@absolutejs/absolute 0.19.0-beta.515 → 0.19.0-beta.516

package/dist/react/ai/index.js

@@ -1647,6 +1647,30 @@ var buildSources = (results) => results.map((result) => ({
   text: result.chunkText,
   title: result.title
 }));
+var buildAnswerGroundingStatus = ({
+  coverage,
+  expectedCount,
+  matchedCount,
+  unresolvedCitationCount,
+  resolvedCitationCount
+}) => {
+  if (expectedCount > 0) {
+    if (matchedCount === expectedCount && unresolvedCitationCount === 0 && coverage !== "ungrounded") {
+      return "pass";
+    }
+    if (matchedCount > 0 || resolvedCitationCount > 0) {
+      return "partial";
+    }
+    return "fail";
+  }
+  if (coverage === "grounded" && unresolvedCitationCount === 0) {
+    return "pass";
+  }
+  if (resolvedCitationCount > 0 || coverage === "partial") {
+    return "partial";
+  }
+  return "fail";
+};
 var buildRAGEvaluationLeaderboard = (runs) => {
   const sorted = [...runs].sort((left, right) => {
     if (right.response.passingRate !== left.response.passingRate) {
@@ -1799,6 +1823,93 @@ var buildRAGEvaluationResponse = (cases) => {
     totalCases
   };
 };
+var evaluateRAGAnswerGroundingCase = ({
+  caseIndex,
+  caseInput
+}) => {
+  const mode = resolveEvaluationMode(caseInput);
+  const expectedIds = normalizeExpectedIds(mode === "chunkId" ? caseInput.expectedChunkIds ?? [] : mode === "source" ? caseInput.expectedSources ?? [] : caseInput.expectedDocumentIds ?? []);
+  const groundedAnswer = buildRAGGroundedAnswer(caseInput.answer, caseInput.sources);
+  const citedReferences = groundedAnswer.parts.flatMap((part) => part.type === "citation" ? part.references : []);
+  const citedIds = normalizeExpectedIds(citedReferences.map((reference) => extractExpectedId(reference, mode)));
+  const expectedSet = new Set(expectedIds);
+  const citedSet = new Set(citedIds);
+  const matchedIds = normalizeExpectedIds([...expectedSet].filter((id) => citedSet.has(id)));
+  const missingIds = normalizeExpectedIds([...expectedSet].filter((id) => !citedSet.has(id)));
+  const extraIds = normalizeExpectedIds([...citedSet].filter((id) => !expectedSet.has(id)));
+  const matchedCount = matchedIds.length;
+  const expectedCount = expectedIds.length;
+  const citedCount = citedIds.length;
+  const precision = citedCount > 0 ? matchedCount / citedCount : 0;
+  const recall = expectedCount > 0 ? matchedCount / expectedCount : 0;
+  const citationF1 = precision + recall > 0 ? 2 * precision * recall / (precision + recall) : 0;
+  const citationCount = groundedAnswer.parts.filter((part) => part.type === "citation").length;
+  const unresolvedCitationCount = new Set(groundedAnswer.ungroundedReferenceNumbers).size;
+  const resolvedCitationCount = citedReferences.length;
+  const resolvedCitationRate = citationCount > 0 ? Math.min(1, resolvedCitationCount / citationCount) : 0;
+  return {
+    answer: caseInput.answer,
+    caseId: caseInput.id ?? `case-${caseIndex + 1}`,
+    citationCount,
+    citationF1,
+    citationPrecision: precision,
+    citationRecall: recall,
+    citedIds,
+    coverage: groundedAnswer.coverage,
+    expectedCount,
+    expectedIds,
+    extraIds,
+    groundedAnswer,
+    hasCitations: groundedAnswer.hasCitations,
+    label: caseInput.label,
+    matchedCount,
+    matchedIds,
+    metadata: caseInput.metadata,
+    missingIds,
+    mode,
+    query: caseInput.query,
+    referenceCount: groundedAnswer.references.length,
+    resolvedCitationCount,
+    resolvedCitationRate,
+    status: buildAnswerGroundingStatus({
+      coverage: groundedAnswer.coverage,
+      expectedCount,
+      matchedCount,
+      resolvedCitationCount,
+      unresolvedCitationCount
+    }),
+    unresolvedCitationCount
+  };
+};
+var buildRAGAnswerGroundingEvaluationResponse = (cases) => {
+  const totalCases = cases.length;
+  const passedCases = cases.filter((entry) => entry.status === "pass").length;
+  const partialCases = cases.filter((entry) => entry.status === "partial").length;
+  const failedCases = cases.filter((entry) => entry.status === "fail").length;
+  const groundedCases = cases.filter((entry) => entry.coverage === "grounded").length;
+  const partiallyGroundedCases = cases.filter((entry) => entry.coverage === "partial").length;
+  const ungroundedCases = cases.filter((entry) => entry.coverage === "ungrounded").length;
+  return {
+    cases,
+    ok: true,
+    passingRate: totalCases > 0 ? passedCases / totalCases * 100 : 0,
+    summary: {
+      averageCitationF1: cases.reduce((sum, result) => sum + result.citationF1, 0) / (totalCases || 1),
+      averageCitationPrecision: cases.reduce((sum, result) => sum + result.citationPrecision, 0) / (totalCases || 1),
+      averageCitationRecall: cases.reduce((sum, result) => sum + result.citationRecall, 0) / (totalCases || 1),
+      averageResolvedCitationRate: cases.reduce((sum, result) => sum + result.resolvedCitationRate, 0) / (totalCases || 1),
+      failedCases,
+      groundedCases,
+      partiallyGroundedCases,
+      passedCases,
+      partialCases,
+      totalCases,
+      ungroundedCases
+    },
+    totalCases
+  };
+};
+var evaluateRAGAnswerGrounding = (input) => buildRAGAnswerGroundingEvaluationResponse(input.cases.map((caseInput, caseIndex) => evaluateRAGAnswerGroundingCase({ caseIndex, caseInput })));
 var compareRAGRerankers = async ({
   collection,
   suite,
@@ -2953,5 +3064,5 @@ export {
   AIStreamProvider
 };
 
-//# debugId=4B5F55063D897A0D64756E2164756E21
+//# debugId=586E6D6CBA45E2DA64756E2164756E21
 //# sourceMappingURL=index.js.map