@voltagent/core 1.1.26 → 1.1.27

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/index.d.mts CHANGED
@@ -1,7 +1,7 @@
1
1
  import { ModelMessage, DataContent as DataContent$1, UserContent, AssistantContent, ToolContent, ProviderOptions as ProviderOptions$1 } from '@ai-sdk/provider-utils';
2
2
  export { AssistantContent, FilePart, ImagePart, TextPart, ToolContent, UserContent } from '@ai-sdk/provider-utils';
3
3
  import { TextStreamPart, generateText, UIMessage, StreamTextResult, LanguageModel, CallSettings, Output, ToolSet, GenerateTextResult, GenerateObjectResult, AsyncIterableStream as AsyncIterableStream$1, CallWarning, LanguageModelUsage, FinishReason, EmbeddingModel } from 'ai';
4
- export { hasToolCall, stepCountIs } from 'ai';
4
+ export { LanguageModel, hasToolCall, stepCountIs } from 'ai';
5
5
  import * as zod from 'zod';
6
6
  import { z } from 'zod';
7
7
  import { AsyncIterableStream } from '@voltagent/internal/utils';
@@ -639,13 +639,7 @@ type GetMessagesOptions = {
639
639
  /**
640
640
  * Memory options for MemoryManager
641
641
  */
642
- type MemoryOptions = {
643
- /**
644
- * Maximum number of messages to store in the database
645
- * @default 100
646
- */
647
- storageLimit?: number;
648
- };
642
+ type MemoryOptions = {};
649
643
  /**
650
644
  * Workflow state entry for suspension and resumption
651
645
  * Stores only the essential state needed to resume a workflow
@@ -748,11 +742,6 @@ interface MemoryConfig {
748
742
  * @default 3600000 (1 hour)
749
743
  */
750
744
  cacheTTL?: number;
751
- /**
752
- * Maximum number of messages to store per conversation
753
- * @default 100
754
- */
755
- storageLimit?: number;
756
745
  /**
757
746
  * Working memory configuration
758
747
  * Enables agents to maintain important context
@@ -1764,6 +1753,114 @@ interface VoltOpsPromptManager {
1764
1753
  entries: string[];
1765
1754
  };
1766
1755
  }
1756
+ type VoltOpsEvalRunStatus = "pending" | "running" | "succeeded" | "failed" | "cancelled";
1757
+ type VoltOpsTerminalEvalRunStatus = "succeeded" | "failed" | "cancelled";
1758
+ type VoltOpsEvalResultStatus = "pending" | "running" | "passed" | "failed" | "error";
1759
+ interface VoltOpsEvalRunSummary {
1760
+ id: string;
1761
+ status: VoltOpsEvalRunStatus | string;
1762
+ triggerSource: string;
1763
+ datasetId?: string | null;
1764
+ datasetVersionId?: string | null;
1765
+ datasetVersionLabel?: string | null;
1766
+ itemCount: number;
1767
+ successCount: number;
1768
+ failureCount: number;
1769
+ meanScore?: number | null;
1770
+ medianScore?: number | null;
1771
+ sumScore?: number | null;
1772
+ passRate?: number | null;
1773
+ startedAt?: string | null;
1774
+ completedAt?: string | null;
1775
+ durationMs?: number | null;
1776
+ tags?: string[] | null;
1777
+ createdAt: string;
1778
+ updatedAt: string;
1779
+ }
1780
+ interface VoltOpsCreateEvalRunRequest {
1781
+ experimentId?: string;
1782
+ datasetVersionId?: string;
1783
+ providerCredentialId?: string;
1784
+ triggerSource?: string;
1785
+ autoQueue?: boolean;
1786
+ }
1787
+ interface VoltOpsEvalRunResultScorePayload {
1788
+ scorerId: string;
1789
+ score?: number | null;
1790
+ threshold?: number | null;
1791
+ thresholdPassed?: boolean | null;
1792
+ metadata?: Record<string, unknown> | null;
1793
+ }
1794
+ interface VoltOpsEvalRunResultLiveMetadata {
1795
+ traceId?: string | null;
1796
+ spanId?: string | null;
1797
+ operationId?: string | null;
1798
+ operationType?: string | null;
1799
+ sampling?: {
1800
+ strategy: string;
1801
+ rate?: number | null;
1802
+ } | null;
1803
+ triggerSource?: string | null;
1804
+ environment?: string | null;
1805
+ }
1806
+ interface VoltOpsAppendEvalRunResultPayload {
1807
+ id?: string;
1808
+ datasetItemId?: string | null;
1809
+ datasetItemHash: string;
1810
+ status?: VoltOpsEvalResultStatus;
1811
+ input?: unknown;
1812
+ expected?: unknown;
1813
+ output?: unknown;
1814
+ durationMs?: number | null;
1815
+ scores?: VoltOpsEvalRunResultScorePayload[];
1816
+ metadata?: Record<string, unknown> | null;
1817
+ traceIds?: string[] | null;
1818
+ liveEval?: VoltOpsEvalRunResultLiveMetadata | null;
1819
+ }
1820
+ interface VoltOpsAppendEvalRunResultsRequest {
1821
+ results: VoltOpsAppendEvalRunResultPayload[];
1822
+ }
1823
+ interface VoltOpsEvalRunCompletionSummaryPayload {
1824
+ itemCount?: number;
1825
+ successCount?: number;
1826
+ failureCount?: number;
1827
+ meanScore?: number | null;
1828
+ medianScore?: number | null;
1829
+ sumScore?: number | null;
1830
+ passRate?: number | null;
1831
+ durationMs?: number | null;
1832
+ metadata?: Record<string, unknown> | null;
1833
+ }
1834
+ interface VoltOpsEvalRunErrorPayload {
1835
+ message: string;
1836
+ code?: string;
1837
+ details?: Record<string, unknown>;
1838
+ }
1839
+ interface VoltOpsCompleteEvalRunRequest {
1840
+ status: VoltOpsTerminalEvalRunStatus;
1841
+ summary?: VoltOpsEvalRunCompletionSummaryPayload;
1842
+ error?: VoltOpsEvalRunErrorPayload;
1843
+ }
1844
+ interface VoltOpsCreateScorerRequest {
1845
+ id: string;
1846
+ name: string;
1847
+ category?: string | null;
1848
+ description?: string | null;
1849
+ defaultThreshold?: number | null;
1850
+ thresholdOperator?: string | null;
1851
+ metadata?: Record<string, unknown> | null;
1852
+ }
1853
+ interface VoltOpsScorerSummary {
1854
+ id: string;
1855
+ name: string;
1856
+ category?: string | null;
1857
+ description?: string | null;
1858
+ defaultThreshold?: number | null;
1859
+ thresholdOperator?: string | null;
1860
+ metadata?: Record<string, unknown> | null;
1861
+ createdAt: string;
1862
+ updatedAt: string;
1863
+ }
1767
1864
  /**
1768
1865
  * Main VoltOps client interface
1769
1866
  */
@@ -1776,6 +1873,14 @@ interface VoltOpsClient$1 {
1776
1873
  };
1777
1874
  /** Create a prompt helper for agent instructions */
1778
1875
  createPromptHelper(agentId: string, historyEntryId?: string): PromptHelper;
1876
+ /** Create a new evaluation run in VoltOps */
1877
+ createEvalRun(payload?: VoltOpsCreateEvalRunRequest): Promise<VoltOpsEvalRunSummary>;
1878
+ /** Append evaluation results to an existing run */
1879
+ appendEvalRunResults(runId: string, payload: VoltOpsAppendEvalRunResultsRequest): Promise<VoltOpsEvalRunSummary>;
1880
+ /** Complete an evaluation run */
1881
+ completeEvalRun(runId: string, payload: VoltOpsCompleteEvalRunRequest): Promise<VoltOpsEvalRunSummary>;
1882
+ /** Upsert a scorer definition */
1883
+ createEvalScorer(payload: VoltOpsCreateScorerRequest): Promise<VoltOpsScorerSummary>;
1779
1884
  /** List managed memory databases available to the project */
1780
1885
  listManagedMemoryDatabases(): Promise<ManagedMemoryDatabaseSummary[]>;
1781
1886
  /** List credentials for a managed memory database */
@@ -2013,6 +2118,10 @@ declare class VoltOpsClient implements VoltOpsClient$1 {
2013
2118
  * Get prompt manager for direct access
2014
2119
  */
2015
2120
  getPromptManager(): VoltOpsPromptManager | undefined;
2121
+ createEvalRun(payload?: VoltOpsCreateEvalRunRequest): Promise<VoltOpsEvalRunSummary>;
2122
+ appendEvalRunResults(runId: string, payload: VoltOpsAppendEvalRunResultsRequest): Promise<VoltOpsEvalRunSummary>;
2123
+ completeEvalRun(runId: string, payload: VoltOpsCompleteEvalRunRequest): Promise<VoltOpsEvalRunSummary>;
2124
+ createEvalScorer(payload: VoltOpsCreateScorerRequest): Promise<VoltOpsScorerSummary>;
2016
2125
  private request;
2017
2126
  private buildQueryString;
2018
2127
  private createManagedMemoryClient;
@@ -2054,6 +2163,9 @@ declare class VoltOpsClient implements VoltOpsClient$1 {
2054
2163
  * Cleanup resources when client is no longer needed
2055
2164
  */
2056
2165
  dispose(): Promise<void>;
2166
+ private normalizeRunSummary;
2167
+ private normalizeDate;
2168
+ private normalizeScorerSummary;
2057
2169
  }
2058
2170
  /**
2059
2171
  * Factory function to create VoltOps client
@@ -2251,6 +2363,203 @@ interface VoltAgentStreamTextResult<TOOLS extends Record<string, any> = Record<s
2251
2363
  readonly fullStream: AsyncIterable<VoltAgentTextStreamPart<TOOLS>>;
2252
2364
  }
2253
2365
 
2366
+ type SamplingPolicy = {
2367
+ type: "always";
2368
+ } | {
2369
+ type: "never";
2370
+ } | {
2371
+ type: "ratio";
2372
+ rate: number;
2373
+ };
2374
+ interface SamplingMetadata {
2375
+ strategy: "always" | "never" | "ratio";
2376
+ rate?: number;
2377
+ applied?: boolean;
2378
+ }
2379
+ interface ScorerContext<Payload extends Record<string, unknown>, Params extends Record<string, unknown> = Record<string, unknown>> {
2380
+ payload: Payload;
2381
+ params: Params;
2382
+ }
2383
+ type ScorerResult = {
2384
+ status?: "success";
2385
+ score?: number | null;
2386
+ metadata?: Record<string, unknown> | null;
2387
+ } | {
2388
+ status: "error";
2389
+ score?: number | null;
2390
+ metadata?: Record<string, unknown> | null;
2391
+ error: unknown;
2392
+ } | {
2393
+ status: "skipped";
2394
+ score?: number | null;
2395
+ metadata?: Record<string, unknown> | null;
2396
+ };
2397
+ interface LocalScorerDefinition<Payload extends Record<string, unknown>, Params extends Record<string, unknown> = Record<string, unknown>> {
2398
+ id: string;
2399
+ name: string;
2400
+ scorer: (context: ScorerContext<Payload, Params>) => ScorerResult | Promise<ScorerResult>;
2401
+ params?: Params | ((payload: Payload) => Params | undefined | Promise<Params | undefined>);
2402
+ metadata?: Record<string, unknown> | null;
2403
+ sampling?: SamplingPolicy;
2404
+ }
2405
+ interface LocalScorerExecutionResult {
2406
+ id: string;
2407
+ name: string;
2408
+ status: "success" | "error" | "skipped";
2409
+ score: number | null;
2410
+ metadata: Record<string, unknown> | null;
2411
+ sampling?: SamplingMetadata;
2412
+ durationMs: number;
2413
+ error?: unknown;
2414
+ }
2415
+ interface ScorerLifecycleScope {
2416
+ run<T>(executor: () => T | Promise<T>): Promise<T>;
2417
+ }
2418
+ interface RunLocalScorersArgs<Payload extends Record<string, unknown>> {
2419
+ payload: Payload;
2420
+ scorers: LocalScorerDefinition<Payload>[];
2421
+ defaultSampling?: SamplingPolicy;
2422
+ baseArgs?: Record<string, unknown> | ((payload: Payload) => Record<string, unknown> | Promise<Record<string, unknown>>);
2423
+ onScorerStart?: (info: {
2424
+ definition: LocalScorerDefinition<Payload>;
2425
+ sampling?: SamplingMetadata;
2426
+ }) => ScorerLifecycleScope | undefined;
2427
+ onScorerComplete?: (info: {
2428
+ definition: LocalScorerDefinition<Payload>;
2429
+ execution: LocalScorerExecutionResult;
2430
+ context?: ScorerLifecycleScope;
2431
+ }) => void;
2432
+ }
2433
+ interface RunLocalScorersResult {
2434
+ results: LocalScorerExecutionResult[];
2435
+ summary: {
2436
+ successCount: number;
2437
+ errorCount: number;
2438
+ skippedCount: number;
2439
+ };
2440
+ }
2441
+ interface NormalizedScorerResult {
2442
+ score?: number | null;
2443
+ metadata?: Record<string, unknown> | null;
2444
+ error?: unknown;
2445
+ status?: "success" | "error" | "skipped";
2446
+ }
2447
+ declare function runLocalScorers<Payload extends Record<string, unknown>>(args: RunLocalScorersArgs<Payload>): Promise<RunLocalScorersResult>;
2448
+ declare function shouldSample(policy?: SamplingPolicy): boolean;
2449
+ declare function buildSamplingMetadata(policy?: SamplingPolicy): SamplingMetadata | undefined;
2450
+ declare function normalizeScorerResult(result: unknown): NormalizedScorerResult;
2451
+
2452
+ interface ScorerPipelineContext<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> {
2453
+ payload: Payload;
2454
+ params: Params;
2455
+ results: Record<string, unknown>;
2456
+ }
2457
+ interface ScorerReasonContext<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> extends ScorerPipelineContext<Payload, Params> {
2458
+ score: number | null;
2459
+ }
2460
+ type PreprocessFunctionStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = (context: ScorerPipelineContext<Payload, Params>) => unknown | Promise<unknown>;
2461
+ type AnalyzeFunctionStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = (context: ScorerPipelineContext<Payload, Params>) => unknown | Promise<unknown>;
2462
+ type GenerateScoreResult = number | {
2463
+ score: number;
2464
+ metadata?: Record<string, unknown> | null;
2465
+ };
2466
+ type PreprocessStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = PreprocessFunctionStep<Payload, Params>;
2467
+ type AnalyzeStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = AnalyzeFunctionStep<Payload, Params>;
2468
+ type GenerateScoreStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = (context: ScorerPipelineContext<Payload, Params>) => GenerateScoreResult | Promise<GenerateScoreResult>;
2469
+ type GenerateReasonResult = string | {
2470
+ reason: string;
2471
+ metadata?: Record<string, unknown> | null;
2472
+ };
2473
+ type GenerateReasonStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = (context: ScorerReasonContext<Payload, Params>) => GenerateReasonResult | Promise<GenerateReasonResult>;
2474
+ interface CreateScorerOptions<Payload extends Record<string, unknown> = Record<string, unknown>, Params extends Record<string, unknown> = Record<string, unknown>> {
2475
+ id: string;
2476
+ name?: string;
2477
+ metadata?: Record<string, unknown> | null;
2478
+ preprocess?: PreprocessStep<Payload, Params>;
2479
+ analyze?: AnalyzeStep<Payload, Params>;
2480
+ generateScore?: GenerateScoreStep<Payload, Params>;
2481
+ generateReason?: GenerateReasonStep<Payload, Params>;
2482
+ }
2483
+ declare function createScorer<Payload extends Record<string, unknown> = Record<string, unknown>, Params extends Record<string, unknown> = Record<string, unknown>>(options: CreateScorerOptions<Payload, Params>): LocalScorerDefinition<Payload, Params>;
2484
+ interface WeightedBlendComponent<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> {
2485
+ id: string;
2486
+ weight: number;
2487
+ step?: GenerateScoreStep<Payload, Params>;
2488
+ }
2489
+ interface WeightedBlendOptions {
2490
+ metadataKey?: string;
2491
+ }
2492
+ declare function weightedBlend<Payload extends Record<string, unknown>, Params extends Record<string, unknown>>(components: WeightedBlendComponent<Payload, Params>[], options?: WeightedBlendOptions): GenerateScoreStep<Payload, Params>;
2493
+
2494
+ interface BuilderResultsSnapshot {
2495
+ prepare?: unknown;
2496
+ analyze?: unknown;
2497
+ score?: number | null;
2498
+ reason?: string | null;
2499
+ raw: Record<string, unknown>;
2500
+ }
2501
+ interface BuilderContextBase<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> {
2502
+ payload: Payload;
2503
+ params: Params;
2504
+ results: BuilderResultsSnapshot;
2505
+ }
2506
+ interface BuilderPrepareContext<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> extends BuilderContextBase<Payload, Params> {
2507
+ kind: "prepare";
2508
+ }
2509
+ interface BuilderAnalyzeContext<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> extends BuilderContextBase<Payload, Params> {
2510
+ kind: "analyze";
2511
+ }
2512
+ interface BuilderScoreContext<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> extends BuilderContextBase<Payload, Params> {
2513
+ kind: "score";
2514
+ }
2515
+ interface BuilderReasonContext<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> extends BuilderContextBase<Payload, Params> {
2516
+ kind: "reason";
2517
+ score: number | null;
2518
+ }
2519
+ type BuilderPrepareStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = (context: BuilderPrepareContext<Payload, Params>) => unknown | Promise<unknown>;
2520
+ type BuilderAnalyzeStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = (context: BuilderAnalyzeContext<Payload, Params>) => unknown | Promise<unknown>;
2521
+ type BuilderScoreStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = (context: BuilderScoreContext<Payload, Params>) => GenerateScoreResult | number | Promise<GenerateScoreResult | number>;
2522
+ type BuilderReasonStep<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> = (context: BuilderReasonContext<Payload, Params>) => GenerateReasonResult | string | Promise<GenerateReasonResult | string>;
2523
+ type BuildScorerCustomOptions<Payload extends Record<string, unknown> = Record<string, unknown>, Params extends Record<string, unknown> = Record<string, unknown>> = {
2524
+ id: string;
2525
+ label?: string;
2526
+ description?: string;
2527
+ metadata?: Record<string, unknown> | null;
2528
+ sampling?: SamplingPolicy;
2529
+ params?: Params | ((payload: Payload) => Params | undefined | Promise<Params | undefined>);
2530
+ };
2531
+ type BuildScorerOptions<Payload extends Record<string, unknown> = Record<string, unknown>, Params extends Record<string, unknown> = Record<string, unknown>> = BuildScorerCustomOptions<Payload, Params>;
2532
+ interface BuildScorerRunArgs<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> {
2533
+ payload: Payload;
2534
+ params?: Params;
2535
+ sampling?: SamplingPolicy;
2536
+ }
2537
+ interface BuildScorerRunResult<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> {
2538
+ id: string;
2539
+ status: "success" | "error" | "skipped";
2540
+ score: number | null;
2541
+ reason?: string;
2542
+ metadata: Record<string, unknown> | null;
2543
+ durationMs: number;
2544
+ sampling?: ReturnType<typeof buildSamplingMetadata>;
2545
+ rawResult: ScorerResult;
2546
+ payload: Payload;
2547
+ params: Params;
2548
+ steps: BuilderResultsSnapshot;
2549
+ }
2550
+ interface ScorerBuilder<Payload extends Record<string, unknown>, Params extends Record<string, unknown>> {
2551
+ prepare(step: BuilderPrepareStep<Payload, Params>): ScorerBuilder<Payload, Params>;
2552
+ analyze(step: BuilderAnalyzeStep<Payload, Params>): ScorerBuilder<Payload, Params>;
2553
+ score(step: BuilderScoreStep<Payload, Params>): ScorerBuilder<Payload, Params>;
2554
+ reason(step: BuilderReasonStep<Payload, Params>): ScorerBuilder<Payload, Params>;
2555
+ build(): LocalScorerDefinition<Payload, Params>;
2556
+ run(args: BuildScorerRunArgs<Payload, Params>): Promise<BuildScorerRunResult<Payload, Params>>;
2557
+ getId(): string;
2558
+ getLabel(): string;
2559
+ getDescription(): string | undefined;
2560
+ }
2561
+ declare function buildScorer<Payload extends Record<string, unknown> = Record<string, unknown>, Params extends Record<string, unknown> = Record<string, unknown>>(options: BuildScorerOptions<Payload, Params>): ScorerBuilder<Payload, Params>;
2562
+
2254
2563
  /**
2255
2564
  * Unified Observability Types for VoltAgent
2256
2565
  *
@@ -3220,6 +3529,7 @@ declare class AgentTraceContext {
3220
3529
  error?: Error | any;
3221
3530
  attributes?: Record<string, any>;
3222
3531
  }): void;
3532
+ private resolveParentSpan;
3223
3533
  /**
3224
3534
  * Get the active context for manual context propagation
3225
3535
  */
@@ -3244,6 +3554,15 @@ interface ApiToolInfo {
3244
3554
  interface ToolWithNodeId extends BaseTool {
3245
3555
  node_id: string;
3246
3556
  }
3557
+ interface AgentScorerState {
3558
+ key: string;
3559
+ id: string;
3560
+ name: string;
3561
+ node_id: string;
3562
+ sampling?: SamplingPolicy;
3563
+ metadata?: Record<string, unknown> | null;
3564
+ params?: Record<string, unknown> | null;
3565
+ }
3247
3566
  /**
3248
3567
  * SubAgent data structure for agent state
3249
3568
  */
@@ -3257,6 +3576,7 @@ interface SubAgentStateData {
3257
3576
  memory?: AgentMemoryState;
3258
3577
  node_id: string;
3259
3578
  subAgents?: SubAgentStateData[];
3579
+ scorers?: AgentScorerState[];
3260
3580
  methodConfig?: {
3261
3581
  method: string;
3262
3582
  schema?: string;
@@ -3304,6 +3624,7 @@ interface AgentFullState {
3304
3624
  tools: ToolWithNodeId[];
3305
3625
  subAgents: SubAgentStateData[];
3306
3626
  memory: AgentMemoryState;
3627
+ scorers?: AgentScorerState[];
3307
3628
  retriever?: {
3308
3629
  name: string;
3309
3630
  description?: string;
@@ -3427,7 +3748,59 @@ type AgentOptions = {
3427
3748
  voltOpsClient?: VoltOpsClient;
3428
3749
  observability?: VoltAgentObservability;
3429
3750
  context?: ContextInput;
3751
+ eval?: AgentEvalConfig;
3430
3752
  };
3753
+ type AgentEvalOperationType = "generateText" | "streamText" | "generateObject" | "streamObject";
3754
+ interface AgentEvalPayload {
3755
+ operationId: string;
3756
+ operationType: AgentEvalOperationType;
3757
+ input?: string | null;
3758
+ output?: string | null;
3759
+ rawInput?: string | UIMessage[] | BaseMessage[];
3760
+ rawOutput?: unknown;
3761
+ userId?: string;
3762
+ conversationId?: string;
3763
+ traceId: string;
3764
+ spanId: string;
3765
+ metadata?: Record<string, unknown>;
3766
+ }
3767
+ type AgentEvalContext = AgentEvalPayload & Record<string, unknown> & {
3768
+ agentId: string;
3769
+ agentName: string;
3770
+ timestamp: string;
3771
+ rawPayload: AgentEvalPayload;
3772
+ };
3773
+ type AgentEvalParams = Record<string, unknown>;
3774
+ type AgentEvalSamplingPolicy = SamplingPolicy;
3775
+ type AgentEvalScorerFactory = () => LocalScorerDefinition<AgentEvalContext, Record<string, unknown>> | Promise<LocalScorerDefinition<AgentEvalContext, Record<string, unknown>>>;
3776
+ type AgentEvalScorerReference = LocalScorerDefinition<AgentEvalContext, Record<string, unknown>> | AgentEvalScorerFactory;
3777
+ interface AgentEvalResult {
3778
+ scorerId: string;
3779
+ scorerName?: string;
3780
+ status: "success" | "error" | "skipped";
3781
+ score?: number | null;
3782
+ metadata?: Record<string, unknown> | null;
3783
+ error?: unknown;
3784
+ durationMs?: number;
3785
+ payload: AgentEvalPayload;
3786
+ rawPayload: AgentEvalPayload;
3787
+ }
3788
+ interface AgentEvalScorerConfig {
3789
+ scorer: AgentEvalScorerReference;
3790
+ params?: AgentEvalParams | ((context: AgentEvalContext) => AgentEvalParams | undefined | Promise<AgentEvalParams | undefined>);
3791
+ sampling?: AgentEvalSamplingPolicy;
3792
+ id?: string;
3793
+ onResult?: (result: AgentEvalResult) => void | Promise<void>;
3794
+ buildPayload?: (context: AgentEvalContext) => Record<string, unknown> | Promise<Record<string, unknown>>;
3795
+ buildParams?: (context: AgentEvalContext) => AgentEvalParams | undefined | Promise<AgentEvalParams | undefined>;
3796
+ }
3797
+ interface AgentEvalConfig {
3798
+ scorers: Record<string, AgentEvalScorerConfig>;
3799
+ triggerSource?: string;
3800
+ environment?: string;
3801
+ sampling?: AgentEvalSamplingPolicy;
3802
+ redact?: (payload: AgentEvalPayload) => AgentEvalPayload;
3803
+ }
3431
3804
  /**
3432
3805
  * Agent status information
3433
3806
  */
@@ -3907,6 +4280,7 @@ declare class Agent {
3907
4280
  private readonly subAgentManager;
3908
4281
  private readonly voltOpsClient?;
3909
4282
  private readonly prompts?;
4283
+ private readonly evalConfig?;
3910
4284
  constructor(options: AgentOptions);
3911
4285
  /**
3912
4286
  * Generate text response
@@ -3947,6 +4321,8 @@ declare class Agent {
3947
4321
  * Calculate delegation depth
3948
4322
  */
3949
4323
  private calculateDelegationDepth;
4324
+ private enqueueEvalScoring;
4325
+ private createEvalHost;
3950
4326
  /**
3951
4327
  * Get observability instance (lazy initialization)
3952
4328
  */
@@ -6104,10 +6480,6 @@ declare class InMemoryStorageAdapter implements StorageAdapter {
6104
6480
  private users;
6105
6481
  private workflowStates;
6106
6482
  private workflowStatesByWorkflow;
6107
- private storageLimit;
6108
- constructor(options?: {
6109
- storageLimit?: number;
6110
- });
6111
6483
  /**
6112
6484
  * Add a single message
6113
6485
  */
@@ -7035,6 +7407,7 @@ declare enum NodeType {
7035
7407
  RETRIEVER = "retriever",
7036
7408
  VECTOR = "vector",
7037
7409
  EMBEDDING = "embedding",
7410
+ SCORER = "scorer",
7038
7411
  WORKFLOW_STEP = "workflow_step",
7039
7412
  WORKFLOW_AGENT_STEP = "workflow_agent_step",
7040
7413
  WORKFLOW_FUNC_STEP = "workflow_func_step",
@@ -7665,4 +8038,4 @@ declare class VoltAgent {
7665
8038
  */
7666
8039
  declare function convertUsage(usage: LanguageModelUsage | undefined): UsageInfo | undefined;
7667
8040
 
7668
- export { A2AServerRegistry, AbortError, Agent, type AgentFullState, type AgentHookOnEnd, type AgentHookOnError, type AgentHookOnHandoff, type AgentHookOnPrepareMessages, type AgentHookOnPrepareModelMessages, type AgentHookOnStart, type AgentHookOnStepFinish, type AgentHookOnToolEnd, type AgentHookOnToolStart, type AgentHooks, type AgentOptions, AgentRegistry, type AgentResponse, type AgentStatus, type AgentTool, AiSdkEmbeddingAdapter, type AllowedVariableValue, type ApiToolInfo, type BaseEventMetadata, type BaseGenerationOptions, type BaseLLMOptions, type BaseMessage, BaseRetriever, type BaseTool, type BaseToolCall, type CachedPrompt, type ChatMessage, ClientHTTPError, type ClientSideToolResult, type CloudflareFetchHandler, type Conversation, ConversationAlreadyExistsError, ConversationNotFoundError, type ConversationQueryOptions, type ConversationQueryOptions as ConversationQueryOptionsV2, type Conversation as ConversationV2, type CreateConversationInput, type CreateConversationInput as CreateConversationInputV2, type CreateReasoningToolsOptions, DEFAULT_INSTRUCTIONS, type DataContent, type Document, type DynamicValue, type DynamicValueOptions, type EmbeddingAdapter$1 as EmbeddingAdapter, EmbeddingAdapterNotConfiguredError, EmbeddingError, type ExtractVariableNames, FEW_SHOT_EXAMPLES, type GenerateObjectOptions, type GenerateObjectSubAgentConfig, type GenerateTextOptions, type GenerateTextSubAgentConfig, type GetMessagesOptions, type IServerProvider, type IServerlessProvider, type VoltOpsClient$1 as IVoltOpsClient, InMemoryStorageAdapter$1 as InMemoryObservabilityAdapter, InMemoryStorageAdapter, InMemoryVectorAdapter, type InferGenerateObjectResponse, type InferGenerateTextResponse, type InferMessage, type InferModel, type InferProviderParams, type InferStreamResponse, type InferTool, type LLMProvider, LazyRemoteExportProcessor, LocalStorageSpanProcessor, type LogFilter, LoggerProxy, MCPConfiguration, type MCPElicitationAdapter, type MCPLoggingAdapter, type MCPPromptsAdapter, type MCPResourcesAdapter, MCPServerRegistry, type ManagedMemoryAddMessageInput, type ManagedMemoryAddMessagesInput, type ManagedMemoryClearMessagesInput, type ManagedMemoryConnectionInfo, type ManagedMemoryConversationsClient, type ManagedMemoryCredentialCreateResult, type ManagedMemoryCredentialListResult, type ManagedMemoryCredentialSummary, type ManagedMemoryDatabaseSummary, type ManagedMemoryGetMessagesInput, type ManagedMemoryMessagesClient, type ManagedMemorySetWorkingMemoryInput, type ManagedMemoryStatus, type ManagedMemoryUpdateConversationInput, type ManagedMemoryVoltOpsClient, type ManagedMemoryWorkflowStateUpdateInput, type ManagedMemoryWorkflowStatesClient, type ManagedMemoryWorkingMemoryClient, type ManagedMemoryWorkingMemoryInput, Memory, type MemoryConfig, type MemoryOptions, type MemoryStorageMetadata, type MemoryUpdateMode, Memory as MemoryV2, MemoryV2Error, type MessageContent, MessageContentBuilder, type MessageRole, type ModelToolCall, NextAction, NodeType, VoltAgentObservability$1 as NodeVoltAgentObservability, type ObservabilityConfig, type ObservabilityLogRecord, type ObservabilitySpan, type ObservabilityStorageAdapter, type ObservabilityWebSocketEvent, type OnEndHookArgs, type OnErrorHookArgs, type OnHandoffHookArgs, type OnPrepareMessagesHookArgs, type OnPrepareMessagesHookResult, type OnPrepareModelMessagesHookArgs, type OnPrepareModelMessagesHookResult, type OnStartHookArgs, type OnStepFinishHookArgs, type OnToolEndHookArgs, type OnToolStartHookArgs, type OperationContext, type PackageUpdateInfo, type PromptApiClient, type PromptApiResponse, type PromptContent, type PromptCreator, type PromptHelper, type PromptReference, type PromptTemplate, type ProviderObjectResponse, type ProviderObjectStreamResponse, type ProviderParams, type ProviderResponse, type ProviderTextResponse, type ProviderTextStreamResponse, type ReadableStreamType, type ReasoningStep, ReasoningStepSchema, type RegisterOptions, type RegisteredWorkflow, type RemoteLogExportConfig, RemoteLogProcessor, type RetrieveOptions, type Retriever, type RetrieverOptions, type SearchOptions, type SearchResult, type ServerAgentResponse, type ServerApiResponse, type ServerProviderDeps, type ServerProviderFactory, type ServerWorkflowResponse, type ServerlessProviderFactory, type ServerlessRemoteEndpointConfig, type ServerlessRemoteExportConfig, type ServerlessRequestHandler, ServerlessVoltAgentObservability, type SpanAttributes, type SpanEvent, type SpanFilterConfig, SpanFilterProcessor, SpanKind, type SpanLink, type SpanStatus, SpanStatusCode, type SpanTreeNode, type StepChunkCallback, type StepFinishCallback, type StepWithContent, type StopWhen, type StorageAdapter, StorageError, StorageLogProcessor, type StoredUIMessage, type StreamObjectFinishResult, type StreamObjectOnFinishCallback, type StreamObjectOptions, type StreamObjectSubAgentConfig, type StreamPart, type StreamTextFinishResult, type StreamTextOnFinishCallback, type StreamTextOptions, type StreamTextSubAgentConfig, type SubAgentConfig, type SubAgentMethod, type SubAgentStateData, type SupervisorConfig, type TemplateVariables, type TimelineEventCoreLevel, type TimelineEventCoreStatus, type TimelineEventCoreType, Tool, type ToolCall, ToolDeniedError, type ToolErrorInfo, type ToolExecuteOptions, ToolManager, type ToolOptions, type ToolSchema, type ToolStatus, type ToolStatusInfo, type ToolWithNodeId, type Toolkit, type Usage, type UsageInfo, type VectorAdapter, VectorAdapterNotConfiguredError, VectorError, type VectorItem, type VectorSearchOptions, type Voice, type VoiceEventData, type VoiceEventType, type VoiceMetadata, type VoiceOptions, VoltAgent, VoltAgentError, VoltAgentObservability, type VoltAgentOptions, type VoltAgentStreamTextResult, type VoltAgentTextStreamPart, VoltOpsClient, type VoltOpsClientOptions, VoltOpsPromptApiClient, type VoltOpsPromptManager, VoltOpsPromptManagerImpl, WebSocketEventEmitter, WebSocketLogProcessor, WebSocketSpanProcessor, type Workflow, type WorkflowConfig, type WorkflowExecutionContext, WorkflowRegistry, type WorkflowStateEntry, type WorkflowStats, type WorkflowStepContext, type WorkflowStepType, type WorkflowTimelineEvent, type WorkingMemoryConfig, type WorkingMemoryScope, type WorkingMemorySummary, type WorkingMemoryUpdateOptions, addTimestampToMessage, andAgent, andAll, andRace, andTap, andThen, andWhen, andWorkflow, appendToMessage, buildRetrieverLogMessage, buildSpanTree, checkForUpdates, convertUsage, cosineSimilarity, createHooks, createNodeId, createPrompt, createReasoningTools, createRetrieverTool, createSimpleTemplateEngine, createSubagent, createSuspendController, createTool, createToolkit, createVoltAgentObservability, createVoltOpsClient, createWorkflow, createWorkflowChain, createWorkflowStepNodeId, VoltAgent as default, extractFileParts, extractImageParts, extractText, extractTextParts, extractWorkflowStepInfo, filterContentParts, getContentLength, getEnvVar, getGlobalLogBuffer, getGlobalLogger, getNodeTypeFromNodeId, getWorkflowStepNodeType, hasContent, hasFilePart, hasImagePart, hasTextPart, isAbortError, isNodeRuntime, isServerlessRuntime, isStructuredContent, isTextContent, isVoltAgentError, mapMessageContent, messageHelpers, normalizeContent, normalizeToArray, prependToMessage, readableLogRecordToObservabilityLog, readableSpanToObservabilitySpan, safeJsonParse, serializeValueForDebug, tool, transformTextContent, updateAllPackages, updateSinglePackage, zodSchemaToJsonUI };
8041
+ export { A2AServerRegistry, AbortError, Agent, type AgentEvalConfig, type AgentEvalContext, type AgentEvalOperationType, type AgentEvalPayload, type AgentEvalResult, type AgentEvalSamplingPolicy, type AgentEvalScorerConfig, type AgentEvalScorerFactory, type AgentEvalScorerReference, type AgentFullState, type AgentHookOnEnd, type AgentHookOnError, type AgentHookOnHandoff, type AgentHookOnPrepareMessages, type AgentHookOnPrepareModelMessages, type AgentHookOnStart, type AgentHookOnStepFinish, type AgentHookOnToolEnd, type AgentHookOnToolStart, type AgentHooks, type AgentOptions, AgentRegistry, type AgentResponse, type AgentScorerState, type AgentStatus, type AgentTool, AiSdkEmbeddingAdapter, type AllowedVariableValue, type ApiToolInfo, type BaseEventMetadata, type BaseGenerationOptions, type BaseLLMOptions, type BaseMessage, BaseRetriever, type BaseTool, type BaseToolCall, type BuildScorerOptions, type BuildScorerRunArgs, type BuildScorerRunResult, type BuilderAnalyzeContext, type BuilderPrepareContext, type BuilderReasonContext, type BuilderScoreContext, type CachedPrompt, type ChatMessage, ClientHTTPError, type ClientSideToolResult, type CloudflareFetchHandler, type Conversation, ConversationAlreadyExistsError, ConversationNotFoundError, type ConversationQueryOptions, type ConversationQueryOptions as ConversationQueryOptionsV2, type Conversation as ConversationV2, type CreateConversationInput, type CreateConversationInput as CreateConversationInputV2, type CreateReasoningToolsOptions, type CreateScorerOptions, DEFAULT_INSTRUCTIONS, type DataContent, type Document, type DynamicValue, type DynamicValueOptions, type EmbeddingAdapter$1 as EmbeddingAdapter, EmbeddingAdapterNotConfiguredError, EmbeddingError, type ExtractVariableNames, FEW_SHOT_EXAMPLES, type GenerateObjectOptions, type GenerateObjectSubAgentConfig, type GenerateReasonResult, type GenerateScoreResult, type GenerateScoreStep, type GenerateTextOptions, type GenerateTextSubAgentConfig, type GetMessagesOptions, type IServerProvider, type IServerlessProvider, type VoltOpsClient$1 as IVoltOpsClient, InMemoryStorageAdapter$1 as InMemoryObservabilityAdapter, InMemoryStorageAdapter, InMemoryVectorAdapter, type InferGenerateObjectResponse, type InferGenerateTextResponse, type InferMessage, type InferModel, type InferProviderParams, type InferStreamResponse, type InferTool, type LLMProvider, LazyRemoteExportProcessor, type LocalScorerDefinition, type LocalScorerExecutionResult, LocalStorageSpanProcessor, type LogFilter, LoggerProxy, MCPConfiguration, type MCPElicitationAdapter, type MCPLoggingAdapter, type MCPPromptsAdapter, type MCPResourcesAdapter, MCPServerRegistry, type ManagedMemoryAddMessageInput, type ManagedMemoryAddMessagesInput, type ManagedMemoryClearMessagesInput, type ManagedMemoryConnectionInfo, type ManagedMemoryConversationsClient, type ManagedMemoryCredentialCreateResult, type ManagedMemoryCredentialListResult, type ManagedMemoryCredentialSummary, type ManagedMemoryDatabaseSummary, type ManagedMemoryGetMessagesInput, type ManagedMemoryMessagesClient, type ManagedMemorySetWorkingMemoryInput, type ManagedMemoryStatus, type ManagedMemoryUpdateConversationInput, type ManagedMemoryVoltOpsClient, type ManagedMemoryWorkflowStateUpdateInput, type ManagedMemoryWorkflowStatesClient, type ManagedMemoryWorkingMemoryClient, type ManagedMemoryWorkingMemoryInput, Memory, type MemoryConfig, type MemoryOptions, type MemoryStorageMetadata, type MemoryUpdateMode, Memory as MemoryV2, MemoryV2Error, type MessageContent, MessageContentBuilder, type MessageRole, type ModelToolCall, NextAction, NodeType, VoltAgentObservability$1 as NodeVoltAgentObservability, type ObservabilityConfig, type ObservabilityLogRecord, type ObservabilitySpan, type ObservabilityStorageAdapter, type ObservabilityWebSocketEvent, type OnEndHookArgs, type OnErrorHookArgs, type OnHandoffHookArgs, type OnPrepareMessagesHookArgs, type OnPrepareMessagesHookResult, type OnPrepareModelMessagesHookArgs, type OnPrepareModelMessagesHookResult, type OnStartHookArgs, type OnStepFinishHookArgs, type OnToolEndHookArgs, type OnToolStartHookArgs, type OperationContext, type PackageUpdateInfo, type PromptApiClient, type PromptApiResponse, type PromptContent, type PromptCreator, type PromptHelper, type PromptReference, type PromptTemplate, type ProviderObjectResponse, type ProviderObjectStreamResponse, type ProviderParams, type ProviderResponse, type ProviderTextResponse, type ProviderTextStreamResponse, type ReadableStreamType, type ReasoningStep, ReasoningStepSchema, type RegisterOptions, type RegisteredWorkflow, type RemoteLogExportConfig, RemoteLogProcessor, type RetrieveOptions, type Retriever, type RetrieverOptions, type RunLocalScorersArgs, type RunLocalScorersResult, type SamplingMetadata, type SamplingPolicy, type ScorerBuilder, type ScorerContext, type ScorerLifecycleScope, type ScorerPipelineContext, type ScorerReasonContext, type ScorerResult, type SearchOptions, type SearchResult, type ServerAgentResponse, type ServerApiResponse, type ServerProviderDeps, type ServerProviderFactory, type ServerWorkflowResponse, type ServerlessProviderFactory, type ServerlessRemoteEndpointConfig, type ServerlessRemoteExportConfig, type ServerlessRequestHandler, ServerlessVoltAgentObservability, type SpanAttributes, type SpanEvent, type SpanFilterConfig, SpanFilterProcessor, SpanKind, type SpanLink, type SpanStatus, SpanStatusCode, type SpanTreeNode, type StepChunkCallback, type StepFinishCallback, type StepWithContent, type StopWhen, type StorageAdapter, StorageError, StorageLogProcessor, type StoredUIMessage, type StreamObjectFinishResult, type StreamObjectOnFinishCallback, type StreamObjectOptions, type StreamObjectSubAgentConfig, type StreamPart, type StreamTextFinishResult, type StreamTextOnFinishCallback, type StreamTextOptions, type StreamTextSubAgentConfig, type SubAgentConfig, type SubAgentMethod, type SubAgentStateData, type SupervisorConfig, type TemplateVariables, type TimelineEventCoreLevel, type TimelineEventCoreStatus, type TimelineEventCoreType, Tool, type ToolCall, ToolDeniedError, type ToolErrorInfo, type ToolExecuteOptions, ToolManager, type ToolOptions, type ToolSchema, type ToolStatus, type ToolStatusInfo, type ToolWithNodeId, type Toolkit, type Usage, type UsageInfo, type VectorAdapter, VectorAdapterNotConfiguredError, VectorError, type VectorItem, type VectorSearchOptions, type Voice, type VoiceEventData, type VoiceEventType, type VoiceMetadata, type VoiceOptions, VoltAgent, VoltAgentError, VoltAgentObservability, type VoltAgentOptions, type VoltAgentStreamTextResult, type VoltAgentTextStreamPart, type VoltOpsAppendEvalRunResultPayload, type VoltOpsAppendEvalRunResultsRequest, VoltOpsClient, type VoltOpsClientOptions, type VoltOpsCompleteEvalRunRequest, type VoltOpsCreateEvalRunRequest, type VoltOpsCreateScorerRequest, type VoltOpsEvalResultStatus, type VoltOpsEvalRunCompletionSummaryPayload, type VoltOpsEvalRunErrorPayload, type VoltOpsEvalRunResultLiveMetadata, type VoltOpsEvalRunResultScorePayload, type VoltOpsEvalRunStatus, type VoltOpsEvalRunSummary, VoltOpsPromptApiClient, type VoltOpsPromptManager, VoltOpsPromptManagerImpl, type VoltOpsScorerSummary, type VoltOpsTerminalEvalRunStatus, WebSocketEventEmitter, WebSocketLogProcessor, WebSocketSpanProcessor, type WeightedBlendComponent, type WeightedBlendOptions, type Workflow, type WorkflowConfig, type WorkflowExecutionContext, WorkflowRegistry, type WorkflowStateEntry, type WorkflowStats, type WorkflowStepContext, type WorkflowStepType, type WorkflowTimelineEvent, type WorkingMemoryConfig, type WorkingMemoryScope, type WorkingMemorySummary, type WorkingMemoryUpdateOptions, addTimestampToMessage, andAgent, andAll, andRace, andTap, andThen, andWhen, andWorkflow, appendToMessage, buildRetrieverLogMessage, buildSamplingMetadata, buildScorer, buildSpanTree, checkForUpdates, convertUsage, cosineSimilarity, createHooks, createNodeId, createPrompt, createReasoningTools, createRetrieverTool, createScorer, createSimpleTemplateEngine, createSubagent, createSuspendController, createTool, createToolkit, createVoltAgentObservability, createVoltOpsClient, createWorkflow, createWorkflowChain, createWorkflowStepNodeId, VoltAgent as default, extractFileParts, extractImageParts, extractText, extractTextParts, extractWorkflowStepInfo, filterContentParts, getContentLength, getEnvVar, getGlobalLogBuffer, getGlobalLogger, getNodeTypeFromNodeId, getWorkflowStepNodeType, hasContent, hasFilePart, hasImagePart, hasTextPart, isAbortError, isNodeRuntime, isServerlessRuntime, isStructuredContent, isTextContent, isVoltAgentError, mapMessageContent, messageHelpers, normalizeContent, normalizeScorerResult, normalizeToArray, prependToMessage, readableLogRecordToObservabilityLog, readableSpanToObservabilitySpan, runLocalScorers, safeJsonParse, serializeValueForDebug, shouldSample, tool, transformTextContent, updateAllPackages, updateSinglePackage, weightedBlend, zodSchemaToJsonUI };