n8n/packages/@n8n/instance-ai/evaluations/summary.ts
Mutasem Aldmour 7d6fd4ca9c
fix: Report an unverifiable unit as notVerified, not a silent pass (#34421)
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-17 14:23:10 +00:00

112 lines
4.1 KiB
TypeScript

import type {
BuildExpectationAggregation,
ExecutionScenarioAggregation,
TestCaseAggregation,
WorkflowTestCase,
WorkflowTestCaseResult,
} from './types';
export type CaseRunStatus = 'checked' | 'built' | 'build_failed';
export interface ScoredCounts {
passCount: number;
totalCount: number;
}
export type AggregatedCaseUnit = ExecutionScenarioAggregation | BuildExpectationAggregation;
/**
* Whether a case is graded on a built workflow. Keyed off execution scenarios only:
* outcome expectations are judged from the rendered build context (workflow AND any
* agent/config-eval artifact), so a case that produces an agent instead of a workflow
* still grades cleanly rather than reporting a missing-workflow build failure.
*/
export function requiresWorkflowOutput(testCase: WorkflowTestCase): boolean {
return (testCase.executionScenarios?.length ?? 0) > 0;
}
export function getAggregatedCaseUnits(tc: TestCaseAggregation): AggregatedCaseUnit[] {
return [...tc.executionScenarios, ...tc.buildExpectations.filter((ea) => ea.evaluatedCount > 0)];
}
export function countAggregatedUnitTrials(units: AggregatedCaseUnit[]): ScoredCounts {
return units.reduce<ScoredCounts>(
(counts, unit) => {
counts.passCount += unit.passCount;
counts.totalCount += unit.evaluatedCount;
return counts;
},
{ passCount: 0, totalCount: 0 },
);
}
export function getCheckedRunCount(tc: TestCaseAggregation): number {
if (requiresWorkflowOutput(tc.testCase)) return tc.buildSuccessCount;
// A scenario-less case is "checked" on any run that produced a scoreable unit — a judged
// process/outcome expectation (which covers the workflow and any agent/config-eval artifact).
return tc.runs.filter((run) => (run.buildExpectationResults?.length ?? 0) > 0).length;
}
export function getRunScoredCounts(result: WorkflowTestCaseResult): ScoredCounts {
// Incomplete units (no verifier/judge verdict) are visible but not scored.
const scoredExpectations = (result.buildExpectationResults ?? []).filter((e) => !e.incomplete);
const scoredScenarios = result.executionScenarioResults.filter((sr) => !sr.incomplete);
return {
passCount:
scoredScenarios.filter((sr) => sr.success).length +
scoredExpectations.filter((e) => e.pass).length,
totalCount: scoredScenarios.length + scoredExpectations.length,
};
}
export function getCaseRunStatus(result: WorkflowTestCaseResult): CaseRunStatus {
// A scenario-less case (e.g. an agent/config-eval build) produces no workflow to score by
// design; treat it as checked (not build-failed) once it has any scoreable unit — a judged
// process/outcome expectation.
const checkedWithoutWorkflow =
!requiresWorkflowOutput(result.testCase) && (result.buildExpectationResults?.length ?? 0) > 0;
if (checkedWithoutWorkflow) return 'checked';
if (result.workflowBuildSuccess) return 'built';
return 'build_failed';
}
/** Per-case verification tallies across an aggregated run. `notVerified` cases —
* where nothing could be scored — are counted apart from passed/failed so the
* reported pass rate reflects only what was actually checked. */
export interface CaseVerificationRollup {
/** Verified cases where every evaluated unit passed. */
passed: number;
/** Verified cases with at least one failing evaluated unit. */
failed: number;
/** Cases where no unit could be scored (all incomplete / skipped). */
notVerified: number;
}
export function rollupCaseVerification(cases: TestCaseAggregation[]): CaseVerificationRollup {
const rollup: CaseVerificationRollup = { passed: 0, failed: 0, notVerified: 0 };
for (const tc of cases) {
if (tc.status === 'notVerified') {
rollup.notVerified++;
continue;
}
const { passCount, totalCount } = countAggregatedUnitTrials(getAggregatedCaseUnits(tc));
if (totalCount > 0 && passCount === totalCount) rollup.passed++;
else rollup.failed++;
}
return rollup;
}
export function getCaseRunStatusLabel(result: WorkflowTestCaseResult): string {
switch (getCaseRunStatus(result)) {
case 'checked':
return 'CHECKED';
case 'built':
return 'BUILT';
case 'build_failed':
return 'BUILD FAILED';
}
}