@oneuptime/common 11.5.3 → 11.5.5
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/Models/DatabaseModels/AIAgentTaskPullRequest.ts +56 -45
- package/Models/DatabaseModels/AIInsight.ts +748 -0
- package/Models/DatabaseModels/AIRun.ts +262 -0
- package/Models/DatabaseModels/CodeRepository.ts +42 -0
- package/Models/DatabaseModels/Dashboard.ts +3 -3
- package/Models/DatabaseModels/GlobalConfig.ts +105 -0
- package/Models/DatabaseModels/Index.ts +4 -8
- package/Models/DatabaseModels/InstanceHealthLog.ts +277 -0
- package/Models/DatabaseModels/LlmLog.ts +33 -0
- package/Models/DatabaseModels/Project.ts +123 -6
- package/Models/DatabaseModels/Runbook.ts +1 -1
- package/Server/API/AIAgentDataAPI.ts +802 -140
- package/Server/API/AIAgentTaskAPI.ts +132 -69
- package/Server/API/AIAgentTaskLogAPI.ts +65 -41
- package/Server/API/AIAgentTaskPullRequestAPI.ts +41 -0
- package/Server/API/AIInsightAPI.ts +220 -0
- package/Server/API/AIInvestigationAPI.ts +363 -1
- package/Server/API/CodeFixRunAPI.ts +251 -0
- package/Server/API/GitHubAPI.ts +246 -3
- package/Server/API/TelemetryExceptionAPI.ts +193 -48
- package/Server/EnvironmentConfig.ts +1 -1
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783943300000-DropServiceCodeRepository.ts +26 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783947444597-CodeFixRunsOnAIRun.ts +73 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783950962813-AddCodeFixTaskType.ts +29 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783954040984-AddInstrumentationFixTasksFlag.ts +31 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783957160597-MigrationName.ts +51 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783958542237-AddTaskContextToAIRun.ts +25 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783965945957-AddInvestigationVerdictAndGrade.ts +47 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783970619301-AddFixGuardrailColumns.ts +53 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783970619302-DropLegacyAIAgentTaskTables.ts +31 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783977781677-AddCiStatusToFixPullRequests.ts +37 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1783990000000-AddCompletionTokensToLlmLog.ts +25 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1784010274993-AddSentinelInsight.ts +75 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1784010274994-AddSentinelInsightFlags.ts +37 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1784030612266-RenameSentinelToAI.ts +203 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/1784033837629-MigrationName.ts +163 -0
- package/Server/Infrastructure/Postgres/SchemaMigrations/Index.ts +30 -0
- package/Server/Infrastructure/Queue.ts +82 -4
- package/Server/Infrastructure/Semaphore.ts +62 -1
- package/Server/Services/AIAgentService.ts +32 -0
- package/Server/Services/AIAgentTaskPullRequestService.ts +117 -0
- package/Server/Services/AIInsightService.ts +174 -0
- package/Server/Services/AIRunEventService.ts +63 -0
- package/Server/Services/AIRunService.ts +259 -1
- package/Server/Services/AIService.ts +249 -7
- package/Server/Services/AlertService.ts +4 -4
- package/Server/Services/CodeRepositoryService.ts +209 -0
- package/Server/Services/DashboardService.ts +11 -2
- package/Server/Services/GlobalConfigService.ts +148 -0
- package/Server/Services/IncidentService.ts +183 -38
- package/Server/Services/IncidentStateTimelineService.ts +24 -4
- package/Server/Services/Index.ts +4 -4
- package/Server/Services/{AIAgentTaskLogService.ts → InstanceHealthLogService.ts} +1 -1
- package/Server/Services/LlmLogService.ts +27 -0
- package/Server/Services/LlmProviderService.ts +108 -0
- package/Server/Services/MetricBaselineService.ts +123 -0
- package/Server/Services/MonitorService.ts +15 -0
- package/Server/Services/TelemetryExceptionService.ts +323 -116
- package/Server/Services/TraceAggregationService.ts +176 -0
- package/Server/Types/Workflow/ComponentCode.ts +6 -0
- package/Server/Types/Workflow/Components/AI/GenerateText.ts +326 -0
- package/Server/Types/Workflow/Components/Index.ts +2 -0
- package/Server/Utils/AI/Chat/ObservabilityAssistant.ts +1 -1
- package/Server/Utils/AI/CodeFix/CodeAgentWorkspaceGuard.ts +71 -0
- package/Server/Utils/AI/CodeFix/CodeFixAgentCompletion.ts +243 -0
- package/Server/Utils/AI/CodeFix/CodeFixRunQueue.ts +131 -0
- package/Server/Utils/AI/CodeFix/FixRunBudget.ts +148 -0
- package/Server/Utils/AI/CodeFix/OpenPullRequestCap.ts +117 -0
- package/Server/Utils/AI/Eval/EvalCorpus.ts +369 -0
- package/Server/Utils/AI/Eval/EvalScores.ts +172 -0
- package/Server/Utils/AI/Eval/ReplayInvestigation.ts +87 -0
- package/Server/Utils/AI/PerfEvidence/SpanTreeAnalyzer.ts +676 -0
- package/Server/Utils/AI/{Sentinel/SentinelInvestigationEngine.ts → SRE/AIInvestigationEngine.ts} +58 -37
- package/Server/Utils/AI/{Sentinel/SentinelMemory.ts → SRE/AIMemory.ts} +3 -3
- package/Server/Utils/AI/{Sentinel → SRE}/AlertInvestigationRunner.ts +50 -20
- package/Server/Utils/AI/SRE/ConfidenceSignal.ts +293 -0
- package/Server/Utils/AI/SRE/FixFromIncidentTaskTrigger.ts +115 -0
- package/Server/Utils/AI/SRE/FixPerformanceTaskTrigger.ts +223 -0
- package/Server/Utils/AI/{Sentinel → SRE}/IncidentInvestigationRunner.ts +71 -21
- package/Server/Utils/AI/{Sentinel → SRE}/IncidentPostmortemRunner.ts +8 -8
- package/Server/Utils/AI/SRE/Insights/Detectors/ErrorLogSpikeDetector.ts +364 -0
- package/Server/Utils/AI/SRE/Insights/Detectors/ExceptionSpikeDetector.ts +317 -0
- package/Server/Utils/AI/SRE/Insights/Detectors/Index.ts +27 -0
- package/Server/Utils/AI/SRE/Insights/Detectors/MetricDriftDetector.ts +221 -0
- package/Server/Utils/AI/SRE/Insights/Detectors/NewExceptionDetector.ts +238 -0
- package/Server/Utils/AI/SRE/Insights/Detectors/TraceLatencyRegressionDetector.ts +521 -0
- package/Server/Utils/AI/SRE/Insights/FixRouting.ts +242 -0
- package/Server/Utils/AI/SRE/Insights/InsightScanner.ts +290 -0
- package/Server/Utils/AI/SRE/Insights/InsightStore.ts +283 -0
- package/Server/Utils/AI/SRE/Insights/InsightTriageRunner.ts +242 -0
- package/Server/Utils/AI/SRE/Insights/Triage.ts +162 -0
- package/Server/Utils/AI/SRE/Insights/Types.ts +49 -0
- package/Server/Utils/AI/SRE/InstrumentationTaskTrigger.ts +212 -0
- package/Server/Utils/AI/SRE/InvestigationGrader.ts +282 -0
- package/Server/Utils/AI/{Sentinel → SRE}/InvestigationQueue.ts +114 -19
- package/Server/Utils/AI/{Sentinel → SRE}/README.md +4 -4
- package/Server/Utils/AI/SRE/SubjectCodeFixRun.ts +172 -0
- package/Server/Utils/AI/Toolbox/{SentinelActionTools.ts → AIActionTools.ts} +22 -1
- package/Server/Utils/AI/Toolbox/Index.ts +2 -2
- package/Server/Utils/AnalyticsDatabase/ClickhouseCapacity.ts +526 -0
- package/Server/Utils/CodeRepository/GitHub/GitHub.ts +319 -1
- package/Server/Utils/CodeRepository/ServiceRepoLinkSuggester.ts +221 -0
- package/Server/Utils/CodeRepository/StackTraceRepoResolver.ts +571 -0
- package/Server/Utils/Execute.ts +34 -13
- package/Server/Utils/LLM/LLMService.ts +160 -30
- package/Server/Utils/Monitor/Criteria/SqlMonitorCriteria.ts +152 -0
- package/Server/Utils/Monitor/MonitorCriteriaEvaluator.ts +29 -0
- package/Tests/Server/API/DashboardMasterPasswordAPI.test.ts +225 -0
- package/Tests/Server/Infrastructure/Queue.test.ts +227 -0
- package/Tests/Server/Infrastructure/SemaphorePermit.test.ts +159 -0
- package/Tests/Server/Services/AIAgentTaskPullRequestOutcomeStats.test.ts +154 -0
- package/Tests/Server/Services/AIInsightService.test.ts +324 -0
- package/Tests/Server/Services/AIRunCodeFixClaim.test.ts +401 -0
- package/Tests/Server/Services/AIRunHumanVerdict.test.ts +161 -0
- package/Tests/Server/Services/AIServiceDailyBudget.test.ts +95 -5
- package/Tests/Server/Services/AIServiceProjectAccess.test.ts +259 -0
- package/Tests/Server/Services/FixFromIncidentTaskTrigger.test.ts +249 -0
- package/Tests/Server/Services/FixPerformanceTaskTrigger.test.ts +274 -0
- package/Tests/Server/Services/GlobalConfigService.test.ts +80 -0
- package/Tests/Server/Services/IncidentInternalNoteAnnouncement.test.ts +1 -1
- package/Tests/Server/Services/InstrumentationTaskTrigger.test.ts +313 -0
- package/Tests/Server/Services/LlmProviderProjectOwned.test.ts +132 -0
- package/Tests/Server/Services/TelemetryExceptionAIFixReadiness.test.ts +405 -0
- package/Tests/Server/Services/TelemetryExceptionCodeFixRun.test.ts +322 -0
- package/Tests/Server/Services/TraceServiceLatencyProfile.test.ts +278 -0
- package/Tests/Server/Types/Workflow/Components/AIGenerateText.test.ts +722 -0
- package/Tests/Server/Utils/AI/{SentinelAlertGating.test.ts → AIAlertGating.test.ts} +19 -19
- package/Tests/Server/Utils/AI/AIConfidenceSignal.test.ts +362 -0
- package/Tests/Server/Utils/AI/{SentinelInvestigationQueue.test.ts → AIInvestigationQueue.test.ts} +19 -19
- package/Tests/Server/Utils/AI/CodeAgentWorkspaceGuard.test.ts +138 -0
- package/Tests/Server/Utils/AI/CodeFixAgentCompletion.test.ts +335 -0
- package/Tests/Server/Utils/AI/CodeFixRunQueue.test.ts +134 -0
- package/Tests/Server/Utils/AI/EvalCorpus.test.ts +209 -0
- package/Tests/Server/Utils/AI/EvalScores.test.ts +343 -0
- package/Tests/Server/Utils/AI/FixRunBudget.test.ts +218 -0
- package/Tests/Server/Utils/AI/Insights/ErrorLogSpikeDetector.test.ts +437 -0
- package/Tests/Server/Utils/AI/Insights/ExceptionSpikeDetector.test.ts +287 -0
- package/Tests/Server/Utils/AI/Insights/InsightDetectorsIndex.test.ts +34 -0
- package/Tests/Server/Utils/AI/Insights/InsightFixRouting.test.ts +572 -0
- package/Tests/Server/Utils/AI/Insights/InsightNeverThrowsHardening.test.ts +344 -0
- package/Tests/Server/Utils/AI/Insights/InsightScanner.test.ts +720 -0
- package/Tests/Server/Utils/AI/Insights/InsightStore.test.ts +467 -0
- package/Tests/Server/Utils/AI/Insights/InsightStoreHardening.test.ts +370 -0
- package/Tests/Server/Utils/AI/Insights/InsightTriage.test.ts +210 -0
- package/Tests/Server/Utils/AI/Insights/InsightTriageRunner.test.ts +221 -0
- package/Tests/Server/Utils/AI/Insights/InvestigationQueueInsightSubject.test.ts +328 -0
- package/Tests/Server/Utils/AI/Insights/MetricBaselineDrift.test.ts +182 -0
- package/Tests/Server/Utils/AI/Insights/MetricDriftDetector.test.ts +280 -0
- package/Tests/Server/Utils/AI/Insights/NewExceptionDetector.test.ts +259 -0
- package/Tests/Server/Utils/AI/Insights/TraceLatencyRegressionDetector.test.ts +770 -0
- package/Tests/Server/Utils/AI/InvestigationGrader.test.ts +343 -0
- package/Tests/Server/Utils/AI/LLMServiceRequestPolicy.test.ts +307 -0
- package/Tests/Server/Utils/AI/OpenPullRequestCap.test.ts +170 -0
- package/Tests/Server/Utils/AI/SpanTreeAnalyzer.test.ts +559 -0
- package/Tests/Server/Utils/AnalyticsDatabase/ClickhouseCapacity.test.ts +158 -0
- package/Tests/Server/Utils/GitHubCheckRunsConclusion.test.ts +100 -0
- package/Tests/Server/Utils/GitHubPullRequestStateMapping.test.ts +50 -0
- package/Tests/Server/Utils/Monitor/Criteria/SqlMonitorCriteria.test.ts +266 -0
- package/Tests/Server/Utils/ServiceRepoLinkSuggester.test.ts +167 -0
- package/Tests/Server/Utils/StackTraceRepoResolver.test.ts +527 -0
- package/Tests/Types/AI/AIChatPermissionMode.test.ts +112 -0
- package/Tests/Types/AI/CodeFixTaskType.test.ts +152 -0
- package/Tests/Types/AI/FixPullRequestCiStatus.test.ts +181 -0
- package/Tests/Types/Monitor/MonitorStepSqlMonitor.test.ts +110 -0
- package/Tests/Types/Monitor/SqlDatabaseType.test.ts +55 -0
- package/Tests/Types/Workflow/AIGenerateTextMetadata.test.ts +180 -0
- package/Tests/UI/Components/TableBulkCsvExport.test.tsx +168 -0
- package/Tests/UI/Utils/TableColumnsToCsv.test.ts +523 -0
- package/Types/AI/AIChatTypes.ts +8 -0
- package/Types/AI/AIInsightEvidence.ts +77 -0
- package/Types/AI/AIInsightHumanVerdict.ts +15 -0
- package/Types/AI/AIInsightSeverity.ts +13 -0
- package/Types/AI/AIInsightStatus.ts +38 -0
- package/Types/AI/AIInsightType.ts +41 -0
- package/Types/AI/AIRunAutoGrade.ts +18 -0
- package/Types/AI/AIRunEventType.ts +5 -0
- package/Types/AI/AIRunHumanVerdict.ts +17 -0
- package/Types/AI/AIRunType.ts +6 -0
- package/Types/AI/CodeFixTaskContext.ts +88 -0
- package/Types/AI/CodeFixTaskType.ts +132 -0
- package/Types/AI/FixPullRequestCiStatus.ts +132 -0
- package/Types/BaseDatabase/AggregationIntervalUtil.ts +67 -0
- package/Types/Incident/IncidentMetricType.ts +7 -0
- package/Types/Monitor/CriteriaFilter.ts +8 -0
- package/Types/Monitor/MonitorCriteriaInstance.ts +67 -0
- package/Types/Monitor/MonitorStep.ts +38 -0
- package/Types/Monitor/MonitorStepSqlMonitor.ts +141 -0
- package/Types/Monitor/MonitorType.ts +17 -0
- package/Types/Monitor/SqlDatabaseType.ts +49 -0
- package/Types/Monitor/SqlMonitor/SqlMonitorResponse.ts +28 -0
- package/Types/Permission.ts +11 -95
- package/Types/Probe/ProbeMonitorResponse.ts +2 -0
- package/Types/Runbook/RunbookStep.ts +26 -2
- package/Types/Runbook/RunbookStepType.ts +1 -0
- package/Types/Workflow/Component.ts +13 -0
- package/Types/Workflow/ComponentID.ts +1 -0
- package/Types/Workflow/Components/AI.ts +148 -0
- package/Types/Workflow/Components.ts +8 -0
- package/UI/Components/AI/GenerateFromAIModal.tsx +1 -1
- package/UI/Components/Markdown.tsx/MarkdownViewer.tsx +11 -0
- package/UI/Components/MoreMenu/MoreMenu.tsx +1 -1
- package/UI/Components/Table/Table.tsx +38 -2
- package/UI/Utils/TableColumnsToCsv.ts +385 -0
- package/Utils/Monitor/MonitorMetricType.ts +1 -0
- package/build/dist/Models/DatabaseModels/AIAgentTaskPullRequest.js +59 -45
- package/build/dist/Models/DatabaseModels/AIAgentTaskPullRequest.js.map +1 -1
- package/build/dist/Models/DatabaseModels/AIInsight.js +795 -0
- package/build/dist/Models/DatabaseModels/AIInsight.js.map +1 -0
- package/build/dist/Models/DatabaseModels/AIRun.js +274 -0
- package/build/dist/Models/DatabaseModels/AIRun.js.map +1 -1
- package/build/dist/Models/DatabaseModels/CodeRepository.js +43 -0
- package/build/dist/Models/DatabaseModels/CodeRepository.js.map +1 -1
- package/build/dist/Models/DatabaseModels/Dashboard.js +3 -3
- package/build/dist/Models/DatabaseModels/Dashboard.js.map +1 -1
- package/build/dist/Models/DatabaseModels/GlobalConfig.js +110 -0
- package/build/dist/Models/DatabaseModels/GlobalConfig.js.map +1 -1
- package/build/dist/Models/DatabaseModels/Index.js +4 -8
- package/build/dist/Models/DatabaseModels/Index.js.map +1 -1
- package/build/dist/Models/DatabaseModels/InstanceHealthLog.js +304 -0
- package/build/dist/Models/DatabaseModels/InstanceHealthLog.js.map +1 -0
- package/build/dist/Models/DatabaseModels/LlmLog.js +35 -0
- package/build/dist/Models/DatabaseModels/LlmLog.js.map +1 -1
- package/build/dist/Models/DatabaseModels/Project.js +127 -6
- package/build/dist/Models/DatabaseModels/Project.js.map +1 -1
- package/build/dist/Models/DatabaseModels/Runbook.js +1 -1
- package/build/dist/Models/DatabaseModels/Runbook.js.map +1 -1
- package/build/dist/Server/API/AIAgentDataAPI.js +568 -85
- package/build/dist/Server/API/AIAgentDataAPI.js.map +1 -1
- package/build/dist/Server/API/AIAgentTaskAPI.js +100 -44
- package/build/dist/Server/API/AIAgentTaskAPI.js.map +1 -1
- package/build/dist/Server/API/AIAgentTaskLogAPI.js +49 -25
- package/build/dist/Server/API/AIAgentTaskLogAPI.js.map +1 -1
- package/build/dist/Server/API/AIAgentTaskPullRequestAPI.js +27 -0
- package/build/dist/Server/API/AIAgentTaskPullRequestAPI.js.map +1 -1
- package/build/dist/Server/API/AIInsightAPI.js +140 -0
- package/build/dist/Server/API/AIInsightAPI.js.map +1 -0
- package/build/dist/Server/API/AIInvestigationAPI.js +257 -1
- package/build/dist/Server/API/AIInvestigationAPI.js.map +1 -1
- package/build/dist/Server/API/CodeFixRunAPI.js +181 -0
- package/build/dist/Server/API/CodeFixRunAPI.js.map +1 -0
- package/build/dist/Server/API/GitHubAPI.js +164 -2
- package/build/dist/Server/API/GitHubAPI.js.map +1 -1
- package/build/dist/Server/API/TelemetryExceptionAPI.js +148 -43
- package/build/dist/Server/API/TelemetryExceptionAPI.js.map +1 -1
- package/build/dist/Server/EnvironmentConfig.js +1 -1
- package/build/dist/Server/EnvironmentConfig.js.map +1 -1
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783943300000-DropServiceCodeRepository.js +23 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783943300000-DropServiceCodeRepository.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783947444597-CodeFixRunsOnAIRun.js +32 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783947444597-CodeFixRunsOnAIRun.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783950962813-AddCodeFixTaskType.js +16 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783950962813-AddCodeFixTaskType.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783954040984-AddInstrumentationFixTasksFlag.js +16 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783954040984-AddInstrumentationFixTasksFlag.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783957160597-MigrationName.js +24 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783957160597-MigrationName.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783958542237-AddTaskContextToAIRun.js +16 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783958542237-AddTaskContextToAIRun.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783965945957-AddInvestigationVerdictAndGrade.js +24 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783965945957-AddInvestigationVerdictAndGrade.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783970619301-AddFixGuardrailColumns.js +24 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783970619301-AddFixGuardrailColumns.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783970619302-DropLegacyAIAgentTaskTables.js +24 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783970619302-DropLegacyAIAgentTaskTables.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783977781677-AddCiStatusToFixPullRequests.js +18 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783977781677-AddCiStatusToFixPullRequests.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783990000000-AddCompletionTokensToLlmLog.js +18 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1783990000000-AddCompletionTokensToLlmLog.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1784010274993-AddSentinelInsight.js +32 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1784010274993-AddSentinelInsight.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1784010274994-AddSentinelInsightFlags.js +18 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1784010274994-AddSentinelInsightFlags.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1784030612266-RenameSentinelToAI.js +125 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1784030612266-RenameSentinelToAI.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1784033837629-MigrationName.js +62 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/1784033837629-MigrationName.js.map +1 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/Index.js +30 -0
- package/build/dist/Server/Infrastructure/Postgres/SchemaMigrations/Index.js.map +1 -1
- package/build/dist/Server/Infrastructure/Queue.js +73 -3
- package/build/dist/Server/Infrastructure/Queue.js.map +1 -1
- package/build/dist/Server/Infrastructure/Semaphore.js +44 -1
- package/build/dist/Server/Infrastructure/Semaphore.js.map +1 -1
- package/build/dist/Server/Services/AIAgentService.js +29 -0
- package/build/dist/Server/Services/AIAgentService.js.map +1 -1
- package/build/dist/Server/Services/AIAgentTaskPullRequestService.js +78 -0
- package/build/dist/Server/Services/AIAgentTaskPullRequestService.js.map +1 -1
- package/build/dist/Server/Services/AIInsightService.js +164 -0
- package/build/dist/Server/Services/AIInsightService.js.map +1 -0
- package/build/dist/Server/Services/AIRunEventService.js +49 -0
- package/build/dist/Server/Services/AIRunEventService.js.map +1 -1
- package/build/dist/Server/Services/AIRunService.js +221 -0
- package/build/dist/Server/Services/AIRunService.js.map +1 -1
- package/build/dist/Server/Services/AIService.js +198 -11
- package/build/dist/Server/Services/AIService.js.map +1 -1
- package/build/dist/Server/Services/AlertService.js +4 -4
- package/build/dist/Server/Services/AlertService.js.map +1 -1
- package/build/dist/Server/Services/CodeRepositoryService.js +180 -0
- package/build/dist/Server/Services/CodeRepositoryService.js.map +1 -1
- package/build/dist/Server/Services/DashboardService.js +8 -2
- package/build/dist/Server/Services/DashboardService.js.map +1 -1
- package/build/dist/Server/Services/GlobalConfigService.js +84 -0
- package/build/dist/Server/Services/GlobalConfigService.js.map +1 -1
- package/build/dist/Server/Services/IncidentService.js +149 -37
- package/build/dist/Server/Services/IncidentService.js.map +1 -1
- package/build/dist/Server/Services/IncidentStateTimelineService.js +24 -4
- package/build/dist/Server/Services/IncidentStateTimelineService.js.map +1 -1
- package/build/dist/Server/Services/Index.js +4 -4
- package/build/dist/Server/Services/Index.js.map +1 -1
- package/build/dist/Server/Services/{AIAgentTaskLogService.js → InstanceHealthLogService.js} +2 -2
- package/build/dist/Server/Services/InstanceHealthLogService.js.map +1 -0
- package/build/dist/Server/Services/LlmLogService.js +23 -0
- package/build/dist/Server/Services/LlmLogService.js.map +1 -1
- package/build/dist/Server/Services/LlmProviderService.js +96 -0
- package/build/dist/Server/Services/LlmProviderService.js.map +1 -1
- package/build/dist/Server/Services/MetricBaselineService.js +65 -0
- package/build/dist/Server/Services/MetricBaselineService.js.map +1 -1
- package/build/dist/Server/Services/MonitorService.js +9 -1
- package/build/dist/Server/Services/MonitorService.js.map +1 -1
- package/build/dist/Server/Services/TelemetryExceptionService.js +243 -82
- package/build/dist/Server/Services/TelemetryExceptionService.js.map +1 -1
- package/build/dist/Server/Services/TraceAggregationService.js +124 -0
- package/build/dist/Server/Services/TraceAggregationService.js.map +1 -1
- package/build/dist/Server/Types/Workflow/ComponentCode.js.map +1 -1
- package/build/dist/Server/Types/Workflow/Components/AI/GenerateText.js +240 -0
- package/build/dist/Server/Types/Workflow/Components/AI/GenerateText.js.map +1 -0
- package/build/dist/Server/Types/Workflow/Components/Index.js +2 -0
- package/build/dist/Server/Types/Workflow/Components/Index.js.map +1 -1
- package/build/dist/Server/Utils/AI/CodeFix/CodeAgentWorkspaceGuard.js +48 -0
- package/build/dist/Server/Utils/AI/CodeFix/CodeAgentWorkspaceGuard.js.map +1 -0
- package/build/dist/Server/Utils/AI/CodeFix/CodeFixAgentCompletion.js +174 -0
- package/build/dist/Server/Utils/AI/CodeFix/CodeFixAgentCompletion.js.map +1 -0
- package/build/dist/Server/Utils/AI/CodeFix/CodeFixRunQueue.js +126 -0
- package/build/dist/Server/Utils/AI/CodeFix/CodeFixRunQueue.js.map +1 -0
- package/build/dist/Server/Utils/AI/CodeFix/FixRunBudget.js +131 -0
- package/build/dist/Server/Utils/AI/CodeFix/FixRunBudget.js.map +1 -0
- package/build/dist/Server/Utils/AI/CodeFix/OpenPullRequestCap.js +99 -0
- package/build/dist/Server/Utils/AI/CodeFix/OpenPullRequestCap.js.map +1 -0
- package/build/dist/Server/Utils/AI/Eval/EvalCorpus.js +245 -0
- package/build/dist/Server/Utils/AI/Eval/EvalCorpus.js.map +1 -0
- package/build/dist/Server/Utils/AI/Eval/EvalScores.js +83 -0
- package/build/dist/Server/Utils/AI/Eval/EvalScores.js.map +1 -0
- package/build/dist/Server/Utils/AI/Eval/ReplayInvestigation.js +12 -0
- package/build/dist/Server/Utils/AI/Eval/ReplayInvestigation.js.map +1 -0
- package/build/dist/Server/Utils/AI/PerfEvidence/SpanTreeAnalyzer.js +439 -0
- package/build/dist/Server/Utils/AI/PerfEvidence/SpanTreeAnalyzer.js.map +1 -0
- package/build/dist/Server/Utils/AI/{Sentinel/SentinelInvestigationEngine.js → SRE/AIInvestigationEngine.js} +47 -31
- package/build/dist/Server/Utils/AI/SRE/AIInvestigationEngine.js.map +1 -0
- package/build/dist/Server/Utils/AI/{Sentinel/SentinelMemory.js → SRE/AIMemory.js} +5 -5
- package/build/dist/Server/Utils/AI/SRE/AIMemory.js.map +1 -0
- package/build/dist/Server/Utils/AI/{Sentinel → SRE}/AlertInvestigationRunner.js +42 -19
- package/build/dist/Server/Utils/AI/SRE/AlertInvestigationRunner.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/ConfidenceSignal.js +226 -0
- package/build/dist/Server/Utils/AI/SRE/ConfidenceSignal.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/FixFromIncidentTaskTrigger.js +100 -0
- package/build/dist/Server/Utils/AI/SRE/FixFromIncidentTaskTrigger.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/FixPerformanceTaskTrigger.js +174 -0
- package/build/dist/Server/Utils/AI/SRE/FixPerformanceTaskTrigger.js.map +1 -0
- package/build/dist/Server/Utils/AI/{Sentinel → SRE}/IncidentInvestigationRunner.js +60 -22
- package/build/dist/Server/Utils/AI/SRE/IncidentInvestigationRunner.js.map +1 -0
- package/build/dist/Server/Utils/AI/{Sentinel → SRE}/IncidentPostmortemRunner.js +9 -9
- package/build/dist/Server/Utils/AI/SRE/IncidentPostmortemRunner.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/ErrorLogSpikeDetector.js +268 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/ErrorLogSpikeDetector.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/ExceptionSpikeDetector.js +226 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/ExceptionSpikeDetector.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/Index.js +26 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/Index.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/MetricDriftDetector.js +159 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/MetricDriftDetector.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/NewExceptionDetector.js +177 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/NewExceptionDetector.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/TraceLatencyRegressionDetector.js +384 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Detectors/TraceLatencyRegressionDetector.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/FixRouting.js +180 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/FixRouting.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/InsightScanner.js +269 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/InsightScanner.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/InsightStore.js +231 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/InsightStore.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/InsightTriageRunner.js +206 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/InsightTriageRunner.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Triage.js +121 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Triage.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Types.js +2 -0
- package/build/dist/Server/Utils/AI/SRE/Insights/Types.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/InstrumentationTaskTrigger.js +148 -0
- package/build/dist/Server/Utils/AI/SRE/InstrumentationTaskTrigger.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/InvestigationGrader.js +231 -0
- package/build/dist/Server/Utils/AI/SRE/InvestigationGrader.js.map +1 -0
- package/build/dist/Server/Utils/AI/{Sentinel → SRE}/InvestigationQueue.js +102 -23
- package/build/dist/Server/Utils/AI/SRE/InvestigationQueue.js.map +1 -0
- package/build/dist/Server/Utils/AI/SRE/SubjectCodeFixRun.js +167 -0
- package/build/dist/Server/Utils/AI/SRE/SubjectCodeFixRun.js.map +1 -0
- package/build/dist/Server/Utils/AI/Toolbox/{SentinelActionTools.js → AIActionTools.js} +18 -2
- package/build/dist/Server/Utils/AI/Toolbox/AIActionTools.js.map +1 -0
- package/build/dist/Server/Utils/AI/Toolbox/Index.js +2 -2
- package/build/dist/Server/Utils/AI/Toolbox/Index.js.map +1 -1
- package/build/dist/Server/Utils/AnalyticsDatabase/ClickhouseCapacity.js +301 -0
- package/build/dist/Server/Utils/AnalyticsDatabase/ClickhouseCapacity.js.map +1 -0
- package/build/dist/Server/Utils/CodeRepository/GitHub/GitHub.js +240 -2
- package/build/dist/Server/Utils/CodeRepository/GitHub/GitHub.js.map +1 -1
- package/build/dist/Server/Utils/CodeRepository/ServiceRepoLinkSuggester.js +127 -0
- package/build/dist/Server/Utils/CodeRepository/ServiceRepoLinkSuggester.js.map +1 -0
- package/build/dist/Server/Utils/CodeRepository/StackTraceRepoResolver.js +336 -0
- package/build/dist/Server/Utils/CodeRepository/StackTraceRepoResolver.js.map +1 -0
- package/build/dist/Server/Utils/Execute.js +21 -11
- package/build/dist/Server/Utils/Execute.js.map +1 -1
- package/build/dist/Server/Utils/LLM/LLMService.js +125 -25
- package/build/dist/Server/Utils/LLM/LLMService.js.map +1 -1
- package/build/dist/Server/Utils/Monitor/Criteria/SqlMonitorCriteria.js +120 -0
- package/build/dist/Server/Utils/Monitor/Criteria/SqlMonitorCriteria.js.map +1 -0
- package/build/dist/Server/Utils/Monitor/MonitorCriteriaEvaluator.js +21 -0
- package/build/dist/Server/Utils/Monitor/MonitorCriteriaEvaluator.js.map +1 -1
- package/build/dist/Types/AI/AIInsightEvidence.js +2 -0
- package/build/dist/Types/AI/AIInsightEvidence.js.map +1 -0
- package/build/dist/Types/AI/AIInsightHumanVerdict.js +16 -0
- package/build/dist/Types/AI/AIInsightHumanVerdict.js.map +1 -0
- package/build/dist/Types/AI/AIInsightSeverity.js +14 -0
- package/build/dist/Types/AI/AIInsightSeverity.js.map +1 -0
- package/build/dist/Types/AI/AIInsightStatus.js +36 -0
- package/build/dist/Types/AI/AIInsightStatus.js.map +1 -0
- package/build/dist/Types/AI/AIInsightType.js +42 -0
- package/build/dist/Types/AI/AIInsightType.js.map +1 -0
- package/build/dist/Types/AI/AIRunAutoGrade.js +19 -0
- package/build/dist/Types/AI/AIRunAutoGrade.js.map +1 -0
- package/build/dist/Types/AI/AIRunEventType.js +5 -0
- package/build/dist/Types/AI/AIRunEventType.js.map +1 -1
- package/build/dist/Types/AI/AIRunHumanVerdict.js +18 -0
- package/build/dist/Types/AI/AIRunHumanVerdict.js.map +1 -0
- package/build/dist/Types/AI/AIRunType.js +6 -0
- package/build/dist/Types/AI/AIRunType.js.map +1 -1
- package/build/dist/Types/AI/CodeFixTaskContext.js +29 -0
- package/build/dist/Types/AI/CodeFixTaskContext.js.map +1 -0
- package/build/dist/Types/AI/CodeFixTaskType.js +122 -0
- package/build/dist/Types/AI/CodeFixTaskType.js.map +1 -0
- package/build/dist/Types/AI/FixPullRequestCiStatus.js +102 -0
- package/build/dist/Types/AI/FixPullRequestCiStatus.js.map +1 -0
- package/build/dist/Types/BaseDatabase/AggregationIntervalUtil.js +55 -0
- package/build/dist/Types/BaseDatabase/AggregationIntervalUtil.js.map +1 -1
- package/build/dist/Types/Incident/IncidentMetricType.js +7 -0
- package/build/dist/Types/Incident/IncidentMetricType.js.map +1 -1
- package/build/dist/Types/Monitor/CriteriaFilter.js +7 -0
- package/build/dist/Types/Monitor/CriteriaFilter.js.map +1 -1
- package/build/dist/Types/Monitor/MonitorCriteriaInstance.js +62 -0
- package/build/dist/Types/Monitor/MonitorCriteriaInstance.js.map +1 -1
- package/build/dist/Types/Monitor/MonitorStep.js +28 -0
- package/build/dist/Types/Monitor/MonitorStep.js.map +1 -1
- package/build/dist/Types/Monitor/MonitorStepSqlMonitor.js +93 -0
- package/build/dist/Types/Monitor/MonitorStepSqlMonitor.js.map +1 -0
- package/build/dist/Types/Monitor/MonitorType.js +15 -0
- package/build/dist/Types/Monitor/MonitorType.js.map +1 -1
- package/build/dist/Types/Monitor/SqlDatabaseType.js +47 -0
- package/build/dist/Types/Monitor/SqlDatabaseType.js.map +1 -0
- package/build/dist/Types/Monitor/SqlMonitor/SqlMonitorResponse.js +2 -0
- package/build/dist/Types/Monitor/SqlMonitor/SqlMonitorResponse.js.map +1 -0
- package/build/dist/Types/Permission.js +10 -84
- package/build/dist/Types/Permission.js.map +1 -1
- package/build/dist/Types/Runbook/RunbookStepType.js +1 -0
- package/build/dist/Types/Runbook/RunbookStepType.js.map +1 -1
- package/build/dist/Types/Workflow/ComponentID.js +1 -0
- package/build/dist/Types/Workflow/ComponentID.js.map +1 -1
- package/build/dist/Types/Workflow/Components/AI.js +135 -0
- package/build/dist/Types/Workflow/Components/AI.js.map +1 -0
- package/build/dist/Types/Workflow/Components.js +7 -0
- package/build/dist/Types/Workflow/Components.js.map +1 -1
- package/build/dist/UI/Components/AI/GenerateFromAIModal.js +1 -1
- package/build/dist/UI/Components/AI/GenerateFromAIModal.js.map +1 -1
- package/build/dist/UI/Components/Markdown.tsx/MarkdownViewer.js +9 -0
- package/build/dist/UI/Components/Markdown.tsx/MarkdownViewer.js.map +1 -1
- package/build/dist/UI/Components/MoreMenu/MoreMenu.js +1 -1
- package/build/dist/UI/Components/Table/Table.js +28 -2
- package/build/dist/UI/Components/Table/Table.js.map +1 -1
- package/build/dist/UI/Utils/TableColumnsToCsv.js +291 -0
- package/build/dist/UI/Utils/TableColumnsToCsv.js.map +1 -0
- package/build/dist/Utils/Monitor/MonitorMetricType.js +1 -0
- package/build/dist/Utils/Monitor/MonitorMetricType.js.map +1 -1
- package/package.json +1 -1
- package/Models/DatabaseModels/AIAgentTask.ts +0 -660
- package/Models/DatabaseModels/AIAgentTaskLog.ts +0 -426
- package/Models/DatabaseModels/AIAgentTaskTelemetryException.ts +0 -399
- package/Models/DatabaseModels/ServiceCodeRepository.ts +0 -647
- package/Server/Services/AIAgentTaskService.ts +0 -240
- package/Server/Services/AIAgentTaskTelemetryExceptionService.ts +0 -39
- package/Server/Services/ServiceCodeRepositoryService.ts +0 -55
- package/Tests/Server/Services/AIAgentTaskServiceClaim.test.ts +0 -146
- package/Types/AI/AIAgentTaskMetadata.ts +0 -25
- package/Types/AI/AIAgentTaskType.ts +0 -40
- package/build/dist/Models/DatabaseModels/AIAgentTask.js +0 -691
- package/build/dist/Models/DatabaseModels/AIAgentTask.js.map +0 -1
- package/build/dist/Models/DatabaseModels/AIAgentTaskLog.js +0 -447
- package/build/dist/Models/DatabaseModels/AIAgentTaskLog.js.map +0 -1
- package/build/dist/Models/DatabaseModels/AIAgentTaskTelemetryException.js +0 -415
- package/build/dist/Models/DatabaseModels/AIAgentTaskTelemetryException.js.map +0 -1
- package/build/dist/Models/DatabaseModels/ServiceCodeRepository.js +0 -665
- package/build/dist/Models/DatabaseModels/ServiceCodeRepository.js.map +0 -1
- package/build/dist/Server/Services/AIAgentTaskLogService.js.map +0 -1
- package/build/dist/Server/Services/AIAgentTaskService.js +0 -217
- package/build/dist/Server/Services/AIAgentTaskService.js.map +0 -1
- package/build/dist/Server/Services/AIAgentTaskTelemetryExceptionService.js +0 -36
- package/build/dist/Server/Services/AIAgentTaskTelemetryExceptionService.js.map +0 -1
- package/build/dist/Server/Services/ServiceCodeRepositoryService.js +0 -54
- package/build/dist/Server/Services/ServiceCodeRepositoryService.js.map +0 -1
- package/build/dist/Server/Utils/AI/Sentinel/AlertInvestigationRunner.js.map +0 -1
- package/build/dist/Server/Utils/AI/Sentinel/IncidentInvestigationRunner.js.map +0 -1
- package/build/dist/Server/Utils/AI/Sentinel/IncidentPostmortemRunner.js.map +0 -1
- package/build/dist/Server/Utils/AI/Sentinel/InvestigationQueue.js.map +0 -1
- package/build/dist/Server/Utils/AI/Sentinel/SentinelInvestigationEngine.js.map +0 -1
- package/build/dist/Server/Utils/AI/Sentinel/SentinelMemory.js.map +0 -1
- package/build/dist/Server/Utils/AI/Toolbox/SentinelActionTools.js.map +0 -1
- package/build/dist/Types/AI/AIAgentTaskMetadata.js +0 -6
- package/build/dist/Types/AI/AIAgentTaskMetadata.js.map +0 -1
- package/build/dist/Types/AI/AIAgentTaskType.js +0 -29
- package/build/dist/Types/AI/AIAgentTaskType.js.map +0 -1
|
@@ -0,0 +1,245 @@
|
|
|
1
|
+
var __decorate = (this && this.__decorate) || function (decorators, target, key, desc) {
|
|
2
|
+
var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
|
|
3
|
+
if (typeof Reflect === "object" && typeof Reflect.decorate === "function") r = Reflect.decorate(decorators, target, key, desc);
|
|
4
|
+
else for (var i = decorators.length - 1; i >= 0; i--) if (d = decorators[i]) r = (c < 3 ? d(r) : c > 3 ? d(target, key, r) : d(target, key)) || r;
|
|
5
|
+
return c > 3 && r && Object.defineProperty(target, key, r), r;
|
|
6
|
+
};
|
|
7
|
+
var __metadata = (this && this.__metadata) || function (k, v) {
|
|
8
|
+
if (typeof Reflect === "object" && typeof Reflect.metadata === "function") return Reflect.metadata(k, v);
|
|
9
|
+
};
|
|
10
|
+
import SortOrder from "../../../../Types/BaseDatabase/SortOrder";
|
|
11
|
+
import LIMIT_MAX from "../../../../Types/Database/LimitMax";
|
|
12
|
+
import AIRunType from "../../../../Types/AI/AIRunType";
|
|
13
|
+
import AIRunStatus from "../../../../Types/AI/AIRunStatus";
|
|
14
|
+
import AIRunEventType from "../../../../Types/AI/AIRunEventType";
|
|
15
|
+
import { IncidentFeedEventType } from "../../../../Models/DatabaseModels/IncidentFeed";
|
|
16
|
+
import { AlertFeedEventType } from "../../../../Models/DatabaseModels/AlertFeed";
|
|
17
|
+
import QueryHelper from "../../../Types/Database/QueryHelper";
|
|
18
|
+
import AIRunService from "../../../Services/AIRunService";
|
|
19
|
+
import AIRunEventService from "../../../Services/AIRunEventService";
|
|
20
|
+
import IncidentFeedService from "../../../Services/IncidentFeedService";
|
|
21
|
+
import AlertFeedService from "../../../Services/AlertFeedService";
|
|
22
|
+
import AIConfidenceSignal from "../SRE/ConfidenceSignal";
|
|
23
|
+
import logger from "../../Logger";
|
|
24
|
+
import CaptureSpan from "../../Telemetry/CaptureSpan";
|
|
25
|
+
/*
|
|
26
|
+
* AI eval harness — golden-incident corpus (G3 bootstrap).
|
|
27
|
+
*
|
|
28
|
+
* A GoldenCase is one LABELED, completed AI investigation, exported as
|
|
29
|
+
* plain JSON: the run's identity, its human/auto label (AIRun.humanVerdict /
|
|
30
|
+
* AIRun.autoGrade — the Phase 2 measurement columns), deterministic per-run
|
|
31
|
+
* evidence stats derived from the recorded AIRunEvent trail, and the trail
|
|
32
|
+
* itself in skeletal form. Everything here is computable WITHOUT re-running
|
|
33
|
+
* any LLM — that is the point of the bootstrap: recorded runs + labels make
|
|
34
|
+
* the four G3 scores (EvalScores.ts) measurable today.
|
|
35
|
+
*
|
|
36
|
+
* WHAT THE TRAIL DOES AND DOES NOT CONTAIN (consequence for replay):
|
|
37
|
+
* AIRunEvent stores the validated tool ARGUMENTS (`toolArguments`) and a
|
|
38
|
+
* result SUMMARY (`resultSummary`: rowCount, durationInMs, isTruncated,
|
|
39
|
+
* bytesSentToLlm, errorMessage) — never the tool's actual result rows. So the
|
|
40
|
+
* corpus supports labeled-run SCORING (this file + EvalScores.ts) but not
|
|
41
|
+
* offline REPLAY; see ReplayInvestigation.ts for exactly what full replay
|
|
42
|
+
* would need.
|
|
43
|
+
*
|
|
44
|
+
* "Data-bearing" is defined ONCE, in
|
|
45
|
+
* AIConfidenceSignal.isDataBearingRowCount (rowCount > 0; rowCount 0 is
|
|
46
|
+
* proof of absence — evidence, but not data-bearing), and reused here so the
|
|
47
|
+
* offline stats can never drift from the live confidence floor's definition.
|
|
48
|
+
*/
|
|
49
|
+
// Default number of labeled runs exported when the caller does not say.
|
|
50
|
+
export const DEFAULT_CORPUS_LIMIT = 200;
|
|
51
|
+
export default class EvalCorpus {
|
|
52
|
+
/*
|
|
53
|
+
* Derive the deterministic evidence stats from a run's recorded event trail
|
|
54
|
+
* (pure, exported for tests). Definitions, kept consistent with the live
|
|
55
|
+
* confidence floor:
|
|
56
|
+
* - citationsMinted: events carrying a citationId — the engine mints one
|
|
57
|
+
* citation per successful tool call, server-side, from validated args.
|
|
58
|
+
* - toolCallsTotal / toolCallsFailed: ToolCallCompleted + ToolCallFailed
|
|
59
|
+
* events. ToolCallStarted is deliberately NOT counted — a call that
|
|
60
|
+
* never finished (pod death mid-call) has no outcome to score.
|
|
61
|
+
* - dataBearingToolCalls: ToolCallCompleted events whose
|
|
62
|
+
* resultSummary.rowCount passes isDataBearingRowCount — the SAME helper
|
|
63
|
+
* the confidence floor applies to citation rowCounts.
|
|
64
|
+
*/
|
|
65
|
+
static deriveConfidenceStats(events) {
|
|
66
|
+
var _a;
|
|
67
|
+
let citationsMinted = 0;
|
|
68
|
+
let dataBearingToolCalls = 0;
|
|
69
|
+
let toolCallsTotal = 0;
|
|
70
|
+
let toolCallsFailed = 0;
|
|
71
|
+
for (const event of events) {
|
|
72
|
+
if (event.citationId) {
|
|
73
|
+
citationsMinted++;
|
|
74
|
+
}
|
|
75
|
+
if (event.eventType === AIRunEventType.ToolCallCompleted) {
|
|
76
|
+
toolCallsTotal++;
|
|
77
|
+
if (AIConfidenceSignal.isDataBearingRowCount((_a = event.resultSummary) === null || _a === void 0 ? void 0 : _a.rowCount)) {
|
|
78
|
+
dataBearingToolCalls++;
|
|
79
|
+
}
|
|
80
|
+
}
|
|
81
|
+
else if (event.eventType === AIRunEventType.ToolCallFailed) {
|
|
82
|
+
toolCallsTotal++;
|
|
83
|
+
toolCallsFailed++;
|
|
84
|
+
}
|
|
85
|
+
}
|
|
86
|
+
return {
|
|
87
|
+
citationsMinted,
|
|
88
|
+
dataBearingToolCalls,
|
|
89
|
+
toolCallsTotal,
|
|
90
|
+
toolCallsFailed,
|
|
91
|
+
};
|
|
92
|
+
}
|
|
93
|
+
// Skeletal event projection for the exported JSON (pure, exported for tests).
|
|
94
|
+
static toGoldenCaseEvents(events) {
|
|
95
|
+
return events.map((event) => {
|
|
96
|
+
let ok = undefined;
|
|
97
|
+
if (event.eventType === AIRunEventType.ToolCallCompleted) {
|
|
98
|
+
ok = true;
|
|
99
|
+
}
|
|
100
|
+
else if (event.eventType === AIRunEventType.ToolCallFailed) {
|
|
101
|
+
ok = false;
|
|
102
|
+
}
|
|
103
|
+
return {
|
|
104
|
+
sequence: event.sequence || 0,
|
|
105
|
+
eventType: (event.eventType || "").toString(),
|
|
106
|
+
toolName: event.toolName,
|
|
107
|
+
ok,
|
|
108
|
+
};
|
|
109
|
+
});
|
|
110
|
+
}
|
|
111
|
+
/*
|
|
112
|
+
* Export the golden corpus: completed Investigation runs that HAVE a label
|
|
113
|
+
* (humanVerdict or autoGrade set — unlabeled runs carry no ground truth and
|
|
114
|
+
* are deliberately excluded), newest first, with their event trails.
|
|
115
|
+
*/
|
|
116
|
+
static async exportCorpus(options) {
|
|
117
|
+
var _a;
|
|
118
|
+
const limit = Math.min(options.limit || DEFAULT_CORPUS_LIMIT, LIMIT_MAX);
|
|
119
|
+
const runSelect = {
|
|
120
|
+
_id: true,
|
|
121
|
+
projectId: true,
|
|
122
|
+
completedAt: true,
|
|
123
|
+
humanVerdict: true,
|
|
124
|
+
autoGrade: true,
|
|
125
|
+
triggeredByIncidentId: true,
|
|
126
|
+
triggeredByAlertId: true,
|
|
127
|
+
};
|
|
128
|
+
const baseQuery = Object.assign({ runType: AIRunType.Investigation, status: AIRunStatus.Completed }, (options.projectId ? { projectId: options.projectId } : {}));
|
|
129
|
+
/*
|
|
130
|
+
* "Has a label" is an OR (humanVerdict set OR autoGrade set); the query
|
|
131
|
+
* layer composes ANDs, so run the two legs separately and union by id.
|
|
132
|
+
*/
|
|
133
|
+
const [humanLabeled, autoGraded] = await Promise.all([
|
|
134
|
+
AIRunService.findBy({
|
|
135
|
+
query: Object.assign(Object.assign({}, baseQuery), { humanVerdict: QueryHelper.notNull() }),
|
|
136
|
+
select: runSelect,
|
|
137
|
+
sort: { completedAt: SortOrder.Descending },
|
|
138
|
+
limit,
|
|
139
|
+
skip: 0,
|
|
140
|
+
props: { isRoot: true },
|
|
141
|
+
}),
|
|
142
|
+
AIRunService.findBy({
|
|
143
|
+
query: Object.assign(Object.assign({}, baseQuery), { autoGrade: QueryHelper.notNull() }),
|
|
144
|
+
select: runSelect,
|
|
145
|
+
sort: { completedAt: SortOrder.Descending },
|
|
146
|
+
limit,
|
|
147
|
+
skip: 0,
|
|
148
|
+
props: { isRoot: true },
|
|
149
|
+
}),
|
|
150
|
+
]);
|
|
151
|
+
const runsById = new Map();
|
|
152
|
+
for (const run of [...humanLabeled, ...autoGraded]) {
|
|
153
|
+
if (run.id) {
|
|
154
|
+
runsById.set(run.id.toString(), run);
|
|
155
|
+
}
|
|
156
|
+
}
|
|
157
|
+
const labeledRuns = Array.from(runsById.values())
|
|
158
|
+
.sort((a, b) => {
|
|
159
|
+
return ((b.completedAt ? b.completedAt.getTime() : 0) -
|
|
160
|
+
(a.completedAt ? a.completedAt.getTime() : 0));
|
|
161
|
+
})
|
|
162
|
+
.slice(0, limit);
|
|
163
|
+
const goldenCases = [];
|
|
164
|
+
for (const run of labeledRuns) {
|
|
165
|
+
const subjectType = this.getSubjectType(run);
|
|
166
|
+
if (!subjectType) {
|
|
167
|
+
// An Investigation run always has a subject; be defensive anyway.
|
|
168
|
+
logger.debug(`Eval corpus: run ${(_a = run.id) === null || _a === void 0 ? void 0 : _a.toString()} has no incident/alert subject — skipped.`);
|
|
169
|
+
continue;
|
|
170
|
+
}
|
|
171
|
+
const events = await AIRunEventService.findBy({
|
|
172
|
+
query: { aiRunId: run.id },
|
|
173
|
+
select: {
|
|
174
|
+
sequence: true,
|
|
175
|
+
eventType: true,
|
|
176
|
+
toolName: true,
|
|
177
|
+
citationId: true,
|
|
178
|
+
resultSummary: true,
|
|
179
|
+
},
|
|
180
|
+
sort: { sequence: SortOrder.Ascending },
|
|
181
|
+
limit: LIMIT_MAX,
|
|
182
|
+
skip: 0,
|
|
183
|
+
props: { isRoot: true },
|
|
184
|
+
});
|
|
185
|
+
goldenCases.push({
|
|
186
|
+
runId: run.id.toString(),
|
|
187
|
+
projectId: run.projectId.toString(),
|
|
188
|
+
subjectType,
|
|
189
|
+
completedAt: run.completedAt
|
|
190
|
+
? new Date(run.completedAt).toISOString()
|
|
191
|
+
: undefined,
|
|
192
|
+
label: {
|
|
193
|
+
humanVerdict: run.humanVerdict,
|
|
194
|
+
autoGrade: run.autoGrade,
|
|
195
|
+
},
|
|
196
|
+
confidence: this.deriveConfidenceStats(events),
|
|
197
|
+
analysisPosted: await this.hasPostedAnalysis(run, subjectType),
|
|
198
|
+
events: this.toGoldenCaseEvents(events),
|
|
199
|
+
});
|
|
200
|
+
}
|
|
201
|
+
return goldenCases;
|
|
202
|
+
}
|
|
203
|
+
static getSubjectType(run) {
|
|
204
|
+
if (run.triggeredByIncidentId) {
|
|
205
|
+
return "Incident";
|
|
206
|
+
}
|
|
207
|
+
if (run.triggeredByAlertId) {
|
|
208
|
+
return "Alert";
|
|
209
|
+
}
|
|
210
|
+
return null;
|
|
211
|
+
}
|
|
212
|
+
/*
|
|
213
|
+
* Whether the AI's analysis was posted for the run's subject: a RootCause
|
|
214
|
+
* feed item exists (postAnalysis is its only writer — the sourcing rule
|
|
215
|
+
* shared with InvestigationGrader). Subject-level, not run-level; see the
|
|
216
|
+
* GoldenCase.analysisPosted doc comment.
|
|
217
|
+
*/
|
|
218
|
+
static async hasPostedAnalysis(run, subjectType) {
|
|
219
|
+
if (subjectType === "Incident") {
|
|
220
|
+
const count = (await IncidentFeedService.countBy({
|
|
221
|
+
query: {
|
|
222
|
+
incidentId: run.triggeredByIncidentId,
|
|
223
|
+
incidentFeedEventType: IncidentFeedEventType.RootCause,
|
|
224
|
+
},
|
|
225
|
+
props: { isRoot: true },
|
|
226
|
+
})).toNumber();
|
|
227
|
+
return count > 0;
|
|
228
|
+
}
|
|
229
|
+
const count = (await AlertFeedService.countBy({
|
|
230
|
+
query: {
|
|
231
|
+
alertId: run.triggeredByAlertId,
|
|
232
|
+
alertFeedEventType: AlertFeedEventType.RootCause,
|
|
233
|
+
},
|
|
234
|
+
props: { isRoot: true },
|
|
235
|
+
})).toNumber();
|
|
236
|
+
return count > 0;
|
|
237
|
+
}
|
|
238
|
+
}
|
|
239
|
+
__decorate([
|
|
240
|
+
CaptureSpan(),
|
|
241
|
+
__metadata("design:type", Function),
|
|
242
|
+
__metadata("design:paramtypes", [Object]),
|
|
243
|
+
__metadata("design:returntype", Promise)
|
|
244
|
+
], EvalCorpus, "exportCorpus", null);
|
|
245
|
+
//# sourceMappingURL=EvalCorpus.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"EvalCorpus.js","sourceRoot":"","sources":["../../../../../../Server/Utils/AI/Eval/EvalCorpus.ts"],"names":[],"mappings":";;;;;;;;;AACA,OAAO,SAAS,MAAM,0CAA0C,CAAC;AACjE,OAAO,SAAS,MAAM,qCAAqC,CAAC;AAC5D,OAAO,SAAS,MAAM,gCAAgC,CAAC;AACvD,OAAO,WAAW,MAAM,kCAAkC,CAAC;AAC3D,OAAO,cAAc,MAAM,qCAAqC,CAAC;AAMjE,OAAO,EAAE,qBAAqB,EAAE,MAAM,gDAAgD,CAAC;AACvF,OAAO,EAAE,kBAAkB,EAAE,MAAM,6CAA6C,CAAC;AACjF,OAAO,WAAW,MAAM,qCAAqC,CAAC;AAE9D,OAAO,YAAY,MAAM,gCAAgC,CAAC;AAC1D,OAAO,iBAAiB,MAAM,qCAAqC,CAAC;AACpE,OAAO,mBAAmB,MAAM,uCAAuC,CAAC;AACxE,OAAO,gBAAgB,MAAM,oCAAoC,CAAC;AAClE,OAAO,kBAAkB,MAAM,yBAAyB,CAAC;AACzD,OAAO,MAAM,MAAM,cAAc,CAAC;AAClC,OAAO,WAAW,MAAM,6BAA6B,CAAC;AAEtD;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AAEH,wEAAwE;AACxE,MAAM,CAAC,MAAM,oBAAoB,GAAW,GAAG,CAAC;AAsEhD,MAAM,CAAC,OAAO,OAAO,UAAU;IAC7B;;;;;;;;;;;;OAYG;IACI,MAAM,CAAC,qBAAqB,CACjC,MAA+B;;QAE/B,IAAI,eAAe,GAAW,CAAC,CAAC;QAChC,IAAI,oBAAoB,GAAW,CAAC,CAAC;QACrC,IAAI,cAAc,GAAW,CAAC,CAAC;QAC/B,IAAI,eAAe,GAAW,CAAC,CAAC;QAEhC,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;YAC3B,IAAI,KAAK,CAAC,UAAU,EAAE,CAAC;gBACrB,eAAe,EAAE,CAAC;YACpB,CAAC;YAED,IAAI,KAAK,CAAC,SAAS,KAAK,cAAc,CAAC,iBAAiB,EAAE,CAAC;gBACzD,cAAc,EAAE,CAAC;gBAEjB,IACE,kBAAkB,CAAC,qBAAqB,CACtC,MAAA,KAAK,CAAC,aAAa,0CAAE,QAAQ,CAC9B,EACD,CAAC;oBACD,oBAAoB,EAAE,CAAC;gBACzB,CAAC;YACH,CAAC;iBAAM,IAAI,KAAK,CAAC,SAAS,KAAK,cAAc,CAAC,cAAc,EAAE,CAAC;gBAC7D,cAAc,EAAE,CAAC;gBACjB,eAAe,EAAE,CAAC;YACpB,CAAC;QACH,CAAC;QAED,OAAO;YACL,eAAe;YACf,oBAAoB;YACpB,cAAc;YACd,eAAe;SAChB,CAAC;IACJ,CAAC;IAED,8EAA8E;IACvE,MAAM,CAAC,kBAAkB,CAC9B,MAA+B;QAE/B,OAAO,MAAM,CAAC,GAAG,CAAC,CAAC,KAAuB,EAAE,EAAE;YAC5C,IAAI,EAAE,GAAwB,SAAS,CAAC;YAExC,IAAI,KAAK,CAAC,SAAS,KAAK,cAAc,CAAC,iBAAiB,EAAE,CAAC;gBACzD,EAAE,GAAG,IAAI,CAAC;YACZ,CAAC;iBAAM,IAAI,KAAK,CAAC,SAAS,KAAK,cAAc,CAAC,cAAc,EAAE,CAAC;gBAC7D,EAAE,GAAG,KAAK,CAAC;YACb,CAAC;YAED,OAAO;gBACL,QAAQ,EAAE,KAAK,CAAC,QAAQ,IAAI,CAAC;gBAC7B,SAAS,EAAE,CAAC,KAAK,CAAC,SAAS,IAAI,EAAE,CAAC,CAAC,QAAQ,EAAE;gBAC7C,QAAQ,EAAE,KAAK,CAAC,QAAQ;gBACxB,EAAE;aACH,CAAC;QACJ,CAAC,CAAC,CAAC;IACL,CAAC;IAED;;;;OAIG;IAEiB,AAAb,MAAM,CAAC,KAAK,CAAC,YAAY,CAC9B,OAA4B;;QAE5B,MAAM,KAAK,GAAW,IAAI,CAAC,GAAG,CAC5B,OAAO,CAAC,KAAK,IAAI,oBAAoB,EACrC,SAAS,CACV,CAAC;QAEF,MAAM,SAAS,GAQX;YACF,GAAG,EAAE,IAAI;YACT,SAAS,EAAE,IAAI;YACf,WAAW,EAAE,IAAI;YACjB,YAAY,EAAE,IAAI;YAClB,SAAS,EAAE,IAAI;YACf,qBAAqB,EAAE,IAAI;YAC3B,kBAAkB,EAAE,IAAI;SACzB,CAAC;QAEF,MAAM,SAAS,mBACb,OAAO,EAAE,SAAS,CAAC,aAAa,EAChC,MAAM,EAAE,WAAW,CAAC,SAAS,IAC1B,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,OAAO,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAC/D,CAAC;QAEF;;;WAGG;QACH,MAAM,CAAC,YAAY,EAAE,UAAU,CAAC,GAC9B,MAAM,OAAO,CAAC,GAAG,CAAC;YAChB,YAAY,CAAC,MAAM,CAAC;gBAClB,KAAK,kCAAO,SAAS,KAAE,YAAY,EAAE,WAAW,CAAC,OAAO,EAAE,GAAE;gBAC5D,MAAM,EAAE,SAAS;gBACjB,IAAI,EAAE,EAAE,WAAW,EAAE,SAAS,CAAC,UAAU,EAAE;gBAC3C,KAAK;gBACL,IAAI,EAAE,CAAC;gBACP,KAAK,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE;aACxB,CAAC;YACF,YAAY,CAAC,MAAM,CAAC;gBAClB,KAAK,kCAAO,SAAS,KAAE,SAAS,EAAE,WAAW,CAAC,OAAO,EAAE,GAAE;gBACzD,MAAM,EAAE,SAAS;gBACjB,IAAI,EAAE,EAAE,WAAW,EAAE,SAAS,CAAC,UAAU,EAAE;gBAC3C,KAAK;gBACL,IAAI,EAAE,CAAC;gBACP,KAAK,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE;aACxB,CAAC;SACH,CAAC,CAAC;QAEL,MAAM,QAAQ,GAAuB,IAAI,GAAG,EAAiB,CAAC;QAE9D,KAAK,MAAM,GAAG,IAAI,CAAC,GAAG,YAAY,EAAE,GAAG,UAAU,CAAC,EAAE,CAAC;YACnD,IAAI,GAAG,CAAC,EAAE,EAAE,CAAC;gBACX,QAAQ,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,QAAQ,EAAE,EAAE,GAAG,CAAC,CAAC;YACvC,CAAC;QACH,CAAC;QAED,MAAM,WAAW,GAAiB,KAAK,CAAC,IAAI,CAAC,QAAQ,CAAC,MAAM,EAAE,CAAC;aAC5D,IAAI,CAAC,CAAC,CAAQ,EAAE,CAAQ,EAAE,EAAE;YAC3B,OAAO,CACL,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;gBAC7C,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAC9C,CAAC;QACJ,CAAC,CAAC;aACD,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC;QAEnB,MAAM,WAAW,GAAsB,EAAE,CAAC;QAE1C,KAAK,MAAM,GAAG,IAAI,WAAW,EAAE,CAAC;YAC9B,MAAM,WAAW,GACf,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,CAAC;YAE3B,IAAI,CAAC,WAAW,EAAE,CAAC;gBACjB,kEAAkE;gBAClE,MAAM,CAAC,KAAK,CACV,oBAAoB,MAAA,GAAG,CAAC,EAAE,0CAAE,QAAQ,EAAE,2CAA2C,CAClF,CAAC;gBACF,SAAS;YACX,CAAC;YAED,MAAM,MAAM,GAAsB,MAAM,iBAAiB,CAAC,MAAM,CAAC;gBAC/D,KAAK,EAAE,EAAE,OAAO,EAAE,GAAG,CAAC,EAAG,EAAE;gBAC3B,MAAM,EAAE;oBACN,QAAQ,EAAE,IAAI;oBACd,SAAS,EAAE,IAAI;oBACf,QAAQ,EAAE,IAAI;oBACd,UAAU,EAAE,IAAI;oBAChB,aAAa,EAAE,IAAI;iBACpB;gBACD,IAAI,EAAE,EAAE,QAAQ,EAAE,SAAS,CAAC,SAAS,EAAE;gBACvC,KAAK,EAAE,SAAS;gBAChB,IAAI,EAAE,CAAC;gBACP,KAAK,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE;aACxB,CAAC,CAAC;YAEH,WAAW,CAAC,IAAI,CAAC;gBACf,KAAK,EAAE,GAAG,CAAC,EAAG,CAAC,QAAQ,EAAE;gBACzB,SAAS,EAAE,GAAG,CAAC,SAAU,CAAC,QAAQ,EAAE;gBACpC,WAAW;gBACX,WAAW,EAAE,GAAG,CAAC,WAAW;oBAC1B,CAAC,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC,WAAW,EAAE;oBACzC,CAAC,CAAC,SAAS;gBACb,KAAK,EAAE;oBACL,YAAY,EAAE,GAAG,CAAC,YAAY;oBAC9B,SAAS,EAAE,GAAG,CAAC,SAAS;iBACzB;gBACD,UAAU,EAAE,IAAI,CAAC,qBAAqB,CAAC,MAAM,CAAC;gBAC9C,cAAc,EAAE,MAAM,IAAI,CAAC,iBAAiB,CAAC,GAAG,EAAE,WAAW,CAAC;gBAC9D,MAAM,EAAE,IAAI,CAAC,kBAAkB,CAAC,MAAM,CAAC;aACxC,CAAC,CAAC;QACL,CAAC;QAED,OAAO,WAAW,CAAC;IACrB,CAAC;IAEO,MAAM,CAAC,cAAc,CAAC,GAAU;QACtC,IAAI,GAAG,CAAC,qBAAqB,EAAE,CAAC;YAC9B,OAAO,UAAU,CAAC;QACpB,CAAC;QAED,IAAI,GAAG,CAAC,kBAAkB,EAAE,CAAC;YAC3B,OAAO,OAAO,CAAC;QACjB,CAAC;QAED,OAAO,IAAI,CAAC;IACd,CAAC;IAED;;;;;OAKG;IACK,MAAM,CAAC,KAAK,CAAC,iBAAiB,CACpC,GAAU,EACV,WAAkC;QAElC,IAAI,WAAW,KAAK,UAAU,EAAE,CAAC;YAC/B,MAAM,KAAK,GAAW,CACpB,MAAM,mBAAmB,CAAC,OAAO,CAAC;gBAChC,KAAK,EAAE;oBACL,UAAU,EAAE,GAAG,CAAC,qBAAsB;oBACtC,qBAAqB,EAAE,qBAAqB,CAAC,SAAS;iBACvD;gBACD,KAAK,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE;aACxB,CAAC,CACH,CAAC,QAAQ,EAAE,CAAC;YAEb,OAAO,KAAK,GAAG,CAAC,CAAC;QACnB,CAAC;QAED,MAAM,KAAK,GAAW,CACpB,MAAM,gBAAgB,CAAC,OAAO,CAAC;YAC7B,KAAK,EAAE;gBACL,OAAO,EAAE,GAAG,CAAC,kBAAmB;gBAChC,kBAAkB,EAAE,kBAAkB,CAAC,SAAS;aACjD;YACD,KAAK,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE;SACxB,CAAC,CACH,CAAC,QAAQ,EAAE,CAAC;QAEb,OAAO,KAAK,GAAG,CAAC,CAAC;IACnB,CAAC;CACF;AA1KqB;IADnB,WAAW,EAAE;;;;oCAyHb"}
|
|
@@ -0,0 +1,83 @@
|
|
|
1
|
+
import AIRunHumanVerdict from "../../../../Types/AI/AIRunHumanVerdict";
|
|
2
|
+
import AIRunAutoGrade from "../../../../Types/AI/AIRunAutoGrade";
|
|
3
|
+
import AIConfidenceSignal from "../SRE/ConfidenceSignal";
|
|
4
|
+
/*
|
|
5
|
+
* Below this many labeled cases the numbers are noise, not measurement. The
|
|
6
|
+
* runner must surface this loudly; nothing downstream (G3 graduation, public
|
|
7
|
+
* scorecards) may consume a report whose sampleTooSmall is true.
|
|
8
|
+
*/
|
|
9
|
+
export const MINIMUM_TRUSTWORTHY_CORPUS_SIZE = 20;
|
|
10
|
+
export default class EvalScores {
|
|
11
|
+
// Build a score with the null-when-unmeasurable rule applied.
|
|
12
|
+
static makeScore(numerator, denominator) {
|
|
13
|
+
return {
|
|
14
|
+
value: denominator === 0 ? null : numerator / denominator,
|
|
15
|
+
numerator,
|
|
16
|
+
denominator,
|
|
17
|
+
};
|
|
18
|
+
}
|
|
19
|
+
/*
|
|
20
|
+
* Would the deterministic evidence floor have flagged this run inconclusive?
|
|
21
|
+
* Reuses AIConfidenceSignal.hasVerifiableEvidence — the LIVE floor —
|
|
22
|
+
* over the corpus stats, so the offline answer can never drift from what
|
|
23
|
+
* production would actually have decided: zero citations minted AND zero
|
|
24
|
+
* data-bearing tool calls.
|
|
25
|
+
*/
|
|
26
|
+
static wouldDeterministicFloorFlagInconclusive(confidence) {
|
|
27
|
+
return !AIConfidenceSignal.hasVerifiableEvidence({
|
|
28
|
+
citationCount: confidence.citationsMinted,
|
|
29
|
+
dataBearingToolCallCount: confidence.dataBearingToolCalls,
|
|
30
|
+
anyToolReturnedData: confidence.dataBearingToolCalls > 0,
|
|
31
|
+
});
|
|
32
|
+
}
|
|
33
|
+
static computeScores(cases) {
|
|
34
|
+
let gradedCount = 0;
|
|
35
|
+
let gradedMatchCount = 0;
|
|
36
|
+
let humanLabeledCount = 0;
|
|
37
|
+
let humanConfirmedCount = 0;
|
|
38
|
+
let analysisPostedCount = 0;
|
|
39
|
+
let analysisPostedWithCitationCount = 0;
|
|
40
|
+
let toolCallsTotal = 0;
|
|
41
|
+
let dataBearingToolCalls = 0;
|
|
42
|
+
let humanRejectedCount = 0;
|
|
43
|
+
let humanRejectedFloorFlaggedCount = 0;
|
|
44
|
+
for (const goldenCase of cases) {
|
|
45
|
+
if (goldenCase.label.autoGrade) {
|
|
46
|
+
gradedCount++;
|
|
47
|
+
if (goldenCase.label.autoGrade === AIRunAutoGrade.Match) {
|
|
48
|
+
gradedMatchCount++;
|
|
49
|
+
}
|
|
50
|
+
}
|
|
51
|
+
if (goldenCase.label.humanVerdict) {
|
|
52
|
+
humanLabeledCount++;
|
|
53
|
+
if (goldenCase.label.humanVerdict === AIRunHumanVerdict.Confirmed) {
|
|
54
|
+
humanConfirmedCount++;
|
|
55
|
+
}
|
|
56
|
+
if (goldenCase.label.humanVerdict === AIRunHumanVerdict.Rejected) {
|
|
57
|
+
humanRejectedCount++;
|
|
58
|
+
if (this.wouldDeterministicFloorFlagInconclusive(goldenCase.confidence)) {
|
|
59
|
+
humanRejectedFloorFlaggedCount++;
|
|
60
|
+
}
|
|
61
|
+
}
|
|
62
|
+
}
|
|
63
|
+
if (goldenCase.analysisPosted) {
|
|
64
|
+
analysisPostedCount++;
|
|
65
|
+
if (goldenCase.confidence.citationsMinted > 0) {
|
|
66
|
+
analysisPostedWithCitationCount++;
|
|
67
|
+
}
|
|
68
|
+
}
|
|
69
|
+
toolCallsTotal += goldenCase.confidence.toolCallsTotal;
|
|
70
|
+
dataBearingToolCalls += goldenCase.confidence.dataBearingToolCalls;
|
|
71
|
+
}
|
|
72
|
+
return {
|
|
73
|
+
caseCount: cases.length,
|
|
74
|
+
sampleTooSmall: cases.length < MINIMUM_TRUSTWORTHY_CORPUS_SIZE,
|
|
75
|
+
topHypothesisPrecision: this.makeScore(gradedMatchCount, gradedCount),
|
|
76
|
+
humanConfirmedRate: this.makeScore(humanConfirmedCount, humanLabeledCount),
|
|
77
|
+
citationGroundingRate: this.makeScore(analysisPostedWithCitationCount, analysisPostedCount),
|
|
78
|
+
toolSelectionAccuracy: this.makeScore(dataBearingToolCalls, toolCallsTotal),
|
|
79
|
+
inconclusiveRecall: this.makeScore(humanRejectedFloorFlaggedCount, humanRejectedCount),
|
|
80
|
+
};
|
|
81
|
+
}
|
|
82
|
+
}
|
|
83
|
+
//# sourceMappingURL=EvalScores.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"EvalScores.js","sourceRoot":"","sources":["../../../../../../Server/Utils/AI/Eval/EvalScores.ts"],"names":[],"mappings":"AAAA,OAAO,iBAAiB,MAAM,wCAAwC,CAAC;AACvE,OAAO,cAAc,MAAM,qCAAqC,CAAC;AACjE,OAAO,kBAAkB,MAAM,yBAAyB,CAAC;AA0BzD;;;;GAIG;AACH,MAAM,CAAC,MAAM,+BAA+B,GAAW,EAAE,CAAC;AAkC1D,MAAM,CAAC,OAAO,OAAO,UAAU;IAC7B,8DAA8D;IACvD,MAAM,CAAC,SAAS,CAAC,SAAiB,EAAE,WAAmB;QAC5D,OAAO;YACL,KAAK,EAAE,WAAW,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,SAAS,GAAG,WAAW;YACzD,SAAS;YACT,WAAW;SACZ,CAAC;IACJ,CAAC;IAED;;;;;;OAMG;IACI,MAAM,CAAC,uCAAuC,CACnD,UAAqC;QAErC,OAAO,CAAC,kBAAkB,CAAC,qBAAqB,CAAC;YAC/C,aAAa,EAAE,UAAU,CAAC,eAAe;YACzC,wBAAwB,EAAE,UAAU,CAAC,oBAAoB;YACzD,mBAAmB,EAAE,UAAU,CAAC,oBAAoB,GAAG,CAAC;SACzD,CAAC,CAAC;IACL,CAAC;IAEM,MAAM,CAAC,aAAa,CAAC,KAAwB;QAClD,IAAI,WAAW,GAAW,CAAC,CAAC;QAC5B,IAAI,gBAAgB,GAAW,CAAC,CAAC;QAEjC,IAAI,iBAAiB,GAAW,CAAC,CAAC;QAClC,IAAI,mBAAmB,GAAW,CAAC,CAAC;QAEpC,IAAI,mBAAmB,GAAW,CAAC,CAAC;QACpC,IAAI,+BAA+B,GAAW,CAAC,CAAC;QAEhD,IAAI,cAAc,GAAW,CAAC,CAAC;QAC/B,IAAI,oBAAoB,GAAW,CAAC,CAAC;QAErC,IAAI,kBAAkB,GAAW,CAAC,CAAC;QACnC,IAAI,8BAA8B,GAAW,CAAC,CAAC;QAE/C,KAAK,MAAM,UAAU,IAAI,KAAK,EAAE,CAAC;YAC/B,IAAI,UAAU,CAAC,KAAK,CAAC,SAAS,EAAE,CAAC;gBAC/B,WAAW,EAAE,CAAC;gBAEd,IAAI,UAAU,CAAC,KAAK,CAAC,SAAS,KAAK,cAAc,CAAC,KAAK,EAAE,CAAC;oBACxD,gBAAgB,EAAE,CAAC;gBACrB,CAAC;YACH,CAAC;YAED,IAAI,UAAU,CAAC,KAAK,CAAC,YAAY,EAAE,CAAC;gBAClC,iBAAiB,EAAE,CAAC;gBAEpB,IAAI,UAAU,CAAC,KAAK,CAAC,YAAY,KAAK,iBAAiB,CAAC,SAAS,EAAE,CAAC;oBAClE,mBAAmB,EAAE,CAAC;gBACxB,CAAC;gBAED,IAAI,UAAU,CAAC,KAAK,CAAC,YAAY,KAAK,iBAAiB,CAAC,QAAQ,EAAE,CAAC;oBACjE,kBAAkB,EAAE,CAAC;oBAErB,IACE,IAAI,CAAC,uCAAuC,CAAC,UAAU,CAAC,UAAU,CAAC,EACnE,CAAC;wBACD,8BAA8B,EAAE,CAAC;oBACnC,CAAC;gBACH,CAAC;YACH,CAAC;YAED,IAAI,UAAU,CAAC,cAAc,EAAE,CAAC;gBAC9B,mBAAmB,EAAE,CAAC;gBAEtB,IAAI,UAAU,CAAC,UAAU,CAAC,eAAe,GAAG,CAAC,EAAE,CAAC;oBAC9C,+BAA+B,EAAE,CAAC;gBACpC,CAAC;YACH,CAAC;YAED,cAAc,IAAI,UAAU,CAAC,UAAU,CAAC,cAAc,CAAC;YACvD,oBAAoB,IAAI,UAAU,CAAC,UAAU,CAAC,oBAAoB,CAAC;QACrE,CAAC;QAED,OAAO;YACL,SAAS,EAAE,KAAK,CAAC,MAAM;YACvB,cAAc,EAAE,KAAK,CAAC,MAAM,GAAG,+BAA+B;YAC9D,sBAAsB,EAAE,IAAI,CAAC,SAAS,CAAC,gBAAgB,EAAE,WAAW,CAAC;YACrE,kBAAkB,EAAE,IAAI,CAAC,SAAS,CAChC,mBAAmB,EACnB,iBAAiB,CAClB;YACD,qBAAqB,EAAE,IAAI,CAAC,SAAS,CACnC,+BAA+B,EAC/B,mBAAmB,CACpB;YACD,qBAAqB,EAAE,IAAI,CAAC,SAAS,CACnC,oBAAoB,EACpB,cAAc,CACf;YACD,kBAAkB,EAAE,IAAI,CAAC,SAAS,CAChC,8BAA8B,EAC9B,kBAAkB,CACnB;SACF,CAAC;IACJ,CAAC;CACF"}
|
|
@@ -0,0 +1,12 @@
|
|
|
1
|
+
import NotImplementedException from "../../../../Types/Exception/NotImplementedException";
|
|
2
|
+
export default class ReplayInvestigation {
|
|
3
|
+
/*
|
|
4
|
+
* NOT IMPLEMENTED — see the module doc block for exactly what is missing
|
|
5
|
+
* (recorded tool RESULT payloads; today's AIRunEvent trail carries argument
|
|
6
|
+
* + summary only).
|
|
7
|
+
*/
|
|
8
|
+
static async replay(_goldenCase) {
|
|
9
|
+
throw new NotImplementedException();
|
|
10
|
+
}
|
|
11
|
+
}
|
|
12
|
+
//# sourceMappingURL=ReplayInvestigation.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"ReplayInvestigation.js","sourceRoot":"","sources":["../../../../../../Server/Utils/AI/Eval/ReplayInvestigation.ts"],"names":[],"mappings":"AAAA,OAAO,uBAAuB,MAAM,qDAAqD,CAAC;AA6E1F,MAAM,CAAC,OAAO,OAAO,mBAAmB;IACtC;;;;OAIG;IACI,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,WAAuB;QAChD,MAAM,IAAI,uBAAuB,EAAE,CAAC;IACtC,CAAC;CACF"}
|