Trace explorer

Trace explorer (sidebar: Trace explorer, shortcut g t) is the searchable table of every trace your SDK has sent. Use it to triage failures and open individual traces.

Filter by failure cell

Chips across the top filter to any of the six failure cells — complete · grounded, complete · ungrounded, incomplete · grounded, incomplete · ungrounded, extra · grounded, extra · ungrounded. Combine chips to see, say, only hallucinations (the ungrounded cells).

Status

Each trace is one of:

• evaluating — the trace arrived; the judges haven't finished (this clears within seconds; see Troubleshooting if it doesn't).
• evaluated — graded, with a failure cell assigned.

Sort & columns

Sort by newest, or by lowest sufficiency / faithfulness to surface the worst answers first. Rows show the trace id, the query, its failure cell, sufficiency/faithfulness scores, latency, and age. Honest abstentions are marked distinctly — a correct "I don't know" is not a failure.

Re-evaluate a trace

Open a trace and use Re-evaluate (also on the Trace detail page) to drop the existing evaluation and re-run the judges — useful after a stuck eval, or to re-grade with updated logic. The row returns to evaluating and updates on its own when grading finishes.

Stuck on "evaluating"?

A trace reads evaluating whenever it has no evaluation yet. This is almost always temporary — a background evaluator picks up stranded traces automatically. It can persist if your LLM provider credit is exhausted (the eval can't run). Top up credit, then Re-evaluate the trace (or the backend re-sweeps stranded traces after a restart). See Troubleshooting for more.