Feedback

Feedback lets human reviewers score outputs, attach corrections, and promote approved corrections into a new dataset version.

Endpoints

POST /v1/feedback

Submit feedback for an inference log.

Key request fields:

  • inference_log_id
  • score of -1, 0, or 1
  • optional correction

Corrections are validated against the task output schema.

GET /v1/feedback

List feedback items, optionally filtered by:

  • status
  • task_id

PUT /v1/feedback/:id/status

Move feedback through the review lifecycle.

Supported public transitions:

  • pending -> staged
  • pending -> rejected
  • staged -> rejected

POST /v1/tasks/:task_id/feedback/promote

Promote staged corrections into a new dataset version.

Notes

  • Only feedback with valid corrections can be staged.
  • Promotion creates a new dataset automatically and marks the promoted feedback rows accordingly.