raindrop-ai 0.7.0 → 0.7.1
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/README.md +25 -0
- package/dist/{chunk-4Z7UHC7X.mjs → chunk-6WL655QL.mjs} +239 -132
- package/dist/{chunk-OX6X4ZWQ.mjs → chunk-SHH4SYVY.mjs} +1470 -1015
- package/dist/evals/cli.js +1688 -1148
- package/dist/evals/cli.mjs +2 -2
- package/dist/{index-n8AIAiRR.d.mts → index-DuJkG47V.d.mts} +29 -3
- package/dist/{index-n8AIAiRR.d.ts → index-DuJkG47V.d.ts} +29 -3
- package/dist/index.d.mts +1 -2
- package/dist/index.d.ts +1 -2
- package/dist/index.js +1694 -1151
- package/dist/index.mjs +5 -6
- package/dist/tracing/index.d.mts +3 -2
- package/dist/tracing/index.d.ts +3 -2
- package/dist/tracing/index.js +1085 -973
- package/dist/tracing/index.mjs +1 -1
- package/package.json +1 -1
package/dist/evals/cli.mjs
CHANGED
|
@@ -5,8 +5,8 @@ import {
|
|
|
5
5
|
publishEvalSuite,
|
|
6
6
|
readEvaluator,
|
|
7
7
|
runEvalSuite
|
|
8
|
-
} from "../chunk-
|
|
9
|
-
import "../chunk-
|
|
8
|
+
} from "../chunk-6WL655QL.mjs";
|
|
9
|
+
import "../chunk-SHH4SYVY.mjs";
|
|
10
10
|
import {
|
|
11
11
|
isEvalSuiteDefinition
|
|
12
12
|
} from "../chunk-RMP6BZSY.mjs";
|
|
@@ -3,8 +3,17 @@ import { Instrumentation } from '@opentelemetry/instrumentation';
|
|
|
3
3
|
import { Span, ContextManager as ContextManager$1, TextMapPropagator } from '@opentelemetry/api';
|
|
4
4
|
import z$1, { z } from 'zod';
|
|
5
5
|
import { SpanExporter, SpanProcessor } from '@opentelemetry/sdk-trace-base';
|
|
6
|
-
import './eval-scope-BYkhP4uI.js';
|
|
7
6
|
|
|
7
|
+
/** Kept structural so public integration declarations do not depend on private schemas. */
|
|
8
|
+
type AppGitOptions = {
|
|
9
|
+
commitSha?: string | null;
|
|
10
|
+
commitDirty?: boolean | null;
|
|
11
|
+
branch?: string | null;
|
|
12
|
+
detectBranch?: boolean;
|
|
13
|
+
sourceDirectory?: string;
|
|
14
|
+
autoDetect?: boolean;
|
|
15
|
+
};
|
|
16
|
+
type AppGitSnapshot = Readonly<Record<string, unknown>>;
|
|
8
17
|
/** The reverse reference a detached child carries on its own spans. */
|
|
9
18
|
type DetachedChildLink = {
|
|
10
19
|
/** The caller's event id. */
|
|
@@ -175,7 +184,7 @@ type OtherAttachment = BaseAttachment & {
|
|
|
175
184
|
type Attachment = CodeAttachment | OtherAttachment;
|
|
176
185
|
type RaindropFeatureFlagValue = string | number | boolean;
|
|
177
186
|
type RaindropFeatureFlags = string[] | Record<string, RaindropFeatureFlagValue>;
|
|
178
|
-
type RaindropPropertyLeaf = string | boolean | number;
|
|
187
|
+
type RaindropPropertyLeaf = string | boolean | number | null;
|
|
179
188
|
type RaindropPropertyValue = RaindropPropertyLeaf | RaindropPropertyValue[] | {
|
|
180
189
|
[key: string]: RaindropPropertyValue;
|
|
181
190
|
};
|
|
@@ -54490,6 +54499,8 @@ type ReplayPlanRow = ReplayRow & {
|
|
|
54490
54499
|
* this, so pass the client you already built.
|
|
54491
54500
|
*/
|
|
54492
54501
|
type ReplayClient = {
|
|
54502
|
+
getAppGitSnapshot?(): AppGitSnapshot;
|
|
54503
|
+
getReplayAppGitSnapshot?(): AppGitSnapshot;
|
|
54493
54504
|
readonly connection: {
|
|
54494
54505
|
readonly apiKey?: string;
|
|
54495
54506
|
readonly projectId: string | undefined;
|
|
@@ -54585,6 +54596,8 @@ type LocalEvaluator<AgentResult = unknown> = {
|
|
|
54585
54596
|
readonly judge: LocalJudge<AgentResult, LocalNumberVerdict>;
|
|
54586
54597
|
});
|
|
54587
54598
|
type CreateReplayOptions = {
|
|
54599
|
+
/** Optional agent source override; metadata discovery never delays a replay. */
|
|
54600
|
+
appGit?: AppGitOptions | false;
|
|
54588
54601
|
/** A project-scoped dataset slug, or a manifest already pinned by discovery. */
|
|
54589
54602
|
dataset: string | EvalDatasetManifest$1;
|
|
54590
54603
|
/** The cases to run. The SDK preserves their order from the pinned manifest. */
|
|
@@ -54638,6 +54651,7 @@ declare function createReplay(client: ReplayClient, options: CreateReplayOptions
|
|
|
54638
54651
|
*/
|
|
54639
54652
|
declare function claimReplay(client: ReplayClient, replayId: string, options?: {
|
|
54640
54653
|
queryUrl?: string;
|
|
54654
|
+
appGit?: AppGitOptions | false;
|
|
54641
54655
|
}): Promise<ReplayHandle>;
|
|
54642
54656
|
declare function readReplay(client: ReplayClient, replayId: string, options?: {
|
|
54643
54657
|
queryUrl?: string;
|
|
@@ -60509,6 +60523,8 @@ declare function createEvalSuiteRun<AgentResult>(client: EvalClient, definition:
|
|
|
60509
60523
|
declare function runEvalSuite<AgentResult>(client: EvalClient, definition: EvalSuiteDefinition<AgentResult>, options?: RunEvalSuiteOptions): Promise<EvalSuiteResult>;
|
|
60510
60524
|
|
|
60511
60525
|
interface AnalyticsConfig {
|
|
60526
|
+
/** Application source revision; optional discovery never waits on Git. */
|
|
60527
|
+
appGit?: AppGitOptions | false;
|
|
60512
60528
|
wizardSession?: string;
|
|
60513
60529
|
/**
|
|
60514
60530
|
* Raindrop write key. When omitted (or empty), the SDK runs in local-only mode:
|
|
@@ -60696,6 +60712,7 @@ type RaindropConnection = {
|
|
|
60696
60712
|
readonly projectId: string | undefined;
|
|
60697
60713
|
};
|
|
60698
60714
|
declare class Raindrop {
|
|
60715
|
+
private appGitContext;
|
|
60699
60716
|
private wizardSession;
|
|
60700
60717
|
private writeKey;
|
|
60701
60718
|
private apiKey;
|
|
@@ -60761,6 +60778,7 @@ declare class Raindrop {
|
|
|
60761
60778
|
event: string;
|
|
60762
60779
|
userId: string;
|
|
60763
60780
|
}): Interaction;
|
|
60781
|
+
private propertiesWithAppGit;
|
|
60764
60782
|
/**
|
|
60765
60783
|
* Returns a tracer object that can be used to create spans and tools that aren't tied to any interaction.
|
|
60766
60784
|
* For chat interaction use-case `begin` should be used.
|
|
@@ -60973,4 +60991,12 @@ declare class Raindrop {
|
|
|
60973
60991
|
private closeWithinDeadline;
|
|
60974
60992
|
}
|
|
60975
60993
|
|
|
60976
|
-
|
|
60994
|
+
/** Evaluation context shared across asynchronous work. */
|
|
60995
|
+
type EvalScope = {
|
|
60996
|
+
readonly correlationId: string;
|
|
60997
|
+
};
|
|
60998
|
+
declare const EVAL_CORRELATION_ID_ATTRIBUTE: "raindrop.eval_correlation_id";
|
|
60999
|
+
declare const currentEvalScope: () => EvalScope | undefined;
|
|
61000
|
+
declare const withEvalScope: <Result>(scope: EvalScope, callback: () => Result) => Result;
|
|
61001
|
+
|
|
61002
|
+
export { type LocalBooleanVerdict as $, type AiTrackUsage as A, type BeginInteractionOptions as B, type CarrierHeaders as C, DEFAULT_EVAL_POLL_INTERVAL_MS as D, EVAL_CORRELATION_ID_ATTRIBUTE as E, type EvalLocalEvaluator as F, type EvalManifest as G, type EvalProgram as H, type InitializeOptions as I, EvalPublishError as J, type EvalRowResult as K, type EvalRowTracePair as L, type EvalRunReading as M, type EvalScope as N, type EvalSuiteDefinition as O, type PartialAiTrackEvent as P, type EvalSuiteDefinitionInput as Q, Raindrop as R, type SpanProcessorOptions as S, type TraceCarrier as T, type EvalSuiteResult as U, type EvalSuiteRunSession as V, type EvalVerdictAssessment as W, type EvaluateReplayOptions as X, type FinishInteractionOptions as Y, type HostedEvaluatorPin as Z, type IdentifyEvent as _, type RaindropSpanProcessor as a, type SpanParams as a$, LocalBooleanVerdictSchema as a0, type LocalDebuggerLiveEventInput as a1, type LocalEvaluator as a2, type LocalEvaluatorContext as a3, type LocalNumberVerdict as a4, LocalNumberVerdictSchema as a5, type LocalScoreVerdict as a6, LocalScoreVerdictSchema as a7, type LocalVerdict as a8, LocalVerdictSchema as a9, ReplayEvaluatorCountsSchema as aA, type ReplayEvaluatorStats as aB, ReplayEvaluatorStatsSchema as aC, type ReplayEvaluatorSummary as aD, ReplayEvaluatorSummarySchema as aE, type ReplayHandle as aF, type LocalEvaluatorContext$1 as aG, type ReplayOptions as aH, type ReplayReading as aI, type ReplayResult as aJ, type ReplayResultRow as aK, ReplayResultRowSchema as aL, type ReplayRow as aM, type ReplayRowOutcome as aN, type ReplayVerdict as aO, ReplayVerdictSchema as aP, type RunEvalSuiteOptions as aQ, type RunReplayOptions as aR, type SelfDiagnoseOptions as aS, type SelfDiagnosticsExecuteContext as aT, type SelfDiagnosticsExecuteResult as aU, type SelfDiagnosticsSignalDefinition as aV, type SelfDiagnosticsSignalDefinitions as aW, type SelfDiagnosticsTool as aX, type SelfDiagnosticsToolInputSchema as aY, type SelfDiagnosticsToolOptions as aZ, type SignalEvent as a_, MAX_EVAL_STATUS_RETRIES as aa, MAX_INGEST_SIZE_BYTES as ab, type NumericThreshold as ac, NumericThresholdSchema as ad, type PortableEvalArtifact as ae, type PortableEvalLimits as af, type PortableEvalOutcome as ag, type PortableEvalResult as ah, type PortableEvaluator as ai, type PortableJudgeRequest as aj, type PortableJudgeVerdict as ak, type PublishEvalDatasetOptions as al, type PublishEvalSuiteOptions as am, type PullEvalOptions as an, type PulledEvalSnapshot as ao, type RaindropConnection as ap, type RaindropFeatureFlagValue as aq, type RaindropFeatureFlags as ar, type RaindropProperties as as, type RaindropPropertyLeaf as at, type RaindropPropertyValue as au, type ReplayClient as av, type ReplayCounts as aw, ReplayCountsSchema as ax, type ReplayEvaluator as ay, type ReplayEvaluatorCounts as az, type Interaction as b, type SubagentDispatch as b0, type SubagentDispatchOptions as b1, type SubagentFinishOptions as b2, type ToolParams as b3, type ToolSpan as b4, type Trace as b5, type TraceContext as b6, TraceSchema as b7, type TrackToolParams as b8, type WorkflowParams as b9, replay as bA, resolveDisableBatching as bB, runEvalSuite as bC, runReplay as bD, traceOutput as bE, traceToolCalls as bF, traceTools as bG, verifyPortableEvalArtifact as bH, withEvalScope as bI, claimReplay as ba, compareEvalRuns as bb, createEvalSuiteRun as bc, createReplay as bd, currentEvalScope as be, defineDataset as bf, defineEvalSuite as bg, defineEvaluatorProgram as bh, defineLocalEvaluator as bi, evaluateEvalRun as bj, evaluatePortableEvaluator as bk, evaluateReplay as bl, importPortableEvaluator as bm, isEvalDataset as bn, isEvalSuiteDefinition as bo, isPublishedEvalSuite as bp, loadEvalSnapshot as bq, projectWorkshopTrace as br, publishEvalDataset as bs, publishEvalSuite as bt, pullEval as bu, readEvalDataset as bv, readEvalManifest as bw, readEvalRun as bx, readEvaluator as by, readReplay as bz, type AppGitSnapshot as c, type SubagentResumeOptions as d, type SubagentRun as e, type Tracer as f, type AiTrackEvent as g, type AppGitOptions as h, type Attachment as i, type AttachmentType as j, type BooleanThreshold as k, BooleanThresholdSchema as l, type CreateReplayOptions as m, DEFAULT_EVAL_WAIT_MS as n, DEFAULT_QUERY_URL as o, DEFAULT_TRACE_WAIT_MS as p, type DefinedEvalDataset as q, type EvalClient as r, type EvalDataset as s, type EvalDatasetInput as t, type EvalDatasetManifest as u, EvalDatasetPublishConflictError as v, type EvalDatasetRowInput as w, EvalDatasetRowInputSchema as x, type EvalDestination as y, type EvalEvaluator as z };
|
|
@@ -3,8 +3,17 @@ import { Instrumentation } from '@opentelemetry/instrumentation';
|
|
|
3
3
|
import { Span, ContextManager as ContextManager$1, TextMapPropagator } from '@opentelemetry/api';
|
|
4
4
|
import z$1, { z } from 'zod';
|
|
5
5
|
import { SpanExporter, SpanProcessor } from '@opentelemetry/sdk-trace-base';
|
|
6
|
-
import './eval-scope-BYkhP4uI.js';
|
|
7
6
|
|
|
7
|
+
/** Kept structural so public integration declarations do not depend on private schemas. */
|
|
8
|
+
type AppGitOptions = {
|
|
9
|
+
commitSha?: string | null;
|
|
10
|
+
commitDirty?: boolean | null;
|
|
11
|
+
branch?: string | null;
|
|
12
|
+
detectBranch?: boolean;
|
|
13
|
+
sourceDirectory?: string;
|
|
14
|
+
autoDetect?: boolean;
|
|
15
|
+
};
|
|
16
|
+
type AppGitSnapshot = Readonly<Record<string, unknown>>;
|
|
8
17
|
/** The reverse reference a detached child carries on its own spans. */
|
|
9
18
|
type DetachedChildLink = {
|
|
10
19
|
/** The caller's event id. */
|
|
@@ -175,7 +184,7 @@ type OtherAttachment = BaseAttachment & {
|
|
|
175
184
|
type Attachment = CodeAttachment | OtherAttachment;
|
|
176
185
|
type RaindropFeatureFlagValue = string | number | boolean;
|
|
177
186
|
type RaindropFeatureFlags = string[] | Record<string, RaindropFeatureFlagValue>;
|
|
178
|
-
type RaindropPropertyLeaf = string | boolean | number;
|
|
187
|
+
type RaindropPropertyLeaf = string | boolean | number | null;
|
|
179
188
|
type RaindropPropertyValue = RaindropPropertyLeaf | RaindropPropertyValue[] | {
|
|
180
189
|
[key: string]: RaindropPropertyValue;
|
|
181
190
|
};
|
|
@@ -54490,6 +54499,8 @@ type ReplayPlanRow = ReplayRow & {
|
|
|
54490
54499
|
* this, so pass the client you already built.
|
|
54491
54500
|
*/
|
|
54492
54501
|
type ReplayClient = {
|
|
54502
|
+
getAppGitSnapshot?(): AppGitSnapshot;
|
|
54503
|
+
getReplayAppGitSnapshot?(): AppGitSnapshot;
|
|
54493
54504
|
readonly connection: {
|
|
54494
54505
|
readonly apiKey?: string;
|
|
54495
54506
|
readonly projectId: string | undefined;
|
|
@@ -54585,6 +54596,8 @@ type LocalEvaluator<AgentResult = unknown> = {
|
|
|
54585
54596
|
readonly judge: LocalJudge<AgentResult, LocalNumberVerdict>;
|
|
54586
54597
|
});
|
|
54587
54598
|
type CreateReplayOptions = {
|
|
54599
|
+
/** Optional agent source override; metadata discovery never delays a replay. */
|
|
54600
|
+
appGit?: AppGitOptions | false;
|
|
54588
54601
|
/** A project-scoped dataset slug, or a manifest already pinned by discovery. */
|
|
54589
54602
|
dataset: string | EvalDatasetManifest$1;
|
|
54590
54603
|
/** The cases to run. The SDK preserves their order from the pinned manifest. */
|
|
@@ -54638,6 +54651,7 @@ declare function createReplay(client: ReplayClient, options: CreateReplayOptions
|
|
|
54638
54651
|
*/
|
|
54639
54652
|
declare function claimReplay(client: ReplayClient, replayId: string, options?: {
|
|
54640
54653
|
queryUrl?: string;
|
|
54654
|
+
appGit?: AppGitOptions | false;
|
|
54641
54655
|
}): Promise<ReplayHandle>;
|
|
54642
54656
|
declare function readReplay(client: ReplayClient, replayId: string, options?: {
|
|
54643
54657
|
queryUrl?: string;
|
|
@@ -60509,6 +60523,8 @@ declare function createEvalSuiteRun<AgentResult>(client: EvalClient, definition:
|
|
|
60509
60523
|
declare function runEvalSuite<AgentResult>(client: EvalClient, definition: EvalSuiteDefinition<AgentResult>, options?: RunEvalSuiteOptions): Promise<EvalSuiteResult>;
|
|
60510
60524
|
|
|
60511
60525
|
interface AnalyticsConfig {
|
|
60526
|
+
/** Application source revision; optional discovery never waits on Git. */
|
|
60527
|
+
appGit?: AppGitOptions | false;
|
|
60512
60528
|
wizardSession?: string;
|
|
60513
60529
|
/**
|
|
60514
60530
|
* Raindrop write key. When omitted (or empty), the SDK runs in local-only mode:
|
|
@@ -60696,6 +60712,7 @@ type RaindropConnection = {
|
|
|
60696
60712
|
readonly projectId: string | undefined;
|
|
60697
60713
|
};
|
|
60698
60714
|
declare class Raindrop {
|
|
60715
|
+
private appGitContext;
|
|
60699
60716
|
private wizardSession;
|
|
60700
60717
|
private writeKey;
|
|
60701
60718
|
private apiKey;
|
|
@@ -60761,6 +60778,7 @@ declare class Raindrop {
|
|
|
60761
60778
|
event: string;
|
|
60762
60779
|
userId: string;
|
|
60763
60780
|
}): Interaction;
|
|
60781
|
+
private propertiesWithAppGit;
|
|
60764
60782
|
/**
|
|
60765
60783
|
* Returns a tracer object that can be used to create spans and tools that aren't tied to any interaction.
|
|
60766
60784
|
* For chat interaction use-case `begin` should be used.
|
|
@@ -60973,4 +60991,12 @@ declare class Raindrop {
|
|
|
60973
60991
|
private closeWithinDeadline;
|
|
60974
60992
|
}
|
|
60975
60993
|
|
|
60976
|
-
|
|
60994
|
+
/** Evaluation context shared across asynchronous work. */
|
|
60995
|
+
type EvalScope = {
|
|
60996
|
+
readonly correlationId: string;
|
|
60997
|
+
};
|
|
60998
|
+
declare const EVAL_CORRELATION_ID_ATTRIBUTE: "raindrop.eval_correlation_id";
|
|
60999
|
+
declare const currentEvalScope: () => EvalScope | undefined;
|
|
61000
|
+
declare const withEvalScope: <Result>(scope: EvalScope, callback: () => Result) => Result;
|
|
61001
|
+
|
|
61002
|
+
export { type LocalBooleanVerdict as $, type AiTrackUsage as A, type BeginInteractionOptions as B, type CarrierHeaders as C, DEFAULT_EVAL_POLL_INTERVAL_MS as D, EVAL_CORRELATION_ID_ATTRIBUTE as E, type EvalLocalEvaluator as F, type EvalManifest as G, type EvalProgram as H, type InitializeOptions as I, EvalPublishError as J, type EvalRowResult as K, type EvalRowTracePair as L, type EvalRunReading as M, type EvalScope as N, type EvalSuiteDefinition as O, type PartialAiTrackEvent as P, type EvalSuiteDefinitionInput as Q, Raindrop as R, type SpanProcessorOptions as S, type TraceCarrier as T, type EvalSuiteResult as U, type EvalSuiteRunSession as V, type EvalVerdictAssessment as W, type EvaluateReplayOptions as X, type FinishInteractionOptions as Y, type HostedEvaluatorPin as Z, type IdentifyEvent as _, type RaindropSpanProcessor as a, type SpanParams as a$, LocalBooleanVerdictSchema as a0, type LocalDebuggerLiveEventInput as a1, type LocalEvaluator as a2, type LocalEvaluatorContext as a3, type LocalNumberVerdict as a4, LocalNumberVerdictSchema as a5, type LocalScoreVerdict as a6, LocalScoreVerdictSchema as a7, type LocalVerdict as a8, LocalVerdictSchema as a9, ReplayEvaluatorCountsSchema as aA, type ReplayEvaluatorStats as aB, ReplayEvaluatorStatsSchema as aC, type ReplayEvaluatorSummary as aD, ReplayEvaluatorSummarySchema as aE, type ReplayHandle as aF, type LocalEvaluatorContext$1 as aG, type ReplayOptions as aH, type ReplayReading as aI, type ReplayResult as aJ, type ReplayResultRow as aK, ReplayResultRowSchema as aL, type ReplayRow as aM, type ReplayRowOutcome as aN, type ReplayVerdict as aO, ReplayVerdictSchema as aP, type RunEvalSuiteOptions as aQ, type RunReplayOptions as aR, type SelfDiagnoseOptions as aS, type SelfDiagnosticsExecuteContext as aT, type SelfDiagnosticsExecuteResult as aU, type SelfDiagnosticsSignalDefinition as aV, type SelfDiagnosticsSignalDefinitions as aW, type SelfDiagnosticsTool as aX, type SelfDiagnosticsToolInputSchema as aY, type SelfDiagnosticsToolOptions as aZ, type SignalEvent as a_, MAX_EVAL_STATUS_RETRIES as aa, MAX_INGEST_SIZE_BYTES as ab, type NumericThreshold as ac, NumericThresholdSchema as ad, type PortableEvalArtifact as ae, type PortableEvalLimits as af, type PortableEvalOutcome as ag, type PortableEvalResult as ah, type PortableEvaluator as ai, type PortableJudgeRequest as aj, type PortableJudgeVerdict as ak, type PublishEvalDatasetOptions as al, type PublishEvalSuiteOptions as am, type PullEvalOptions as an, type PulledEvalSnapshot as ao, type RaindropConnection as ap, type RaindropFeatureFlagValue as aq, type RaindropFeatureFlags as ar, type RaindropProperties as as, type RaindropPropertyLeaf as at, type RaindropPropertyValue as au, type ReplayClient as av, type ReplayCounts as aw, ReplayCountsSchema as ax, type ReplayEvaluator as ay, type ReplayEvaluatorCounts as az, type Interaction as b, type SubagentDispatch as b0, type SubagentDispatchOptions as b1, type SubagentFinishOptions as b2, type ToolParams as b3, type ToolSpan as b4, type Trace as b5, type TraceContext as b6, TraceSchema as b7, type TrackToolParams as b8, type WorkflowParams as b9, replay as bA, resolveDisableBatching as bB, runEvalSuite as bC, runReplay as bD, traceOutput as bE, traceToolCalls as bF, traceTools as bG, verifyPortableEvalArtifact as bH, withEvalScope as bI, claimReplay as ba, compareEvalRuns as bb, createEvalSuiteRun as bc, createReplay as bd, currentEvalScope as be, defineDataset as bf, defineEvalSuite as bg, defineEvaluatorProgram as bh, defineLocalEvaluator as bi, evaluateEvalRun as bj, evaluatePortableEvaluator as bk, evaluateReplay as bl, importPortableEvaluator as bm, isEvalDataset as bn, isEvalSuiteDefinition as bo, isPublishedEvalSuite as bp, loadEvalSnapshot as bq, projectWorkshopTrace as br, publishEvalDataset as bs, publishEvalSuite as bt, pullEval as bu, readEvalDataset as bv, readEvalManifest as bw, readEvalRun as bx, readEvaluator as by, readReplay as bz, type AppGitSnapshot as c, type SubagentResumeOptions as d, type SubagentRun as e, type Tracer as f, type AiTrackEvent as g, type AppGitOptions as h, type Attachment as i, type AttachmentType as j, type BooleanThreshold as k, BooleanThresholdSchema as l, type CreateReplayOptions as m, DEFAULT_EVAL_WAIT_MS as n, DEFAULT_QUERY_URL as o, DEFAULT_TRACE_WAIT_MS as p, type DefinedEvalDataset as q, type EvalClient as r, type EvalDataset as s, type EvalDatasetInput as t, type EvalDatasetManifest as u, EvalDatasetPublishConflictError as v, type EvalDatasetRowInput as w, EvalDatasetRowInputSchema as x, type EvalDestination as y, type EvalEvaluator as z };
|
package/dist/index.d.mts
CHANGED
|
@@ -1,6 +1,5 @@
|
|
|
1
1
|
import '@opentelemetry/instrumentation';
|
|
2
|
-
export {
|
|
3
|
-
export { E as EVAL_CORRELATION_ID_ATTRIBUTE, a as EvalScope, c as currentEvalScope, w as withEvalScope } from './eval-scope-BYkhP4uI.js';
|
|
2
|
+
export { g as AiTrackEvent, A as AiTrackUsage, h as AppGitOptions, c as AppGitSnapshot, i as Attachment, j as AttachmentType, B as BeginInteractionOptions, k as BooleanThreshold, l as BooleanThresholdSchema, m as CreateReplayOptions, D as DEFAULT_EVAL_POLL_INTERVAL_MS, n as DEFAULT_EVAL_WAIT_MS, o as DEFAULT_QUERY_URL, p as DEFAULT_TRACE_WAIT_MS, q as DefinedEvalDataset, E as EVAL_CORRELATION_ID_ATTRIBUTE, r as EvalClient, s as EvalDataset, t as EvalDatasetInput, u as EvalDatasetManifest, v as EvalDatasetPublishConflictError, w as EvalDatasetRowInput, x as EvalDatasetRowInputSchema, y as EvalDestination, z as EvalEvaluator, F as EvalLocalEvaluator, G as EvalManifest, H as EvalProgram, J as EvalPublishError, K as EvalRowResult, L as EvalRowTracePair, M as EvalRunReading, N as EvalScope, O as EvalSuiteDefinition, Q as EvalSuiteDefinitionInput, U as EvalSuiteResult, V as EvalSuiteRunSession, W as EvalVerdictAssessment, X as EvaluateReplayOptions, Y as FinishInteractionOptions, Z as HostedEvaluatorPin, _ as IdentifyEvent, b as Interaction, $ as LocalBooleanVerdict, a0 as LocalBooleanVerdictSchema, a1 as LocalDebuggerLiveEventInput, a2 as LocalEvaluator, a3 as LocalEvaluatorContext, a4 as LocalNumberVerdict, a5 as LocalNumberVerdictSchema, a6 as LocalScoreVerdict, a7 as LocalScoreVerdictSchema, a8 as LocalVerdict, a9 as LocalVerdictSchema, aa as MAX_EVAL_STATUS_RETRIES, ab as MAX_INGEST_SIZE_BYTES, ac as NumericThreshold, ad as NumericThresholdSchema, P as PartialAiTrackEvent, ae as PortableEvalArtifact, af as PortableEvalLimits, ag as PortableEvalOutcome, ah as PortableEvalResult, ai as PortableEvaluator, aj as PortableJudgeRequest, ak as PortableJudgeVerdict, al as PublishEvalDatasetOptions, am as PublishEvalSuiteOptions, an as PullEvalOptions, ao as PulledEvalSnapshot, R as Raindrop, ap as RaindropConnection, aq as RaindropFeatureFlagValue, ar as RaindropFeatureFlags, as as RaindropProperties, at as RaindropPropertyLeaf, au as RaindropPropertyValue, a as RaindropSpanProcessor, av as ReplayClient, aw as ReplayCounts, ax as ReplayCountsSchema, ay as ReplayEvaluator, az as ReplayEvaluatorCounts, aA as ReplayEvaluatorCountsSchema, aB as ReplayEvaluatorStats, aC as ReplayEvaluatorStatsSchema, aD as ReplayEvaluatorSummary, aE as ReplayEvaluatorSummarySchema, aF as ReplayHandle, aG as ReplayLocalEvaluatorContext, aH as ReplayOptions, aI as ReplayReading, aJ as ReplayResult, aK as ReplayResultRow, aL as ReplayResultRowSchema, aM as ReplayRow, aN as ReplayRowOutcome, aO as ReplayVerdict, aP as ReplayVerdictSchema, aQ as RunEvalSuiteOptions, aR as RunReplayOptions, aS as SelfDiagnoseOptions, aT as SelfDiagnosticsExecuteContext, aU as SelfDiagnosticsExecuteResult, aV as SelfDiagnosticsSignalDefinition, aW as SelfDiagnosticsSignalDefinitions, aX as SelfDiagnosticsTool, aY as SelfDiagnosticsToolInputSchema, aZ as SelfDiagnosticsToolOptions, a_ as SignalEvent, a$ as SpanParams, b0 as SubagentDispatch, b1 as SubagentDispatchOptions, b2 as SubagentFinishOptions, d as SubagentResumeOptions, e as SubagentRun, b3 as ToolParams, b4 as ToolSpan, b5 as Trace, b6 as TraceContext, b7 as TraceSchema, f as Tracer, b8 as TrackToolParams, b9 as WorkflowParams, ba as claimReplay, bb as compareEvalRuns, bc as createEvalSuiteRun, bd as createReplay, be as currentEvalScope, R as default, bf as defineDataset, bg as defineEvalSuite, bh as defineEvaluatorProgram, bi as defineLocalEvaluator, bj as evaluateEvalRun, bk as evaluatePortableEvaluator, bl as evaluateReplay, bm as importPortableEvaluator, bn as isEvalDataset, bo as isEvalSuiteDefinition, bp as isPublishedEvalSuite, bq as loadEvalSnapshot, br as projectWorkshopTrace, bs as publishEvalDataset, bt as publishEvalSuite, bu as pullEval, bv as readEvalDataset, bw as readEvalManifest, bx as readEvalRun, by as readEvaluator, bz as readReplay, bA as replay, bB as resolveDisableBatching, bC as runEvalSuite, bD as runReplay, bE as traceOutput, bF as traceToolCalls, bG as traceTools, bH as verifyPortableEvalArtifact, bI as withEvalScope } from './index-DuJkG47V.mjs';
|
|
4
3
|
import '@opentelemetry/api';
|
|
5
4
|
import 'zod';
|
|
6
5
|
import '@opentelemetry/sdk-trace-base';
|
package/dist/index.d.ts
CHANGED
|
@@ -1,6 +1,5 @@
|
|
|
1
1
|
import '@opentelemetry/instrumentation';
|
|
2
|
-
export {
|
|
3
|
-
export { E as EVAL_CORRELATION_ID_ATTRIBUTE, a as EvalScope, c as currentEvalScope, w as withEvalScope } from './eval-scope-BYkhP4uI.js';
|
|
2
|
+
export { g as AiTrackEvent, A as AiTrackUsage, h as AppGitOptions, c as AppGitSnapshot, i as Attachment, j as AttachmentType, B as BeginInteractionOptions, k as BooleanThreshold, l as BooleanThresholdSchema, m as CreateReplayOptions, D as DEFAULT_EVAL_POLL_INTERVAL_MS, n as DEFAULT_EVAL_WAIT_MS, o as DEFAULT_QUERY_URL, p as DEFAULT_TRACE_WAIT_MS, q as DefinedEvalDataset, E as EVAL_CORRELATION_ID_ATTRIBUTE, r as EvalClient, s as EvalDataset, t as EvalDatasetInput, u as EvalDatasetManifest, v as EvalDatasetPublishConflictError, w as EvalDatasetRowInput, x as EvalDatasetRowInputSchema, y as EvalDestination, z as EvalEvaluator, F as EvalLocalEvaluator, G as EvalManifest, H as EvalProgram, J as EvalPublishError, K as EvalRowResult, L as EvalRowTracePair, M as EvalRunReading, N as EvalScope, O as EvalSuiteDefinition, Q as EvalSuiteDefinitionInput, U as EvalSuiteResult, V as EvalSuiteRunSession, W as EvalVerdictAssessment, X as EvaluateReplayOptions, Y as FinishInteractionOptions, Z as HostedEvaluatorPin, _ as IdentifyEvent, b as Interaction, $ as LocalBooleanVerdict, a0 as LocalBooleanVerdictSchema, a1 as LocalDebuggerLiveEventInput, a2 as LocalEvaluator, a3 as LocalEvaluatorContext, a4 as LocalNumberVerdict, a5 as LocalNumberVerdictSchema, a6 as LocalScoreVerdict, a7 as LocalScoreVerdictSchema, a8 as LocalVerdict, a9 as LocalVerdictSchema, aa as MAX_EVAL_STATUS_RETRIES, ab as MAX_INGEST_SIZE_BYTES, ac as NumericThreshold, ad as NumericThresholdSchema, P as PartialAiTrackEvent, ae as PortableEvalArtifact, af as PortableEvalLimits, ag as PortableEvalOutcome, ah as PortableEvalResult, ai as PortableEvaluator, aj as PortableJudgeRequest, ak as PortableJudgeVerdict, al as PublishEvalDatasetOptions, am as PublishEvalSuiteOptions, an as PullEvalOptions, ao as PulledEvalSnapshot, R as Raindrop, ap as RaindropConnection, aq as RaindropFeatureFlagValue, ar as RaindropFeatureFlags, as as RaindropProperties, at as RaindropPropertyLeaf, au as RaindropPropertyValue, a as RaindropSpanProcessor, av as ReplayClient, aw as ReplayCounts, ax as ReplayCountsSchema, ay as ReplayEvaluator, az as ReplayEvaluatorCounts, aA as ReplayEvaluatorCountsSchema, aB as ReplayEvaluatorStats, aC as ReplayEvaluatorStatsSchema, aD as ReplayEvaluatorSummary, aE as ReplayEvaluatorSummarySchema, aF as ReplayHandle, aG as ReplayLocalEvaluatorContext, aH as ReplayOptions, aI as ReplayReading, aJ as ReplayResult, aK as ReplayResultRow, aL as ReplayResultRowSchema, aM as ReplayRow, aN as ReplayRowOutcome, aO as ReplayVerdict, aP as ReplayVerdictSchema, aQ as RunEvalSuiteOptions, aR as RunReplayOptions, aS as SelfDiagnoseOptions, aT as SelfDiagnosticsExecuteContext, aU as SelfDiagnosticsExecuteResult, aV as SelfDiagnosticsSignalDefinition, aW as SelfDiagnosticsSignalDefinitions, aX as SelfDiagnosticsTool, aY as SelfDiagnosticsToolInputSchema, aZ as SelfDiagnosticsToolOptions, a_ as SignalEvent, a$ as SpanParams, b0 as SubagentDispatch, b1 as SubagentDispatchOptions, b2 as SubagentFinishOptions, d as SubagentResumeOptions, e as SubagentRun, b3 as ToolParams, b4 as ToolSpan, b5 as Trace, b6 as TraceContext, b7 as TraceSchema, f as Tracer, b8 as TrackToolParams, b9 as WorkflowParams, ba as claimReplay, bb as compareEvalRuns, bc as createEvalSuiteRun, bd as createReplay, be as currentEvalScope, R as default, bf as defineDataset, bg as defineEvalSuite, bh as defineEvaluatorProgram, bi as defineLocalEvaluator, bj as evaluateEvalRun, bk as evaluatePortableEvaluator, bl as evaluateReplay, bm as importPortableEvaluator, bn as isEvalDataset, bo as isEvalSuiteDefinition, bp as isPublishedEvalSuite, bq as loadEvalSnapshot, br as projectWorkshopTrace, bs as publishEvalDataset, bt as publishEvalSuite, bu as pullEval, bv as readEvalDataset, bw as readEvalManifest, bx as readEvalRun, by as readEvaluator, bz as readReplay, bA as replay, bB as resolveDisableBatching, bC as runEvalSuite, bD as runReplay, bE as traceOutput, bF as traceToolCalls, bG as traceTools, bH as verifyPortableEvalArtifact, bI as withEvalScope } from './index-DuJkG47V.js';
|
|
4
3
|
import '@opentelemetry/api';
|
|
5
4
|
import 'zod';
|
|
6
5
|
import '@opentelemetry/sdk-trace-base';
|