Returns
Returns a dictionary containing retry results:retried(int): Number of evaluations successfully queued for retry. A human-evaluation row counts here too: it is sent back to the annotation queue rather than to the evaluator.skipped(int): Number of evaluations not retried because they are not retryable, are over the retry limit, still miss the data their metric needs, or the caller’s organization lacks credits.errorsnames the cause.errors(list[str]): One message per evaluation that was not retried, naming why.
A single call can mix retryable and non-retryable IDs. Each row is judged on its own, except the
credit check: the caller’s organization must fund every evaluator row in the call, or none of them
retries. A row that cannot be retried is counted in
skipped with its reason in errors, and every
other row still retries normally.Example
Parameters
Single evaluation ID or list of evaluation IDs to retry. An evaluation can be retried only when its
canRetry field is true, which only a FAILED or SKIPPED evaluation carries. An evaluation of a metric with source: "self_hosted" is never retried, whatever its status: send its score again in a new evaluation instead. A retry re-scores the output already stored on the session, so it cannot fix a skip whose cause is that output.Retry Limits
Non-admin users are limited to a fixed number of retries per evaluation (default: 5). Once exceeded, the evaluation cannot be retried by regular users. For additional support or to increase retry limits, contact support@galtea.ai.