Use filters when a project has more traces than you can review manually. Filters narrow the table to the traffic that matters for debugging, evaluation, and improvement work.
Common filters
Use the filter controls to combine:
- Time range.
- Trace or span status.
- Latency and cost ranges.
- Token ranges.
- Tags.
- Attributes.
- Prompt.
- Behavior, when you arrive with Behavior context.
- Session or user identifiers when your integration sends them.
- Evaluator score or pass/fail signals when continuous evaluation is configured.
Filters are most useful when your instrumentation attaches meaningful tags and attributes, such as environment, release, customer segment, route, region, or feature flag.
If those fields are missing, use Instrumentation metadata to update the integration before relying on filters for incident review.
Search workflow
- Start with a time range that matches the incident or release window.
- Add one high-signal filter, such as a tag, attribute, status, or evaluator result.
- Sort or scan the table for latency, cost, or token outliers.
- Open representative traces and inspect the spans.
- Save useful examples into a dataset or use them as context for an Improve cycle.
Filter design tips
Use filters in layers:
| Layer | Example | Why it helps |
|---|
| Window | Last 24 hours, last week, release window | Keeps analysis tied to the period where behavior changed. |
| Segment | Environment, route, customer tier, feature flag | Prevents unrelated traffic from hiding the issue. |
| Object | Prompt, Behavior, tool, evaluator | Connects trace evidence to the platform object you can change. |
| Outcome | Error status, high latency, low eval score | Focuses review on traces that need action. |
Exporting traces
Use export when you need to share filtered trace evidence outside the app or analyze the results in another system. Before exporting, remove filters that are not part of the question you are trying to answer.
Trace exports can contain user content and application metadata. Share them only with people who should have access to that production data.
Create reusable evidence
After filtering, turn important traces into reusable material:
- Add representative cases to a dataset.
- Add metadata patterns to instrumentation standards.
- Create or update evaluators for the failure mode.
- Link the trace to a Behavior investigation.
- Start an Improve cycle when the fix belongs in the prompt.
See Add spans to datasets when a model span should become a regression case.
When filters are not enough
Use Deep search when you know the meaning of the message you want, but not the exact tags, attributes, or words. 
