@absolutejs/absolute 0.19.0-beta.532 → 0.19.0-beta.534

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/ai/index.js CHANGED
@@ -5300,6 +5300,23 @@ var buildTraceStageCounts = (traces) => {
5300
5300
  }
5301
5301
  return counts;
5302
5302
  };
5303
+ var diffTraceStageCounts = ({
5304
+ current,
5305
+ previous
5306
+ }) => {
5307
+ const next = {};
5308
+ const stages = new Set([
5309
+ ...Object.keys(current),
5310
+ ...Object.keys(previous)
5311
+ ]);
5312
+ for (const stage of stages) {
5313
+ const delta = (current[stage] ?? 0) - (previous[stage] ?? 0);
5314
+ if (delta !== 0) {
5315
+ next[stage] = delta;
5316
+ }
5317
+ }
5318
+ return next;
5319
+ };
5303
5320
  var roundTraceAverage = (value, total) => total > 0 ? Number((value / total).toFixed(2)) : 0;
5304
5321
  var summarizeRetrievalTraces = (traces) => {
5305
5322
  if (traces.length === 0) {
@@ -5661,6 +5678,166 @@ var buildGroundingCaseSnapshots = ({
5661
5678
  });
5662
5679
  };
5663
5680
  var getStatusRank = (status) => status === "pass" ? 2 : status === "partial" ? 1 : 0;
5681
+ var formatSignedDelta = (value, decimals = 0, suffix = "") => `${value >= 0 ? "+" : ""}${value.toFixed(decimals)}${suffix}`;
5682
+ var formatEvaluationSummary = (response) => `${response.summary.passedCases}/${response.totalCases} pass \xB7 f1 ${response.summary.averageF1.toFixed(3)} \xB7 latency ${response.summary.averageLatencyMs.toFixed(1)}ms`;
5683
+ var formatHistoryCaseLabels = (cases) => cases.length > 0 ? cases.map((entry) => entry.label ?? entry.caseId).join(", ") : "none";
5684
+ var formatTraceModes = (modes) => modes.length > 0 ? modes.join(" / ") : "n/a";
5685
+ var formatTraceStageSummary = (stageCounts) => {
5686
+ const topStages = Object.entries(stageCounts).sort((left, right) => right[1] - left[1]).slice(0, 3);
5687
+ return topStages.length > 0 ? topStages.map(([stage, count]) => `${stage} ${count}`).join(" \xB7 ") : "n/a";
5688
+ };
5689
+ var formatTraceRatio = (count, total) => `${count}/${total}`;
5690
+ var formatTraceCountDelta = (value) => `${value >= 0 ? "+" : ""}${value}`;
5691
+ var buildRAGComparisonTraceSummaryRows = (entry) => {
5692
+ const trace = entry.traceSummary;
5693
+ if (!trace) {
5694
+ return [{ label: "Trace", value: "Unavailable" }];
5695
+ }
5696
+ return [
5697
+ { label: "Modes", value: formatTraceModes(trace.modes) },
5698
+ { label: "Avg final", value: trace.averageFinalCount.toFixed(1) },
5699
+ { label: "Avg vector", value: trace.averageVectorCount.toFixed(1) },
5700
+ { label: "Avg lexical", value: trace.averageLexicalCount.toFixed(1) },
5701
+ {
5702
+ label: "Transforms",
5703
+ value: formatTraceRatio(trace.transformedCases, trace.totalCases)
5704
+ },
5705
+ {
5706
+ label: "Variants",
5707
+ value: formatTraceRatio(trace.variantCases, trace.totalCases)
5708
+ },
5709
+ {
5710
+ label: "TopK",
5711
+ value: `${trace.averageCandidateTopK.toFixed(1)} / ${trace.averageLexicalTopK.toFixed(1)}`
5712
+ },
5713
+ {
5714
+ label: "Stages",
5715
+ value: formatTraceStageSummary(trace.stageCounts)
5716
+ }
5717
+ ];
5718
+ };
5719
+ var buildRAGComparisonTraceDiffRows = (entry, leader) => {
5720
+ const trace = entry.traceSummary;
5721
+ if (!trace) {
5722
+ return [{ label: "Trace", value: "Unavailable for comparison" }];
5723
+ }
5724
+ const leaderTrace = leader?.traceSummary;
5725
+ if (!leaderTrace) {
5726
+ return [{ label: "Baseline", value: "Leader trace unavailable" }];
5727
+ }
5728
+ if (entry === leader) {
5729
+ return [{ label: "Baseline", value: "Leader strategy" }];
5730
+ }
5731
+ const stageDelta = Object.keys({
5732
+ ...leaderTrace.stageCounts,
5733
+ ...trace.stageCounts
5734
+ }).map((stage) => {
5735
+ const typedStage = stage;
5736
+ const delta = (trace.stageCounts[typedStage] ?? 0) - (leaderTrace.stageCounts[typedStage] ?? 0);
5737
+ return delta === 0 ? null : `${typedStage} ${formatTraceCountDelta(delta)}`;
5738
+ }).filter((value) => Boolean(value)).slice(0, 3).join(" \xB7 ");
5739
+ const rows = [
5740
+ { label: "Baseline", value: leader.label }
5741
+ ];
5742
+ if (formatTraceModes(trace.modes) !== formatTraceModes(leaderTrace.modes)) {
5743
+ rows.push({
5744
+ label: "Modes vs leader",
5745
+ value: `${formatTraceModes(trace.modes)} vs ${formatTraceModes(leaderTrace.modes)}`
5746
+ });
5747
+ }
5748
+ rows.push({
5749
+ label: "Final delta",
5750
+ value: formatSignedDelta(trace.averageFinalCount - leaderTrace.averageFinalCount, 1)
5751
+ }, {
5752
+ label: "Vector delta",
5753
+ value: formatSignedDelta(trace.averageVectorCount - leaderTrace.averageVectorCount, 1)
5754
+ }, {
5755
+ label: "Lexical delta",
5756
+ value: formatSignedDelta(trace.averageLexicalCount - leaderTrace.averageLexicalCount, 1)
5757
+ }, {
5758
+ label: "Transform delta",
5759
+ value: formatTraceCountDelta(trace.transformedCases - leaderTrace.transformedCases)
5760
+ });
5761
+ if (stageDelta) {
5762
+ rows.push({ label: "Stage delta", value: stageDelta });
5763
+ }
5764
+ return rows;
5765
+ };
5766
+ var buildRAGEvaluationHistoryRows = (history) => {
5767
+ if (!history?.latestRun) {
5768
+ return [
5769
+ { label: "History", value: "No persisted benchmark runs yet." }
5770
+ ];
5771
+ }
5772
+ const rows = [
5773
+ { label: "Runs recorded", value: String(history.runs.length) },
5774
+ {
5775
+ label: "Latest",
5776
+ value: `${history.latestRun.label} \xB7 ${formatEvaluationSummary(history.latestRun.response)}`
5777
+ }
5778
+ ];
5779
+ if (history.latestRun.traceSummary) {
5780
+ rows.push({
5781
+ label: "Latest trace",
5782
+ value: `${formatTraceModes(history.latestRun.traceSummary.modes)} \xB7 final ${history.latestRun.traceSummary.averageFinalCount.toFixed(1)} \xB7 vector ${history.latestRun.traceSummary.averageVectorCount.toFixed(1)} \xB7 lexical ${history.latestRun.traceSummary.averageLexicalCount.toFixed(1)}`
5783
+ });
5784
+ }
5785
+ if (history.previousRun) {
5786
+ rows.push({
5787
+ label: "Previous",
5788
+ value: `${history.previousRun.label} \xB7 ${formatEvaluationSummary(history.previousRun.response)}`
5789
+ });
5790
+ }
5791
+ if (!history.diff) {
5792
+ rows.push({
5793
+ label: "History diff",
5794
+ value: "Run the benchmark again to diff regressions over time."
5795
+ });
5796
+ return rows;
5797
+ }
5798
+ rows.push({
5799
+ label: "Passing delta",
5800
+ value: formatSignedDelta(history.diff.summaryDelta.passingRate, 1, "%")
5801
+ }, {
5802
+ label: "Average F1 delta",
5803
+ value: formatSignedDelta(history.diff.summaryDelta.averageF1, 3)
5804
+ }, {
5805
+ label: "Latency delta",
5806
+ value: formatSignedDelta(history.diff.summaryDelta.averageLatencyMs, 1, "ms")
5807
+ }, {
5808
+ label: "Improved",
5809
+ value: formatHistoryCaseLabels(history.diff.improvedCases)
5810
+ }, {
5811
+ label: "Regressed",
5812
+ value: formatHistoryCaseLabels(history.diff.regressedCases)
5813
+ });
5814
+ if (history.diff.traceSummaryDelta) {
5815
+ rows.push({
5816
+ label: "Trace mode shift",
5817
+ value: history.diff.traceSummaryDelta.modesChanged ? "changed" : "stable"
5818
+ }, {
5819
+ label: "Trace final delta",
5820
+ value: formatSignedDelta(history.diff.traceSummaryDelta.averageFinalCount, 1)
5821
+ }, {
5822
+ label: "Trace vector delta",
5823
+ value: formatSignedDelta(history.diff.traceSummaryDelta.averageVectorCount, 1)
5824
+ }, {
5825
+ label: "Trace lexical delta",
5826
+ value: formatSignedDelta(history.diff.traceSummaryDelta.averageLexicalCount, 1)
5827
+ }, {
5828
+ label: "Trace transform delta",
5829
+ value: formatTraceCountDelta(history.diff.traceSummaryDelta.transformedCases)
5830
+ }, {
5831
+ label: "Trace variant delta",
5832
+ value: formatTraceCountDelta(history.diff.traceSummaryDelta.variantCases)
5833
+ });
5834
+ const stageDelta = Object.entries(history.diff.traceSummaryDelta.stageCounts ?? {}).map(([stage, count]) => `${stage} ${formatTraceCountDelta(count)}`).join(", ");
5835
+ if (stageDelta) {
5836
+ rows.push({ label: "Trace stage delta", value: stageDelta });
5837
+ }
5838
+ }
5839
+ return rows;
5840
+ };
5664
5841
  var buildRAGEvaluationRunDiff = ({
5665
5842
  current,
5666
5843
  previous
@@ -5684,6 +5861,22 @@ var buildRAGEvaluationRunDiff = ({
5684
5861
  passingRate: current.response.passingRate - (previous?.response.passingRate ?? 0),
5685
5862
  partialCases: current.response.summary.partialCases - (previous?.response.summary.partialCases ?? 0)
5686
5863
  },
5864
+ traceSummaryDelta: current.traceSummary || previous?.traceSummary ? {
5865
+ averageCandidateTopK: (current.traceSummary?.averageCandidateTopK ?? 0) - (previous?.traceSummary?.averageCandidateTopK ?? 0),
5866
+ averageFinalCount: (current.traceSummary?.averageFinalCount ?? 0) - (previous?.traceSummary?.averageFinalCount ?? 0),
5867
+ averageLexicalCount: (current.traceSummary?.averageLexicalCount ?? 0) - (previous?.traceSummary?.averageLexicalCount ?? 0),
5868
+ averageLexicalTopK: (current.traceSummary?.averageLexicalTopK ?? 0) - (previous?.traceSummary?.averageLexicalTopK ?? 0),
5869
+ averageVectorCount: (current.traceSummary?.averageVectorCount ?? 0) - (previous?.traceSummary?.averageVectorCount ?? 0),
5870
+ lexicalCases: (current.traceSummary?.lexicalCases ?? 0) - (previous?.traceSummary?.lexicalCases ?? 0),
5871
+ modesChanged: (current.traceSummary?.modes ?? []).join("|") !== (previous?.traceSummary?.modes ?? []).join("|"),
5872
+ stageCounts: diffTraceStageCounts({
5873
+ current: current.traceSummary?.stageCounts ?? {},
5874
+ previous: previous?.traceSummary?.stageCounts ?? {}
5875
+ }),
5876
+ transformedCases: (current.traceSummary?.transformedCases ?? 0) - (previous?.traceSummary?.transformedCases ?? 0),
5877
+ variantCases: (current.traceSummary?.variantCases ?? 0) - (previous?.traceSummary?.variantCases ?? 0),
5878
+ vectorCases: (current.traceSummary?.vectorCases ?? 0) - (previous?.traceSummary?.vectorCases ?? 0)
5879
+ } : undefined,
5687
5880
  unchangedCases
5688
5881
  };
5689
5882
  };
@@ -6057,7 +6250,8 @@ var compareRAGRerankers = async ({
6057
6250
  label: entry.label,
6058
6251
  response: entry.response,
6059
6252
  startedAt: 0,
6060
- suiteId: suite.id
6253
+ suiteId: suite.id,
6254
+ traceSummary: entry.traceSummary
6061
6255
  })));
6062
6256
  return {
6063
6257
  entries,
@@ -6142,7 +6336,8 @@ var compareRAGRetrievalStrategies = async ({
6142
6336
  label: entry.label,
6143
6337
  response: entry.response,
6144
6338
  startedAt: 0,
6145
- suiteId: suite.id
6339
+ suiteId: suite.id,
6340
+ traceSummary: entry.traceSummary
6146
6341
  })));
6147
6342
  return {
6148
6343
  entries,
@@ -11387,7 +11582,10 @@ export {
11387
11582
  buildRAGEvaluationRunDiff,
11388
11583
  buildRAGEvaluationResponse,
11389
11584
  buildRAGEvaluationLeaderboard,
11585
+ buildRAGEvaluationHistoryRows,
11390
11586
  buildRAGContext,
11587
+ buildRAGComparisonTraceSummaryRows,
11588
+ buildRAGComparisonTraceDiffRows,
11391
11589
  buildRAGCitations,
11392
11590
  buildRAGCitationReferenceMap,
11393
11591
  buildRAGAnswerGroundingEvaluationRunDiff,
@@ -11403,5 +11601,5 @@ export {
11403
11601
  aiChat
11404
11602
  };
11405
11603
 
11406
- //# debugId=3553069554C7849B64756E2164756E21
11604
+ //# debugId=3488A515C459B2F064756E2164756E21
11407
11605
  //# sourceMappingURL=index.js.map