@absolutejs/absolute 0.19.0-beta.538 → 0.19.0-beta.539

package/dist/types/ai.d.ts

@@ -1311,6 +1311,16 @@ export type RAGEvaluationCaseTracePresentation = {
     traceChange: RAGEvaluationCaseTraceSnapshot['traceChange'];
     rows: RAGLabelValueRow[];
 };
+export type RAGEvaluationHistoryPresentation = {
+    summary: string;
+    rows: RAGLabelValueRow[];
+    caseTraces: RAGEvaluationCaseTracePresentation[];
+};
+export type RAGAnswerGroundingHistoryPresentation = {
+    summary: string;
+    rows: RAGLabelValueRow[];
+    caseSnapshots: RAGAnswerGroundingCaseSnapshotPresentation[];
+};
 export type RAGEvaluationLeaderboardEntry = {
     runId: string;
     suiteId: string;

package/dist/vue/ai/index.js

@@ -2311,8 +2311,11 @@ 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 ${response.summary.averageCitationF1.toFixed(3)}`;
 var formatHistoryCaseLabels = (cases) => cases.length > 0 ? cases.map((entry) => entry.label ?? entry.caseId).join(", ") : "none";
+var formatGroundingHistoryCaseLabels = (cases) => cases.length > 0 ? cases.map((entry) => entry.label ?? entry.caseId).join(", ") : "none";
 var formatTraceModes = (modes) => modes.length > 0 ? modes.join(" / ") : "n/a";
 var formatTraceStageSummary = (stageCounts) => {
   const topStages = Object.entries(stageCounts).sort((left, right) => right[1] - left[1]).slice(0, 3);
@@ -2535,6 +2538,11 @@ var buildRAGEvaluationCaseTracePresentations = (history) => {
     };
   });
 };
+var buildRAGEvaluationHistoryPresentation = (history) => ({
+  caseTraces: buildRAGEvaluationCaseTracePresentations(history),
+  rows: buildRAGEvaluationHistoryRows(history),
+  summary: history?.latestRun ? history.latestRun.label : "No persisted benchmark runs yet."
+});
 var buildRAGEvaluationRunDiff = ({
   current,
   previous
@@ -2666,6 +2674,66 @@ var buildRAGAnswerGroundingCaseSnapshotPresentations = (history) => {
     };
   });
 };
+var buildRAGAnswerGroundingHistoryRows = (history) => {
+  if (!history?.latestRun) {
+    return [{ label: "History", value: "No persisted provider runs yet." }];
+  }
+  const rows = [
+    { label: "Runs recorded", value: String(history.runs.length) },
+    {
+      label: "Latest",
+      value: `${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 = [];
@@ -3889,5 +3957,5 @@ export {
   AIStreamKey
 };
 
-//# debugId=69A98548515264F064756E2164756E21
+//# debugId=557BB3208F1A1E8F64756E2164756E21
 //# sourceMappingURL=index.js.map