Choose a time window and evaluate recorded chat answers
Chat answers and generated reports are evaluated. Task and other non-chat traces are excluded before selection. Failed chats with no answer receive an automatic zero. Missing archives or unknown interaction types remain skips. Complete inputs are retained; inputs beyond the 500 KB evaluation limit are skipped.
Langfuse runs the selected judges. Each answer produces one score per selected dimension. The evaluator, model and input versions are recorded with the run.