Trace detail

The Trace detail page opens one trace in full. Reach it by clicking any row in the Trace explorer.

Diagnosis

If the trace is evaluated, the top of the page shows the diagnosis: the assigned failure cell, whether it was an honest abstention, and the judge scores. If it's still evaluating, the page updates itself once grading finishes — no refresh needed.

Query & Response

The two panes show exactly what your system did:

• Query — the user's question, annotated with the sufficiency score (did retrieval surface what was needed?).
• Response — the generated answer, annotated with the faithfulness score (did it stay grounded, or hallucinate?).

Retrieved context

The retrieved chunks you sent as context, in order — so you can see exactly what the model had to work with when it answered. This is where a retrieval gap (incomplete_grounded) or a hallucination (*_ungrounded) becomes obvious.

Actions

• Re-evaluate — drop the current evaluation and re-run the judges (see Trace explorer → Re-evaluate).
• Copy as Markdown — copy the full trace (query, context, response, diagnosis) as Markdown, e.g. to paste into an issue or a chat with your coding agent.