Skip to main content
When an answer isn’t right, a thumbs-down does more than register a complaint. If you have builder access, ClarityQ can diagnose what went wrong — investigating the conversation that produced the answer against your Context Layer — and suggest a concrete fix you can apply on the spot. Diagnosis is aimed at the people who maintain the model. Anyone can rate an answer, but the diagnosis flow appears only for users with builder access. Other users’ thumbs-down ratings are still recorded to help improve accuracy.
Diagnosis runs on regular Ask Anything chats. It isn’t available in Builder chats or in an embedded ClarityQ. A thumbs-down there records your feedback as normal.

​
Starting a Diagnosis

  1. Give an answer a thumbs-down in Ask Anything.
  2. In the feedback form, select the reasons that describe what went wrong (for example a logic error or a missing filter) and optionally add a note.
  3. Click Send and suggest fix.
That hands the answer to a dedicated diagnosis agent. The agent works in the background — it can take a little while, and you’re free to keep chatting while it runs. You’ll be notified when the results are ready to review.

​
What the Diagnosis Investigates

The diagnosis agent re-reads the conversation up to the answer you rated and inspects your Context Layer — the metrics, entities, and definitions the answer relied on. It checks whether those components exist and whether their definitions are correct, so it can tell a fixable modeling problem apart from a one-off reasoning error. The investigation is read-only. It looks at your Context Layer but never changes it — any fix is applied separately, through the Builder, only if you choose to.

​
What You Get Back

A diagnosis ends in one of three ways:
  • A fix to apply. The diagnosis explains what went wrong and recommends a change to your Context Layer. Click Fix now to jump into the Context Builder with the suggested fix pre-loaded, or Dismiss to leave things as they are.
  • Nothing to fix. The answer matched your Context Layer, so there’s no definition to change — the failure was a one-off. Your feedback is still recorded.
  • Couldn’t finish. If the investigation errors out, your feedback is saved and you can Try again.

​
Applying the Fix

Fix now doesn’t change anything on its own — it opens the Builder pre-loaded with the suggested correction so you can review it, then release it through your draft and deploy. See Context Layer Versioning for how changes are released. This works for versioned components — entities, metrics, segments, and entity dimensions — which are the components the Builder edits.

​
When the fix is on a feature

Features are managed live rather than through the Builder, so they can’t be fixed from a Builder prompt. When every part of a recommended fix targets a feature, the button changes to Fix manually and the panel notes that features are managed manually in the Semantic Catalog. Clicking it takes you to Semantic Catalog → Features, where you make the change on the feature itself. Nothing is pre-filled, so keep the diagnosis text in view as you go — it says exactly what to change. If a fix spans a feature and something versioned, you get the normal Fix now handoff into the Builder.

​
Diagnoses in Shared Conversations

When a conversation is shared, any answer that was diagnosed is marked with an indicator on its rating control. Opening it shows the diagnosis and its recommended fix read-only — there are no Fix now, Dismiss, or Try again controls, since the shared view isn’t the place to act on them. This lets you send someone the evidence of what went wrong, not just the bad answer.