Troubleshooting

Resolve common setup, wiring, execution, and review problems.

“Nothing binds input Space …”

The version reads a Space that is neither a session source, an earlier task output, nor a registered workspace Space. Bind the external dataset for this run, add the producing task, or register the Space.

A structured response is repeatedly cut off

Reduce batch size or increase max_output_tokens. The latter is a per-call ceiling, not a task-wide token budget.

Too few calls appear in session metrics

Open the task breakdown. Deterministic gates may eliminate candidate pairs before semantic work, prior results may be reused, or the displayed aggregate may cover only one stage. Failed attempts and cached attempts should still appear in task observability.

A reviewer produced no verdicts

Inspect failed calls and the checkpoint's model response status. Common causes are a response ceiling that is too small for the requested sample or a prompt/schema mismatch. Unreviewed samples are excluded from quality counts rather than silently treated as correct.

A page is slow while improvement work runs

Opening a version or report should not wait for background improvement. If it does, record the page, pipeline, version, and approximate time; this is a product defect rather than expected behavior.

A scheduled session did not run

Check that the version is declared, the deployment is enabled, the timezone and start date are correct, required sources still resolve, and no per-run budget was exhausted.

Get help

For a restricted-beta problem, include the pipeline ID, version ID, session ID, and the visible error. Never send passwords, database URLs, model keys, or one-time Relifold API-key secrets.