Skip to main content
Two feedback surfaces, both authenticated by the trigger secret for that Flow. They mean different things — don’t mix them.

​
feedback(traceId, {...})

Rating the answer itself: 👍 / 👎, an optional comment, or a before → after correction. This is the existing end-user feedback endpoint, exposed here for your app to forward.
  • A dislike promotes into a confirmed human signal on that trace.
  • A like retracts any pending signals attributed to that trace.
Aliased as traces.feedback — same call, spec-canonical name.
Options:
  • rating (required) — "like" or "dislike".
  • comment — free-text note; surfaces on the trace and in cluster review.
  • correction — { before, after } pair. The after is what the user says the answer should have been. Fed into the improvement loop as candidate fix content.

​
signals.report(traceId, {...})

An external outcome the answer-quality detectors can’t see: refunded, reopened, escalated to human, resolved. These land as candidate signals with origin=sdk_report and go through the same cluster / dataset path as detected signals.
Options:
  • outcome (required) — free-form string. Common vocab: resolved, reopened, escalated, refunded, abandoned. Pick a set and stick to it — it’s how your clusters group.
  • reason — free-text note.

​
Why keep the two separate

Mixing feedback and signals.report makes detectors and clusters noisy:
  • feedback speaks to whether the answer was right. It weighs into answer-quality metrics.
  • signals.report speaks to what actually happened downstream. It weighs into outcome metrics and can surface failures the answer-quality detectors miss.
An answer can be technically correct (👍 feedback) but the user still churns (escalated signal). Both should land, on different tracks.

​
Where the traceId comes from

  • From flows.run — r.traceId on the return value. Preferred: you have it inline.
  • From your own logs — if you captured traceId earlier and are now posting feedback on it, that also works.
  • Correlation with newTraceRef — for pre-run correlation across your own logs.

​
MCP / CLI equivalents

The MCP tool and CLI surface the same operation:
See nora feedback for the full CLI shape.