Skip to main content

Results table

After a test run completes, the Evaluation Results page shows two layers of insight:

  1. An Aggregation Summary Panel at the top — quick judgement of overall run quality.
  2. An Excel-like results table below — scannable, sortable, filterable rows of individual evaluations.

Aggregation Summary Panel​

The summary panel sits at the top of every results page. It surfaces aggregated metrics across all rows in the run.

Cards​

  • Total test cases run
  • Count and % by label — for label-based (Categorical / Binary) evaluators:
    • Pass / Adequate
    • Inadequate
    • Failed / Error
  • Average score — for evaluators with a numeric score.
  • Min score — for evaluators with a numeric score.
  • Max score — for evaluators with a numeric score.

Counts match the underlying result rows exactly. Missing scores are handled gracefully — they don't break the aggregates, they're just excluded from numeric calculations.

Aggregation rules​

  • Label aggregation is a simple count.
  • The Min, Max, and Avg numeric values appear in the summary cards, not in the per-evaluator table — the per-evaluator table is for scanning individual rows.

The Excel-like results table​

The main results table is built for spreadsheet-style scanning, sorting, and triage. Once you're past 50 rows, this is what makes the difference between usable and unusable.

Table UX​

  • Fixed header row — column headers stay in place as you scroll.
  • Horizontal scrolling — long rows scroll cleanly without breaking layout.
  • Row hover — clear visual indication of the row under the cursor.
  • Reverse chronological order — applies to trace-based evaluator results (which come in over time). For other evaluators (run synchronously alongside the test run), row order doesn't carry meaning.

Column interactions​

  • Sort by any column — score, label, status, and others.
  • Resize columns — drag column dividers.
  • Column grouping — columns are organized into three groups: Inputs, Outputs, Metric.
    • Collapse an entire group.
    • Reopen a collapsed group.

Visual cues​

  • Color-coded labels for Categorical / Binary evaluators:
    • Green — Adequate / Pass.
    • Yellow — Mixed.
    • Red — Inadequate / Error.
  • Status icons instead of text-only — quicker to scan.
  • Truncated reasoning with View more for inline expansion; Show less to collapse again.

Filter the table by:

  • Label
  • Status
  • Evaluator

Search by:

  • Trace ID — find a specific run.
  • Reasoning text — full-text search across the explanations.

Scale​

The table is designed to remain usable with 500+ rows. Sorting, filtering, and scroll performance hold up at that scale.

Common workflows​

  • Triage failures. Filter by Label = "Inadequate", sort by score ascending. Hover-expand reasoning to read why each row failed.
  • Compare two evaluators. Use the column-group collapse to hide outputs and focus on the metric columns side by side.
  • Reproduce a bad row. Search by Trace ID to find the row, then jump to the underlying trace in Observability.

Exporting​

To pull rows out of the platform for offline analysis or sharing, use Exporting results — large result sets are exported asynchronously via email.