@absolutejs/absolute 0.19.0-beta.537 → 0.19.0-beta.539

package/dist/ai/index.js

@@ -5834,8 +5834,11 @@ var buildEvaluationCaseTraceSnapshots = ({
 };
 var getStatusRank = (status) => status === "pass" ? 2 : status === "partial" ? 1 : 0;
 var formatSignedDelta = (value, decimals = 0, suffix = "") => `${value >= 0 ? "+" : ""}${value.toFixed(decimals)}${suffix}`;
+var formatEvaluationPassingRate = (value) => `${value.toFixed(1)}%`;
 var formatEvaluationSummary = (response) => `${response.summary.passedCases}/${response.totalCases} pass \xB7 f1 ${response.summary.averageF1.toFixed(3)} \xB7 latency ${response.summary.averageLatencyMs.toFixed(1)}ms`;
+var formatGroundingHistorySummaryValue = (response) => `${response.summary.passedCases}/${response.summary.totalCases} pass \xB7 grounded ${response.summary.groundedCases} \xB7 partial ${response.summary.partialCases} \xB7 ungrounded ${response.summary.ungroundedCases} \xB7 resolved citations ${(response.summary.averageResolvedCitationRate * 100).toFixed(1)}% \xB7 citation f1 ${response.summary.averageCitationF1.toFixed(3)}`;
 var formatHistoryCaseLabels = (cases) => cases.length > 0 ? cases.map((entry) => entry.label ?? entry.caseId).join(", ") : "none";
+var formatGroundingHistoryCaseLabels = (cases) => cases.length > 0 ? cases.map((entry) => entry.label ?? entry.caseId).join(", ") : "none";
 var formatTraceModes = (modes) => modes.length > 0 ? modes.join(" / ") : "n/a";
 var formatTraceStageSummary = (stageCounts) => {
   const topStages = Object.entries(stageCounts).sort((left, right) => right[1] - left[1]).slice(0, 3);
@@ -6058,6 +6061,11 @@ var buildRAGEvaluationCaseTracePresentations = (history) => {
     };
   });
 };
+var buildRAGEvaluationHistoryPresentation = (history) => ({
+  caseTraces: buildRAGEvaluationCaseTracePresentations(history),
+  rows: buildRAGEvaluationHistoryRows(history),
+  summary: history?.latestRun ? history.latestRun.label : "No persisted benchmark runs yet."
+});
 var buildRAGEvaluationRunDiff = ({
   current,
   previous
@@ -6126,6 +6134,129 @@ var buildRAGAnswerGroundingEvaluationRunDiff = ({
     unchangedCases
   };
 };
+var buildRAGAnswerGroundingCaseSnapshotPresentations = (history) => {
+  if (!history?.caseSnapshots.length) {
+    return [];
+  }
+  return history.caseSnapshots.map((entry) => {
+    const label = entry.label ?? entry.caseId;
+    return {
+      answerChange: entry.answerChange,
+      caseId: entry.caseId,
+      label,
+      rows: [
+        {
+          label: "Query",
+          value: entry.query?.trim().length ? entry.query : "n/a"
+        },
+        { label: "Answer change", value: entry.answerChange },
+        { label: "Coverage", value: entry.coverage },
+        {
+          label: "Resolved citations",
+          value: `${entry.resolvedCitationCount}/${entry.citationCount}`
+        },
+        {
+          label: "Resolved citation rate",
+          value: entry.resolvedCitationRate.toFixed(3)
+        },
+        { label: "Citation F1", value: entry.citationF1.toFixed(3) },
+        {
+          label: "Reference count",
+          value: String(entry.referenceCount)
+        },
+        {
+          label: "Cited IDs",
+          value: entry.citedIds.length > 0 ? entry.citedIds.join(", ") : "none"
+        },
+        {
+          label: "Matched IDs",
+          value: entry.matchedIds.length > 0 ? entry.matchedIds.join(", ") : "none"
+        },
+        {
+          label: "Missing IDs",
+          value: entry.missingIds.length > 0 ? entry.missingIds.join(", ") : "none"
+        },
+        {
+          label: "Extra IDs",
+          value: entry.extraIds.length > 0 ? entry.extraIds.join(", ") : "none"
+        },
+        {
+          label: "Unresolved refs",
+          value: entry.ungroundedReferenceNumbers.length > 0 ? entry.ungroundedReferenceNumbers.join(", ") : "none"
+        },
+        {
+          label: "Answer",
+          value: entry.answer.trim().length > 0 ? entry.answer : "n/a"
+        },
+        {
+          label: "Previous answer",
+          value: entry.previousAnswer && entry.previousAnswer.trim().length > 0 ? entry.previousAnswer : "n/a"
+        }
+      ],
+      summary: `${entry.answerChange} \xB7 ${entry.coverage} \xB7 resolved ${entry.resolvedCitationCount}/${entry.citationCount} \xB7 refs ${entry.referenceCount}`
+    };
+  });
+};
+var buildRAGAnswerGroundingHistoryRows = (history) => {
+  if (!history?.latestRun) {
+    return [{ label: "History", value: "No persisted provider runs yet." }];
+  }
+  const rows = [
+    { label: "Runs recorded", value: String(history.runs.length) },
+    {
+      label: "Latest",
+      value: `${history.latestRun.label} \xB7 ${formatGroundingHistorySummaryValue(history.latestRun.response)}`
+    }
+  ];
+  if (history.previousRun) {
+    rows.push({
+      label: "Previous",
+      value: `${history.previousRun.label} \xB7 ${formatGroundingHistorySummaryValue(history.previousRun.response)}`
+    });
+  }
+  if (history.leaderboard[0]) {
+    rows.push({
+      label: "Best recorded",
+      value: `#${history.leaderboard[0].rank} \xB7 ${history.leaderboard[0].label} \xB7 passing ${formatEvaluationPassingRate(history.leaderboard[0].passingRate)} \xB7 citation f1 ${history.leaderboard[0].averageCitationF1.toFixed(3)} \xB7 resolved ${formatEvaluationPassingRate(history.leaderboard[0].averageResolvedCitationRate)}`
+    });
+  }
+  if (history.caseSnapshots.length > 0) {
+    const changedAnswers = history.caseSnapshots.filter((entry) => entry.answerChange === "changed").length;
+    rows.push({
+      label: "Answer drift",
+      value: `${changedAnswers}/${history.caseSnapshots.length} changed`
+    });
+  }
+  if (!history.diff) {
+    rows.push({
+      label: "History diff",
+      value: "Run the provider comparison again to diff grounding regressions over time."
+    });
+    return rows;
+  }
+  rows.push({
+    label: "Passing delta",
+    value: formatSignedDelta(history.diff.summaryDelta.passingRate, 1, "%")
+  }, {
+    label: "Citation F1 delta",
+    value: formatSignedDelta(history.diff.summaryDelta.averageCitationF1, 3)
+  }, {
+    label: "Resolved citation delta",
+    value: formatSignedDelta(history.diff.summaryDelta.averageResolvedCitationRate * 100, 1, "%")
+  }, {
+    label: "Improved",
+    value: formatGroundingHistoryCaseLabels(history.diff.improvedCases)
+  }, {
+    label: "Regressed",
+    value: formatGroundingHistoryCaseLabels(history.diff.regressedCases)
+  });
+  return rows;
+};
+var buildRAGAnswerGroundingHistoryPresentation = (history) => ({
+  caseSnapshots: buildRAGAnswerGroundingCaseSnapshotPresentations(history),
+  rows: buildRAGAnswerGroundingHistoryRows(history),
+  summary: history?.latestRun ? history.latestRun.label : "No persisted provider runs yet."
+});
 var createRAGFileEvaluationHistoryStore = (path) => ({
   listRuns: async ({ limit, suiteId } = {}) => {
     let parsed = [];
@@ -11814,15 +11945,19 @@ export {
   buildRAGEvaluationResponse,
   buildRAGEvaluationLeaderboard,
   buildRAGEvaluationHistoryRows,
+  buildRAGEvaluationHistoryPresentation,
   buildRAGEvaluationCaseTracePresentations,
   buildRAGContext,
   buildRAGComparisonTraceSummaryRows,
   buildRAGComparisonTraceDiffRows,
   buildRAGCitations,
   buildRAGCitationReferenceMap,
+  buildRAGAnswerGroundingHistoryRows,
+  buildRAGAnswerGroundingHistoryPresentation,
   buildRAGAnswerGroundingEvaluationRunDiff,
   buildRAGAnswerGroundingEvaluationResponse,
   buildRAGAnswerGroundingEvaluationLeaderboard,
+  buildRAGAnswerGroundingCaseSnapshotPresentations,
   buildRAGAnswerGroundingCaseDifficultyRunDiff,
   buildRAGAnswerGroundingCaseDifficultyLeaderboard,
   applyRAGReranking,
@@ -11833,5 +11968,5 @@ export {
   aiChat
 };
 
-//# debugId=BBC0B1F804D3DF1E64756E2164756E21
+//# debugId=188A5A929B4F9DCA64756E2164756E21
 //# sourceMappingURL=index.js.map