@absolutejs/absolute 0.19.0-beta.534 → 0.19.0-beta.535

package/dist/types/ai.d.ts

@@ -1195,6 +1195,7 @@ export type RAGEvaluationSuiteRun = {
     elapsedMs: number;
     response: RAGEvaluationResponse;
     traceSummary?: RAGRetrievalTraceComparisonSummary;
+    caseTraceSnapshots?: RAGEvaluationCaseTraceSnapshot[];
     metadata?: Record<string, unknown>;
 };
 export type RAGEvaluationHistoryStore = {
@@ -1246,11 +1247,37 @@ export type RAGEvaluationRunDiff = {
         stageCounts: Partial<Record<RAGRetrievalTraceStage, number>>;
     };
 };
+export type RAGEvaluationCaseTraceSnapshot = {
+    caseId: string;
+    label?: string;
+    query: string;
+    status: RAGEvaluationCaseResult['status'];
+    traceMode?: RAGHybridRetrievalMode;
+    previousTraceMode?: RAGHybridRetrievalMode;
+    transformedQuery?: string;
+    previousTransformedQuery?: string;
+    variantQueries: string[];
+    previousVariantQueries: string[];
+    finalCount: number;
+    previousFinalCount?: number;
+    vectorCount: number;
+    previousVectorCount?: number;
+    lexicalCount: number;
+    previousLexicalCount?: number;
+    candidateTopK: number;
+    previousCandidateTopK?: number;
+    lexicalTopK: number;
+    previousLexicalTopK?: number;
+    stageCounts: Partial<Record<RAGRetrievalTraceStage, number>>;
+    previousStageCounts: Partial<Record<RAGRetrievalTraceStage, number>>;
+    traceChange: 'new' | 'changed' | 'unchanged';
+};
 export type RAGEvaluationHistory = {
     suiteId: string;
     suiteLabel?: string;
     runs: RAGEvaluationSuiteRun[];
     leaderboard: RAGEvaluationLeaderboardEntry[];
+    caseTraceSnapshots: RAGEvaluationCaseTraceSnapshot[];
     latestRun?: RAGEvaluationSuiteRun;
     previousRun?: RAGEvaluationSuiteRun;
     diff?: RAGEvaluationRunDiff;
@@ -1301,6 +1328,7 @@ export type RAGRerankerComparisonEntry = {
     providerName?: string;
     response: RAGEvaluationResponse;
     traceSummary?: RAGRetrievalTraceComparisonSummary;
+    caseTraceSnapshots?: RAGEvaluationCaseTraceSnapshot[];
 };
 export type RAGRerankerComparisonSummary = {
     bestByPassingRate?: string;
@@ -1320,6 +1348,7 @@ export type RAGRetrievalComparisonEntry = {
     retrievalMode: RAGHybridRetrievalMode;
     response: RAGEvaluationResponse;
     traceSummary?: RAGRetrievalTraceComparisonSummary;
+    caseTraceSnapshots?: RAGEvaluationCaseTraceSnapshot[];
 };
 export type RAGRetrievalComparisonSummary = {
     bestByPassingRate?: string;

package/dist/vue/ai/index.js

@@ -2154,6 +2154,98 @@ var buildGroundingCaseSnapshots = ({
     };
   });
 };
+var areStageCountsEqual = (left, right) => {
+  const keys = new Set([
+    ...Object.keys(left),
+    ...Object.keys(right)
+  ]);
+  for (const key of keys) {
+    if ((left[key] ?? 0) !== (right[key] ?? 0)) {
+      return false;
+    }
+  }
+  return true;
+};
+var buildEvaluationCaseTraceSnapshot = ({
+  caseResult,
+  currentTrace,
+  previousTrace
+}) => {
+  const stageCounts = currentTrace ? buildTraceStageCounts([currentTrace]) : {};
+  const previousStageCounts = previousTrace?.stageCounts ?? {};
+  const traceChange = !previousTrace ? currentTrace ? "new" : "unchanged" : previousTrace.traceMode !== currentTrace?.mode || previousTrace.transformedQuery !== (currentTrace?.transformedQuery || undefined) || previousTrace.variantQueries.join("|") !== (currentTrace?.variantQueries ?? []).join("|") || previousTrace.finalCount !== (currentTrace?.resultCounts.final ?? 0) || previousTrace.vectorCount !== (currentTrace?.resultCounts.vector ?? 0) || previousTrace.lexicalCount !== (currentTrace?.resultCounts.lexical ?? 0) || previousTrace.candidateTopK !== (currentTrace?.candidateTopK ?? 0) || previousTrace.lexicalTopK !== (currentTrace?.lexicalTopK ?? 0) || !areStageCountsEqual(previousStageCounts, stageCounts) ? "changed" : "unchanged";
+  return {
+    candidateTopK: currentTrace?.candidateTopK ?? 0,
+    caseId: caseResult.caseId,
+    finalCount: currentTrace?.resultCounts.final ?? 0,
+    label: caseResult.label,
+    lexicalCount: currentTrace?.resultCounts.lexical ?? 0,
+    lexicalTopK: currentTrace?.lexicalTopK ?? 0,
+    previousCandidateTopK: previousTrace?.candidateTopK,
+    previousFinalCount: previousTrace?.finalCount,
+    previousLexicalCount: previousTrace?.lexicalCount,
+    previousLexicalTopK: previousTrace?.lexicalTopK,
+    previousStageCounts,
+    previousTraceMode: previousTrace?.traceMode,
+    previousTransformedQuery: previousTrace?.transformedQuery,
+    previousVariantQueries: previousTrace?.variantQueries ?? [],
+    previousVectorCount: previousTrace?.vectorCount,
+    query: caseResult.query,
+    stageCounts,
+    status: caseResult.status,
+    traceChange,
+    traceMode: currentTrace?.mode,
+    transformedQuery: currentTrace?.transformedQuery || undefined,
+    variantQueries: currentTrace?.variantQueries ?? [],
+    vectorCount: currentTrace?.resultCounts.vector ?? 0
+  };
+};
+var buildEvaluationCaseTraceSnapshotsFromEvaluated = (evaluated) => evaluated.map(({ caseResult, trace }) => buildEvaluationCaseTraceSnapshot({
+  caseResult,
+  currentTrace: trace
+}));
+var buildEvaluationCaseTraceSnapshots = ({
+  current,
+  previous
+}) => {
+  if (!current) {
+    return [];
+  }
+  const currentTraces = new Map((current.caseTraceSnapshots ?? []).map((entry) => [entry.caseId, entry]));
+  const previousTraces = new Map((previous?.caseTraceSnapshots ?? []).map((entry) => [
+    entry.caseId,
+    entry
+  ]));
+  return current.response.cases.map((caseResult) => buildEvaluationCaseTraceSnapshot({
+    caseResult,
+    currentTrace: (() => {
+      const currentSnapshot = currentTraces.get(caseResult.caseId);
+      if (!currentSnapshot) {
+        return;
+      }
+      return {
+        candidateTopK: currentSnapshot.candidateTopK,
+        lexicalTopK: currentSnapshot.lexicalTopK,
+        mode: currentSnapshot.traceMode ?? "vector",
+        query: caseResult.query,
+        resultCounts: {
+          final: currentSnapshot.finalCount,
+          fused: currentSnapshot.finalCount,
+          lexical: currentSnapshot.lexicalCount,
+          reranked: currentSnapshot.finalCount,
+          vector: currentSnapshot.vectorCount
+        },
+        runLexical: currentSnapshot.lexicalCount > 0,
+        runVector: currentSnapshot.vectorCount > 0,
+        steps: [],
+        topK: caseResult.topK,
+        transformedQuery: currentSnapshot.transformedQuery ?? caseResult.query,
+        variantQueries: currentSnapshot.variantQueries
+      };
+    })(),
+    previousTrace: previousTraces.get(caseResult.caseId)
+  }));
+};
 var getStatusRank = (status) => status === "pass" ? 2 : status === "partial" ? 1 : 0;
 var formatSignedDelta = (value, decimals = 0, suffix = "") => `${value >= 0 ? "+" : ""}${value.toFixed(decimals)}${suffix}`;
 var formatEvaluationSummary = (response) => `${response.summary.passedCases}/${response.totalCases} pass \xB7 f1 ${response.summary.averageF1.toFixed(3)} \xB7 latency ${response.summary.averageLatencyMs.toFixed(1)}ms`;
@@ -2313,6 +2405,13 @@ var buildRAGEvaluationHistoryRows = (history) => {
       rows.push({ label: "Trace stage delta", value: stageDelta });
     }
   }
+  if (history.caseTraceSnapshots.length > 0) {
+    const changedCases = history.caseTraceSnapshots.filter((entry) => entry.traceChange === "changed");
+    rows.push({
+      label: "Trace drift cases",
+      value: changedCases.length > 0 ? changedCases.map((entry) => entry.label ?? entry.caseId).slice(0, 4).join(", ") : "none"
+    });
+  }
   return rows;
 };
 var buildRAGEvaluationRunDiff = ({
@@ -2502,6 +2601,10 @@ var loadRAGEvaluationHistory = async ({
   const latestRun = runs[0];
   const previousRun = runs[1];
   return {
+    caseTraceSnapshots: buildEvaluationCaseTraceSnapshots({
+      current: latestRun,
+      previous: previousRun
+    }),
     diff: latestRun && previousRun ? buildRAGEvaluationRunDiff({
       current: latestRun,
       previous: previousRun
@@ -2713,6 +2816,7 @@ var compareRAGRerankers = async ({
     });
     const response = buildRAGEvaluationResponse(evaluated.map((entry) => entry.caseResult));
     return {
+      caseTraceSnapshots: buildEvaluationCaseTraceSnapshotsFromEvaluated(evaluated),
       label: candidate.label ?? candidate.id,
       providerName: typeof candidate.rerank === "function" ? undefined : candidate.rerank?.providerName,
       response,
@@ -2799,6 +2903,7 @@ var compareRAGRetrievalStrategies = async ({
     });
     const response = buildRAGEvaluationResponse(evaluated.map((entry) => entry.caseResult));
     return {
+      caseTraceSnapshots: buildEvaluationCaseTraceSnapshotsFromEvaluated(evaluated),
       label: candidate.label ?? candidate.id,
       response,
       retrievalId: candidate.id,
@@ -2867,7 +2972,8 @@ var executeDryRunRAGEvaluation = (input, defaultTopK = DEFAULT_TOP_K) => input.c
 var runRAGEvaluationSuite = async ({
   suite,
   evaluate,
-  overrides
+  overrides,
+  artifacts
 }) => {
   const startedAt = Date.now();
   const response = await evaluate({
@@ -2877,6 +2983,7 @@ var runRAGEvaluationSuite = async ({
   });
   const finishedAt = Date.now();
   return {
+    caseTraceSnapshots: artifacts?.caseTraceSnapshots,
     elapsedMs: finishedAt - startedAt,
     finishedAt,
     id: generateId(),
@@ -2884,7 +2991,8 @@ var runRAGEvaluationSuite = async ({
     metadata: suite.metadata,
     response,
     startedAt,
-    suiteId: suite.id
+    suiteId: suite.id,
+    traceSummary: artifacts?.traceSummary
   };
 };
 var summarizeRAGEvaluationCase = ({
@@ -3597,5 +3705,5 @@ export {
   AIStreamKey
 };
 
-//# debugId=49142334D6D3EC2564756E2164756E21
+//# debugId=E235F4588786F00E64756E2164756E21
 //# sourceMappingURL=index.js.map