Feedback

When someone rates an answer, that rating lands in the Feedback tab together with the question, any comment they left, and the conversation it came from. From there you can recompute the answer with your assistant's current data, compare it against the one that was rated, and turn a bad answer into a Q&A so it never happens again.

The Feedback tab of an AI assistant

Opening the Feedback tab

Open the assistant, go to Analytics and pick the Feedback tab. The other tab is Logs, which holds the conversations these ratings came from.

The Feedback tab of the Analytics page
Feedback and Logs are now one page
In the previous design Feedback was a separate entry in the assistant menu. It is now the second tab of the Analytics page, so you can jump between a rating and the conversation behind it without leaving the page.

How feedback is collected

You do not create feedback yourself — users do, from the chat:

  1. Under every answer, the assistant shows a thumbs-up and a thumbs-down button.
  2. Clicking one records the rating straight away. Clicking the same button again withdraws it.
  3. Users can add a written comment explaining what was wrong, which shows up in the Comment column.
No storage, no feedback
The rating buttons only appear when the assistant is recording conversations. If storage is set to Discard all, users never see them and this tab stays empty. See Storage and visibility.
The thumbs up and thumbs down buttons under an answer

The feedback table

One row per rated answer:

  • Questionwhat the user asked.
  • Commentthe note the user left, or a dash when there is none. Long comments are truncated — hover to read the whole thing.
  • Feedbacka green thumbs-up or a red thumbs-down. A small globe icon marks feedback that arrived from an external source rather than a signed-in user.
  • Actionsrecompute this answer, delete the feedback, and create a Q&A from it.
  • Similarity scorethe last, unlabelled column. After a recompute it shows a ring with a percentage — how close the new answer is to the one that was rated.

With nothing rated yet the table shows "No Data Found".

The feedback table

Reading the row colours

Rows are tinted so you can scan the table for problems rather than reading every row:

RowWhat it means
Greenpositive feedback, and a recompute produced an answer at least 75% similar — the assistant still answers this well.
Amberpositive feedback with a similarity between 50% and 75% — the answer has drifted somewhat since it was praised.
Redpositive feedback with a similarity below 50% — the assistant no longer gives the answer the user liked. Worth investigating.
Red (thumbs-down)negative feedback, regardless of any score.
Feedback rows with their similarity scores
The score only applies to positive feedback
A similarity score compares today's answer against an answer a user approved of, so it is only computed for thumbs-up rows. A thumbs-down row is red because the user said it was wrong, not because of any measurement. If a recompute fails, the column shows "Error" instead of a ring.

Recomputing answers

Recomputing re-asks the question using the assistant's current data, prompt and model, so you can tell whether a change you made fixed or broke something:

  • One rowthe circular refresh button in the Actions column recomputes just that answer.
  • Recompute all answersthe button below the table re-runs every question that has feedback. QAnswer describes it as recomputing "the answer for all questions where feedback was given, using the current data and settings of the assistant."

Recomputation runs in the background and the table updates live as each answer comes back — a spinner marks the rows still in flight, and the buttons stay disabled until the run finishes.

Tip
Recompute after you add documents, change the prompt or switch the language model. The similarity scores then tell you at a glance whether the answers users liked survived the change.
The Recompute all answers button

The conversation panel

Clicking a row opens the original conversation underneath the table, in read-only form:

  • Old and Currentonce an answer has been recomputed, both versions are shown side by side so you can judge the difference yourself.
  • Prefer this answeraccepts the recomputed answer as the better one; it is then marked "Answer accepted".
  • The refresh buttonrecomputes just this conversation's answer without going back to the table.
  • The ✕ buttoncloses the panel and returns the table to full height.
The conversation panel below the feedback table

Turning feedback into a Q&A

The most useful thing you can do with a bad answer is make sure it is answered correctly next time. Create a Q&A does exactly that:

  1. Click the Create a Q&A action on the row.
  2. The question is filled in for you; write the answer you wanted.
  3. Pick the Source type — Q&A, Document or Website — and, for the last two, point at the source document or URL.
  4. Save. The entry joins the assistant's data and is used from then on.
Info
The entries you create here land in the assistant's Q&A data source, where you can edit them later — see Q&A.
The Create a Q&A dialog

Exporting and deleting

  • Exportthe Export button above the table offers Download as CSV, which saves every feedback row to a file, and Print.
  • Deletethe red bin on a row removes that piece of feedback. It removes the rating, not the conversation — the conversation stays in the Logs tab.
The feedback export menu

Reviewing feedback as an administrator

Feedback shown here belongs to a single assistant. Organization administrators additionally get a consolidated view covering every assistant in the organization, which is the better starting point when you are looking after many of them at once.

Info
See Conversation feedback for the organization-wide view.

What changed from the previous design

If you knew the old Feedback page, here is what moved:

  • Feedback is no longer its own entry in the assistant menu — it is the second tab of the Analytics page, next to Logs.
  • Row colours are now driven by the similarity score, with distinct green, amber and red bands, instead of only marking positive and negative.
  • The similarity score is shown as a percentage ring in its own column, and recomputation streams its progress into the table live rather than requiring a reload.
  • The selected conversation opens in a panel below the table with the old and current answers side by side, rather than on a separate screen.