usually SQL Generation or Chart Generation.
  4. Fix it at the source: a System Prompt rule if it should always apply, a Golden Query or Chart if it’s a recurring question with one right answer.

​
Next steps

Sharing an Analysis

Let users send a single answer or a whole conversation to a colleague.

Agent Observability

Trace the answer behind a thumbs down, step by step.