Read your first diagnosis
What the trace page tells you, from the top: the question, the verdict, the findings for each stage, and the execution.
Open any trace in the dashboard. The page reads from the top down, from what went wrong to the evidence for it.
The verdict#
Under the question you'll see the probable cause, for example Retrieval miss, with
the reason in a sentence. When the LLM judge decided, the reason is the judge's own
words; for an unsupported claim it quotes the claim the context doesn't back, such as
Unsupported: "On the request path, middlewares run from innermost to outermost."
Next to it is a mark that says who decided:
- Final: the LLM judge's verdict.
- Preliminary: an estimate from term overlap or embedding similarity, used until the judge answers, or when no judge is configured.
Estimates and judge verdicts explains the difference and why it matters.
Findings by stage#
Three columns, one per stage: Retrieval (did the context match the question?), Grounding (is every claim supported by that context?) and Answer (does it address the question?). Each shows the check's result, its reason, and who decided. The column the diagnosis blames is tinted red and tagged "Probable cause". Selecting a column opens that step below.
The execution#
Every step in order, drawn to time where the SDK measured it: the query, retrieval, the prompt, generation, evaluation, feedback and the raw trace. Select a step to see what went in and what came out: for retrieval, each chunk with its rank, source and score; for generation, the model, parameters, tokens and cost.
What to do next#
- If the retrieval column failed, look at the chunks: was the right document there?
- If grounding failed, compare the unsupported claim with the chunks.
- If it's part of a recovery run, the strip at the top shows every attempt and what changed between them.
Investigate failures covers going from one trace to the fix.