@absolutejs/absolute 0.19.0-beta.520 → 0.19.0-beta.522

package/dist/ai/index.js

@@ -5129,6 +5129,72 @@ var buildRAGAnswerGroundingEvaluationLeaderboard = (runs) => {
     totalCases: run.response.totalCases
   }));
 };
+var buildRAGAnswerGroundingCaseDifficultyLeaderboard = (entries) => {
+  const grouped = new Map;
+  for (const entry of entries) {
+    for (const result of entry.response.cases) {
+      const current = grouped.get(result.caseId) ?? {
+        caseId: result.caseId,
+        failCount: 0,
+        groundedCount: 0,
+        label: result.label,
+        passCount: 0,
+        partialCount: 0,
+        query: result.query,
+        totalCitationF1: 0,
+        totalEvaluations: 0,
+        totalResolvedCitationRate: 0
+      };
+      current.label ??= result.label;
+      current.query ??= result.query;
+      current.totalEvaluations += 1;
+      current.totalCitationF1 += result.citationF1;
+      current.totalResolvedCitationRate += result.resolvedCitationRate;
+      if (result.status === "pass") {
+        current.passCount += 1;
+      } else if (result.status === "partial") {
+        current.partialCount += 1;
+      } else {
+        current.failCount += 1;
+      }
+      if (result.coverage === "grounded") {
+        current.groundedCount += 1;
+      }
+      grouped.set(result.caseId, current);
+    }
+  }
+  const ranked = Array.from(grouped.values()).sort((left, right) => {
+    const leftPassRate = left.passCount / left.totalEvaluations;
+    const rightPassRate = right.passCount / right.totalEvaluations;
+    if (leftPassRate !== rightPassRate) {
+      return leftPassRate - rightPassRate;
+    }
+    const leftCitationF1 = left.totalCitationF1 / left.totalEvaluations;
+    const rightCitationF1 = right.totalCitationF1 / right.totalEvaluations;
+    if (leftCitationF1 !== rightCitationF1) {
+      return leftCitationF1 - rightCitationF1;
+    }
+    const leftResolved = left.totalResolvedCitationRate / left.totalEvaluations;
+    const rightResolved = right.totalResolvedCitationRate / right.totalEvaluations;
+    if (leftResolved !== rightResolved) {
+      return leftResolved - rightResolved;
+    }
+    return left.caseId.localeCompare(right.caseId);
+  });
+  return ranked.map((entry, index) => ({
+    averageCitationF1: entry.totalCitationF1 / entry.totalEvaluations,
+    averageResolvedCitationRate: entry.totalResolvedCitationRate / entry.totalEvaluations,
+    caseId: entry.caseId,
+    failRate: entry.failCount / entry.totalEvaluations * 100,
+    groundedRate: entry.groundedCount / entry.totalEvaluations * 100,
+    label: entry.label,
+    passRate: entry.passCount / entry.totalEvaluations * 100,
+    partialRate: entry.partialCount / entry.totalEvaluations * 100,
+    query: entry.query,
+    rank: index + 1,
+    totalEvaluations: entry.totalEvaluations
+  }));
+};
 var toHistorySortOrder = (left, right) => right.finishedAt - left.finishedAt;
 var normalizeHistoryRuns = (runs) => [...runs].sort(toHistorySortOrder);
 var toGroundingHistorySortOrder = (left, right) => right.finishedAt - left.finishedAt;
@@ -10828,6 +10894,7 @@ export {
   buildRAGAnswerGroundingEvaluationRunDiff,
   buildRAGAnswerGroundingEvaluationResponse,
   buildRAGAnswerGroundingEvaluationLeaderboard,
+  buildRAGAnswerGroundingCaseDifficultyLeaderboard,
   applyRAGReranking,
   applyRAGQueryTransform,
   anthropicOCR,
@@ -10836,5 +10903,5 @@ export {
   aiChat
 };
 
-//# debugId=ECF363055286173764756E2164756E21
+//# debugId=9E8ADC0BDAB94FDA64756E2164756E21
 //# sourceMappingURL=index.js.map