Discovery & Tracking

Findings finds the specific behaviors in your traces that are worth attention: a failure, an accuracy gap, or a behavior you want to watch. Discovery and Tracking are the two tabs of Findings; Traces is its own page alongside them. The usual path is to run a Discovery search, pin what matters into Tracking, and open a single trace when you need the detail.

​

Discovery

When you don’t yet know exactly what you’re looking for, start with Discovery. Discovery is where you run investigations into your traces at scale. Describe in plain language what you want to look into, and Discovery scans your traces and groups related ones into findings, including patterns you didn’t name. Pin the interesting ones to keep watching them in Tracking. When you run a search you set the prompt, the model quality and the time range; the rest sits under Advanced:

• Discovery prompt: describe the behavior in plain words. It does not have to be a failure.
• Model quality: Cheapest and Balanced find simpler patterns at lower cost; High-quality catches issues that are hard to find.
• Time range: the window of traces to search, the last 30 days by default.
• Sampling %: the share of traces the search reads, 10% by default. Raise it for coverage, lower it to keep cost and time down.
• Trace-name regex: limit the search to traces whose name matches a pattern.
• Explanations: optionally get a short reason, with citations from the trace, for why each trace matched. Off by default since it costs more.

A Discovery is saved with its settings, so you can rerun it later. A rerun reads only traces that arrived since the last run and keeps its findings. If you change the prompt or any setting, the next run discards that Discovery’s findings and rebuilds them; pinned findings are not affected. A run takes minutes; the page shows how many traces it read and what it cost.

​

Tracking

When you know a pattern matters, pin it to Tracking. Tracking shows the findings you pinned. On every refresh, manual or on the Auto refresh schedule you choose (off by default), DataFramer checks the traces that arrived since the last refresh against each pinned finding and shows how many traces matched it over time. Slack alerts: press the bell on any pinned finding, and after each refresh one Slack message lists the alerted findings that gained new traces. A tracked finding marked as a failure also gets its root causes and suggested fixes; see Root Causes & Remediation.

​

Traces

Findings works on the traces you pulled in. See Connect your traces for setup, the Traces page, and how to route traces to reviews and datasets.

​

Next steps