Skip to main content

​
Explore the logs

The Logs explorer (/logs) is where you investigate production behavior. It does not copy your trace payloads — every row is derived from the authoritative trace, span, model-call, cost, and score records.

​
Step 1 — Start with a bounded query

Open /logs. The query bar accepts one clause per field, joined by AND. Start narrow:
A screenshot of the logs explorer with a bounded query entered. Common fields you will use:
Use free text (text ~ ...) after structured filters have narrowed the result set. Free text first is slow and broad.

​
Step 2 — Group the results

Use Group by to cluster by status, provider, model, failure category, alert state, or tag. You can compare count, known cost, average duration, p95 model latency, and average score without leaving the query. A screenshot of the results grouped by failure category.

​
Step 3 — Sort and page

Pick one cursor-safe sort (for example, by created descending). Pages contain at most 100 traces; use Load more to continue. Detail collections (spans, tools, scores) continue independently after 100 records.

​
Step 4 — Act on what you found

Select traces and:
  • Promote to case — turn a real failure into a test case.
  • Open in Playground — reproduce the failure against a prompt variant.
  • Create review task — hand the trace to a teammate.
  • Create draft monitor — start a Continuous Eval monitor from this query.
  • Export — download a bounded, redacted CSV or JSONL file.

​
Limits to know

  • At most 20 structured filters per query.
  • One primary cursor-safe sort.
  • Exports stop at 5,000 rows or 5 MiB.
  • Raw payload export is a privileged, audited operation.

​
Up next

Inspect a single trace in detail.