Evaluation engine

Start with a quality signal that answers a real operating question: did the answer follow policy, use the right evidence, or complete the requested task?

client.feedback(runId, true, { kind: "thumb" });
client.score(runId, "groundedness", 0.92, { dataType: "numeric" });

See Feedback and scores for the full set of options, including how to get the run id.

Use production traces to build datasets, run prompt or model candidates against the same examples, and compare quality, cost, latency, and failure rate before promotion. Online evaluators run asynchronously so ingestion remains independent of scoring.