@artemiskit/core 0.2.4 → 0.4.0
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/CHANGELOG.md +137 -0
- package/README.md +4 -0
- package/dist/adapters/registry.d.ts.map +1 -1
- package/dist/adapters/types.d.ts +50 -2
- package/dist/adapters/types.d.ts.map +1 -1
- package/dist/agent-evaluation/index.d.ts +3 -0
- package/dist/agent-evaluation/index.d.ts.map +1 -0
- package/dist/agent-evaluation/scorer.d.ts +35 -0
- package/dist/agent-evaluation/scorer.d.ts.map +1 -0
- package/dist/agent-evaluation/types.d.ts +37 -0
- package/dist/agent-evaluation/types.d.ts.map +1 -0
- package/dist/artifacts/manifest.d.ts.map +1 -1
- package/dist/artifacts/types.d.ts +54 -0
- package/dist/artifacts/types.d.ts.map +1 -1
- package/dist/evaluators/index.d.ts +1 -0
- package/dist/evaluators/index.d.ts.map +1 -1
- package/dist/evaluators/json-schema.d.ts +0 -1
- package/dist/evaluators/json-schema.d.ts.map +1 -1
- package/dist/evaluators/llm-grader.d.ts +2 -0
- package/dist/evaluators/llm-grader.d.ts.map +1 -1
- package/dist/evaluators/tool-trace.d.ts +7 -0
- package/dist/evaluators/tool-trace.d.ts.map +1 -0
- package/dist/evaluators/types.d.ts +20 -0
- package/dist/evaluators/types.d.ts.map +1 -1
- package/dist/index.d.ts +2 -0
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +19964 -12482
- package/dist/runner/executor.d.ts.map +1 -1
- package/dist/runner/runner.d.ts.map +1 -1
- package/dist/runner/types.d.ts +4 -0
- package/dist/runner/types.d.ts.map +1 -1
- package/dist/scenario/schema.d.ts +789 -99
- package/dist/scenario/schema.d.ts.map +1 -1
- package/dist/storage/supabase.d.ts +26 -5
- package/dist/storage/supabase.d.ts.map +1 -1
- package/dist/storage/types.d.ts +167 -1
- package/dist/storage/types.d.ts.map +1 -1
- package/dist/tools/fixture-executor.d.ts +10 -0
- package/dist/tools/fixture-executor.d.ts.map +1 -0
- package/dist/tools/index.d.ts +4 -0
- package/dist/tools/index.d.ts.map +1 -0
- package/dist/tools/schema-validator.d.ts +10 -0
- package/dist/tools/schema-validator.d.ts.map +1 -0
- package/dist/tools/types.d.ts +50 -0
- package/dist/tools/types.d.ts.map +1 -0
- package/package.json +2 -1
- package/src/adapters/registry.ts +45 -0
- package/src/adapters/types.test.ts +21 -0
- package/src/adapters/types.ts +56 -0
- package/src/agent-evaluation/index.ts +2 -0
- package/src/agent-evaluation/scorer.test.ts +1194 -0
- package/src/agent-evaluation/scorer.ts +640 -0
- package/src/agent-evaluation/types.test.ts +27 -0
- package/src/agent-evaluation/types.ts +43 -0
- package/src/artifacts/manifest.test.ts +66 -19
- package/src/artifacts/manifest.ts +18 -5
- package/src/artifacts/types.ts +65 -0
- package/src/evaluators/index.ts +3 -0
- package/src/evaluators/json-schema.test.ts +130 -0
- package/src/evaluators/json-schema.ts +38 -63
- package/src/evaluators/llm-grader.test.ts +80 -0
- package/src/evaluators/llm-grader.ts +44 -6
- package/src/evaluators/tool-trace.test.ts +46 -0
- package/src/evaluators/tool-trace.ts +50 -0
- package/src/evaluators/types.ts +20 -0
- package/src/index.ts +6 -0
- package/src/runner/executor.test.ts +340 -0
- package/src/runner/executor.ts +289 -20
- package/src/runner/release-validation.test.ts +169 -0
- package/src/runner/runner.ts +5 -1
- package/src/runner/types.ts +4 -0
- package/src/scenario/schema.ts +66 -1
- package/src/storage/supabase.test.ts +1052 -0
- package/src/storage/supabase.ts +614 -5
- package/src/storage/types.ts +207 -1
- package/src/tools/fixture-executor.test.ts +88 -0
- package/src/tools/fixture-executor.ts +112 -0
- package/src/tools/index.ts +3 -0
- package/src/tools/schema-validator.test.ts +32 -0
- package/src/tools/schema-validator.ts +56 -0
- package/src/tools/types.ts +80 -0
|
@@ -5,7 +5,7 @@ import { z } from 'zod';
|
|
|
5
5
|
/**
|
|
6
6
|
* Provider schema - supports all providers
|
|
7
7
|
*/
|
|
8
|
-
export declare const ProviderSchema: z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>;
|
|
8
|
+
export declare const ProviderSchema: z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>;
|
|
9
9
|
/**
|
|
10
10
|
* Provider config schema - optional overrides for provider settings
|
|
11
11
|
* Supports ${ENV_VAR} and ${ENV_VAR:-default} syntax for values
|
|
@@ -24,7 +24,24 @@ export declare const ProviderConfigSchema: z.ZodOptional<z.ZodObject<{
|
|
|
24
24
|
embeddingDeploymentName: z.ZodOptional<z.ZodString>;
|
|
25
25
|
modelFamily: z.ZodOptional<z.ZodString>;
|
|
26
26
|
underlyingProvider: z.ZodOptional<z.ZodEnum<["openai", "azure", "anthropic", "google", "mistral"]>>;
|
|
27
|
+
name: z.ZodOptional<z.ZodString>;
|
|
28
|
+
runnableType: z.ZodOptional<z.ZodEnum<["chain", "agent", "llm", "runnable"]>>;
|
|
29
|
+
captureTraces: z.ZodOptional<z.ZodBoolean>;
|
|
30
|
+
captureMessages: z.ZodOptional<z.ZodBoolean>;
|
|
31
|
+
thinking: z.ZodOptional<z.ZodObject<{
|
|
32
|
+
type: z.ZodEnum<["enabled", "disabled"]>;
|
|
33
|
+
}, "strip", z.ZodTypeAny, {
|
|
34
|
+
type: "enabled" | "disabled";
|
|
35
|
+
}, {
|
|
36
|
+
type: "enabled" | "disabled";
|
|
37
|
+
}>>;
|
|
38
|
+
enableSearch: z.ZodOptional<z.ZodBoolean>;
|
|
39
|
+
searchOptions: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
27
40
|
}, "strip", z.ZodTypeAny, {
|
|
41
|
+
thinking?: {
|
|
42
|
+
type: "enabled" | "disabled";
|
|
43
|
+
} | undefined;
|
|
44
|
+
name?: string | undefined;
|
|
28
45
|
apiKey?: string | undefined;
|
|
29
46
|
baseUrl?: string | undefined;
|
|
30
47
|
defaultModel?: string | undefined;
|
|
@@ -37,7 +54,16 @@ export declare const ProviderConfigSchema: z.ZodOptional<z.ZodObject<{
|
|
|
37
54
|
embeddingDeploymentName?: string | undefined;
|
|
38
55
|
modelFamily?: string | undefined;
|
|
39
56
|
underlyingProvider?: "openai" | "anthropic" | "google" | "mistral" | "azure" | undefined;
|
|
57
|
+
runnableType?: "chain" | "agent" | "llm" | "runnable" | undefined;
|
|
58
|
+
captureTraces?: boolean | undefined;
|
|
59
|
+
captureMessages?: boolean | undefined;
|
|
60
|
+
enableSearch?: boolean | undefined;
|
|
61
|
+
searchOptions?: Record<string, unknown> | undefined;
|
|
40
62
|
}, {
|
|
63
|
+
thinking?: {
|
|
64
|
+
type: "enabled" | "disabled";
|
|
65
|
+
} | undefined;
|
|
66
|
+
name?: string | undefined;
|
|
41
67
|
apiKey?: string | undefined;
|
|
42
68
|
baseUrl?: string | undefined;
|
|
43
69
|
defaultModel?: string | undefined;
|
|
@@ -50,6 +76,11 @@ export declare const ProviderConfigSchema: z.ZodOptional<z.ZodObject<{
|
|
|
50
76
|
embeddingDeploymentName?: string | undefined;
|
|
51
77
|
modelFamily?: string | undefined;
|
|
52
78
|
underlyingProvider?: "openai" | "anthropic" | "google" | "mistral" | "azure" | undefined;
|
|
79
|
+
runnableType?: "chain" | "agent" | "llm" | "runnable" | undefined;
|
|
80
|
+
captureTraces?: boolean | undefined;
|
|
81
|
+
captureMessages?: boolean | undefined;
|
|
82
|
+
enableSearch?: boolean | undefined;
|
|
83
|
+
searchOptions?: Record<string, unknown> | undefined;
|
|
53
84
|
}>>;
|
|
54
85
|
/**
|
|
55
86
|
* Expected result types - how to evaluate responses
|
|
@@ -95,18 +126,22 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
95
126
|
type: z.ZodLiteral<"llm_grader">;
|
|
96
127
|
rubric: z.ZodString;
|
|
97
128
|
model: z.ZodOptional<z.ZodString>;
|
|
98
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
129
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
99
130
|
threshold: z.ZodDefault<z.ZodNumber>;
|
|
131
|
+
/** Require exact, validated JSON from the judge for assurance assessments. */
|
|
132
|
+
strict: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
100
133
|
}, "strip", z.ZodTypeAny, {
|
|
101
134
|
type: "llm_grader";
|
|
135
|
+
strict: boolean;
|
|
102
136
|
threshold: number;
|
|
103
137
|
rubric: string;
|
|
104
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
138
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
105
139
|
model?: string | undefined;
|
|
106
140
|
}, {
|
|
107
141
|
type: "llm_grader";
|
|
108
142
|
rubric: string;
|
|
109
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
143
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
144
|
+
strict?: boolean | undefined;
|
|
110
145
|
threshold?: number | undefined;
|
|
111
146
|
model?: string | undefined;
|
|
112
147
|
}>, z.ZodObject<{
|
|
@@ -154,6 +189,24 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
154
189
|
type: "custom";
|
|
155
190
|
evaluator: string;
|
|
156
191
|
config?: Record<string, unknown> | undefined;
|
|
192
|
+
}>, z.ZodObject<{
|
|
193
|
+
type: z.ZodLiteral<"tool_trace">;
|
|
194
|
+
requiredTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
195
|
+
forbiddenTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
196
|
+
ordered: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
197
|
+
maxCalls: z.ZodOptional<z.ZodNumber>;
|
|
198
|
+
}, "strip", z.ZodTypeAny, {
|
|
199
|
+
type: "tool_trace";
|
|
200
|
+
ordered: boolean;
|
|
201
|
+
requiredTools?: string[] | undefined;
|
|
202
|
+
forbiddenTools?: string[] | undefined;
|
|
203
|
+
maxCalls?: number | undefined;
|
|
204
|
+
}, {
|
|
205
|
+
type: "tool_trace";
|
|
206
|
+
requiredTools?: string[] | undefined;
|
|
207
|
+
forbiddenTools?: string[] | undefined;
|
|
208
|
+
ordered?: boolean | undefined;
|
|
209
|
+
maxCalls?: number | undefined;
|
|
157
210
|
}>, z.ZodObject<{
|
|
158
211
|
type: z.ZodLiteral<"similarity">;
|
|
159
212
|
value: z.ZodString;
|
|
@@ -169,14 +222,14 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
169
222
|
type: "similarity";
|
|
170
223
|
threshold: number;
|
|
171
224
|
model?: string | undefined;
|
|
172
|
-
mode?: "
|
|
225
|
+
mode?: "llm" | "embedding" | undefined;
|
|
173
226
|
embeddingModel?: string | undefined;
|
|
174
227
|
}, {
|
|
175
228
|
value: string;
|
|
176
229
|
type: "similarity";
|
|
177
230
|
threshold?: number | undefined;
|
|
178
231
|
model?: string | undefined;
|
|
179
|
-
mode?: "
|
|
232
|
+
mode?: "llm" | "embedding" | undefined;
|
|
180
233
|
embeddingModel?: string | undefined;
|
|
181
234
|
}>, z.ZodObject<{
|
|
182
235
|
type: z.ZodLiteral<"inline">;
|
|
@@ -233,18 +286,22 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
233
286
|
type: z.ZodLiteral<"llm_grader">;
|
|
234
287
|
rubric: z.ZodString;
|
|
235
288
|
model: z.ZodOptional<z.ZodString>;
|
|
236
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
289
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
237
290
|
threshold: z.ZodDefault<z.ZodNumber>;
|
|
291
|
+
/** Require exact, validated JSON from the judge for assurance assessments. */
|
|
292
|
+
strict: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
238
293
|
}, "strip", z.ZodTypeAny, {
|
|
239
294
|
type: "llm_grader";
|
|
295
|
+
strict: boolean;
|
|
240
296
|
threshold: number;
|
|
241
297
|
rubric: string;
|
|
242
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
298
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
243
299
|
model?: string | undefined;
|
|
244
300
|
}, {
|
|
245
301
|
type: "llm_grader";
|
|
246
302
|
rubric: string;
|
|
247
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
303
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
304
|
+
strict?: boolean | undefined;
|
|
248
305
|
threshold?: number | undefined;
|
|
249
306
|
model?: string | undefined;
|
|
250
307
|
}>, z.ZodObject<{
|
|
@@ -292,6 +349,24 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
292
349
|
type: "custom";
|
|
293
350
|
evaluator: string;
|
|
294
351
|
config?: Record<string, unknown> | undefined;
|
|
352
|
+
}>, z.ZodObject<{
|
|
353
|
+
type: z.ZodLiteral<"tool_trace">;
|
|
354
|
+
requiredTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
355
|
+
forbiddenTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
356
|
+
ordered: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
357
|
+
maxCalls: z.ZodOptional<z.ZodNumber>;
|
|
358
|
+
}, "strip", z.ZodTypeAny, {
|
|
359
|
+
type: "tool_trace";
|
|
360
|
+
ordered: boolean;
|
|
361
|
+
requiredTools?: string[] | undefined;
|
|
362
|
+
forbiddenTools?: string[] | undefined;
|
|
363
|
+
maxCalls?: number | undefined;
|
|
364
|
+
}, {
|
|
365
|
+
type: "tool_trace";
|
|
366
|
+
requiredTools?: string[] | undefined;
|
|
367
|
+
forbiddenTools?: string[] | undefined;
|
|
368
|
+
ordered?: boolean | undefined;
|
|
369
|
+
maxCalls?: number | undefined;
|
|
295
370
|
}>, z.ZodObject<{
|
|
296
371
|
type: z.ZodLiteral<"similarity">;
|
|
297
372
|
value: z.ZodString;
|
|
@@ -307,14 +382,14 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
307
382
|
type: "similarity";
|
|
308
383
|
threshold: number;
|
|
309
384
|
model?: string | undefined;
|
|
310
|
-
mode?: "
|
|
385
|
+
mode?: "llm" | "embedding" | undefined;
|
|
311
386
|
embeddingModel?: string | undefined;
|
|
312
387
|
}, {
|
|
313
388
|
value: string;
|
|
314
389
|
type: "similarity";
|
|
315
390
|
threshold?: number | undefined;
|
|
316
391
|
model?: string | undefined;
|
|
317
|
-
mode?: "
|
|
392
|
+
mode?: "llm" | "embedding" | undefined;
|
|
318
393
|
embeddingModel?: string | undefined;
|
|
319
394
|
}>, z.ZodObject<{
|
|
320
395
|
type: z.ZodLiteral<"inline">;
|
|
@@ -346,9 +421,10 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
346
421
|
threshold: number;
|
|
347
422
|
} | {
|
|
348
423
|
type: "llm_grader";
|
|
424
|
+
strict: boolean;
|
|
349
425
|
threshold: number;
|
|
350
426
|
rubric: string;
|
|
351
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
427
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
352
428
|
model?: string | undefined;
|
|
353
429
|
} | {
|
|
354
430
|
values: string[];
|
|
@@ -365,12 +441,18 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
365
441
|
type: "custom";
|
|
366
442
|
evaluator: string;
|
|
367
443
|
config?: Record<string, unknown> | undefined;
|
|
444
|
+
} | {
|
|
445
|
+
type: "tool_trace";
|
|
446
|
+
ordered: boolean;
|
|
447
|
+
requiredTools?: string[] | undefined;
|
|
448
|
+
forbiddenTools?: string[] | undefined;
|
|
449
|
+
maxCalls?: number | undefined;
|
|
368
450
|
} | {
|
|
369
451
|
value: string;
|
|
370
452
|
type: "similarity";
|
|
371
453
|
threshold: number;
|
|
372
454
|
model?: string | undefined;
|
|
373
|
-
mode?: "
|
|
455
|
+
mode?: "llm" | "embedding" | undefined;
|
|
374
456
|
embeddingModel?: string | undefined;
|
|
375
457
|
} | {
|
|
376
458
|
type: "inline";
|
|
@@ -395,7 +477,8 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
395
477
|
} | {
|
|
396
478
|
type: "llm_grader";
|
|
397
479
|
rubric: string;
|
|
398
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
480
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
481
|
+
strict?: boolean | undefined;
|
|
399
482
|
threshold?: number | undefined;
|
|
400
483
|
model?: string | undefined;
|
|
401
484
|
} | {
|
|
@@ -413,12 +496,18 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
413
496
|
type: "custom";
|
|
414
497
|
evaluator: string;
|
|
415
498
|
config?: Record<string, unknown> | undefined;
|
|
499
|
+
} | {
|
|
500
|
+
type: "tool_trace";
|
|
501
|
+
requiredTools?: string[] | undefined;
|
|
502
|
+
forbiddenTools?: string[] | undefined;
|
|
503
|
+
ordered?: boolean | undefined;
|
|
504
|
+
maxCalls?: number | undefined;
|
|
416
505
|
} | {
|
|
417
506
|
value: string;
|
|
418
507
|
type: "similarity";
|
|
419
508
|
threshold?: number | undefined;
|
|
420
509
|
model?: string | undefined;
|
|
421
|
-
mode?: "
|
|
510
|
+
mode?: "llm" | "embedding" | undefined;
|
|
422
511
|
embeddingModel?: string | undefined;
|
|
423
512
|
} | {
|
|
424
513
|
type: "inline";
|
|
@@ -430,14 +519,64 @@ export declare const ExpectedSchema: z.ZodUnion<[z.ZodDiscriminatedUnion<"type",
|
|
|
430
519
|
* Chat message schema
|
|
431
520
|
*/
|
|
432
521
|
export declare const ChatMessageSchema: z.ZodObject<{
|
|
433
|
-
role: z.ZodEnum<["system", "user", "assistant"]>;
|
|
522
|
+
role: z.ZodEnum<["system", "user", "assistant", "tool"]>;
|
|
434
523
|
content: z.ZodString;
|
|
524
|
+
name: z.ZodOptional<z.ZodString>;
|
|
525
|
+
toolCallId: z.ZodOptional<z.ZodString>;
|
|
526
|
+
tool_calls: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
527
|
+
id: z.ZodString;
|
|
528
|
+
type: z.ZodLiteral<"function">;
|
|
529
|
+
function: z.ZodObject<{
|
|
530
|
+
name: z.ZodString;
|
|
531
|
+
arguments: z.ZodString;
|
|
532
|
+
}, "strip", z.ZodTypeAny, {
|
|
533
|
+
name: string;
|
|
534
|
+
arguments: string;
|
|
535
|
+
}, {
|
|
536
|
+
name: string;
|
|
537
|
+
arguments: string;
|
|
538
|
+
}>;
|
|
539
|
+
}, "strip", z.ZodTypeAny, {
|
|
540
|
+
function: {
|
|
541
|
+
name: string;
|
|
542
|
+
arguments: string;
|
|
543
|
+
};
|
|
544
|
+
type: "function";
|
|
545
|
+
id: string;
|
|
546
|
+
}, {
|
|
547
|
+
function: {
|
|
548
|
+
name: string;
|
|
549
|
+
arguments: string;
|
|
550
|
+
};
|
|
551
|
+
type: "function";
|
|
552
|
+
id: string;
|
|
553
|
+
}>, "many">>;
|
|
435
554
|
}, "strip", z.ZodTypeAny, {
|
|
436
|
-
role: "system" | "user" | "assistant";
|
|
555
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
437
556
|
content: string;
|
|
557
|
+
tool_calls?: {
|
|
558
|
+
function: {
|
|
559
|
+
name: string;
|
|
560
|
+
arguments: string;
|
|
561
|
+
};
|
|
562
|
+
type: "function";
|
|
563
|
+
id: string;
|
|
564
|
+
}[] | undefined;
|
|
565
|
+
name?: string | undefined;
|
|
566
|
+
toolCallId?: string | undefined;
|
|
438
567
|
}, {
|
|
439
|
-
role: "system" | "user" | "assistant";
|
|
568
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
440
569
|
content: string;
|
|
570
|
+
tool_calls?: {
|
|
571
|
+
function: {
|
|
572
|
+
name: string;
|
|
573
|
+
arguments: string;
|
|
574
|
+
};
|
|
575
|
+
type: "function";
|
|
576
|
+
id: string;
|
|
577
|
+
}[] | undefined;
|
|
578
|
+
name?: string | undefined;
|
|
579
|
+
toolCallId?: string | undefined;
|
|
441
580
|
}>;
|
|
442
581
|
/**
|
|
443
582
|
* Variables schema - key-value pairs for template substitution
|
|
@@ -477,14 +616,64 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
477
616
|
name: z.ZodOptional<z.ZodString>;
|
|
478
617
|
description: z.ZodOptional<z.ZodString>;
|
|
479
618
|
prompt: z.ZodUnion<[z.ZodString, z.ZodArray<z.ZodObject<{
|
|
480
|
-
role: z.ZodEnum<["system", "user", "assistant"]>;
|
|
619
|
+
role: z.ZodEnum<["system", "user", "assistant", "tool"]>;
|
|
481
620
|
content: z.ZodString;
|
|
621
|
+
name: z.ZodOptional<z.ZodString>;
|
|
622
|
+
toolCallId: z.ZodOptional<z.ZodString>;
|
|
623
|
+
tool_calls: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
624
|
+
id: z.ZodString;
|
|
625
|
+
type: z.ZodLiteral<"function">;
|
|
626
|
+
function: z.ZodObject<{
|
|
627
|
+
name: z.ZodString;
|
|
628
|
+
arguments: z.ZodString;
|
|
629
|
+
}, "strip", z.ZodTypeAny, {
|
|
630
|
+
name: string;
|
|
631
|
+
arguments: string;
|
|
632
|
+
}, {
|
|
633
|
+
name: string;
|
|
634
|
+
arguments: string;
|
|
635
|
+
}>;
|
|
636
|
+
}, "strip", z.ZodTypeAny, {
|
|
637
|
+
function: {
|
|
638
|
+
name: string;
|
|
639
|
+
arguments: string;
|
|
640
|
+
};
|
|
641
|
+
type: "function";
|
|
642
|
+
id: string;
|
|
643
|
+
}, {
|
|
644
|
+
function: {
|
|
645
|
+
name: string;
|
|
646
|
+
arguments: string;
|
|
647
|
+
};
|
|
648
|
+
type: "function";
|
|
649
|
+
id: string;
|
|
650
|
+
}>, "many">>;
|
|
482
651
|
}, "strip", z.ZodTypeAny, {
|
|
483
|
-
role: "system" | "user" | "assistant";
|
|
652
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
484
653
|
content: string;
|
|
654
|
+
tool_calls?: {
|
|
655
|
+
function: {
|
|
656
|
+
name: string;
|
|
657
|
+
arguments: string;
|
|
658
|
+
};
|
|
659
|
+
type: "function";
|
|
660
|
+
id: string;
|
|
661
|
+
}[] | undefined;
|
|
662
|
+
name?: string | undefined;
|
|
663
|
+
toolCallId?: string | undefined;
|
|
485
664
|
}, {
|
|
486
|
-
role: "system" | "user" | "assistant";
|
|
665
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
487
666
|
content: string;
|
|
667
|
+
tool_calls?: {
|
|
668
|
+
function: {
|
|
669
|
+
name: string;
|
|
670
|
+
arguments: string;
|
|
671
|
+
};
|
|
672
|
+
type: "function";
|
|
673
|
+
id: string;
|
|
674
|
+
}[] | undefined;
|
|
675
|
+
name?: string | undefined;
|
|
676
|
+
toolCallId?: string | undefined;
|
|
488
677
|
}>, "many">]>;
|
|
489
678
|
expected: z.ZodUnion<[z.ZodDiscriminatedUnion<"type", [z.ZodObject<{
|
|
490
679
|
type: z.ZodLiteral<"exact">;
|
|
@@ -526,18 +715,22 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
526
715
|
type: z.ZodLiteral<"llm_grader">;
|
|
527
716
|
rubric: z.ZodString;
|
|
528
717
|
model: z.ZodOptional<z.ZodString>;
|
|
529
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
718
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
530
719
|
threshold: z.ZodDefault<z.ZodNumber>;
|
|
720
|
+
/** Require exact, validated JSON from the judge for assurance assessments. */
|
|
721
|
+
strict: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
531
722
|
}, "strip", z.ZodTypeAny, {
|
|
532
723
|
type: "llm_grader";
|
|
724
|
+
strict: boolean;
|
|
533
725
|
threshold: number;
|
|
534
726
|
rubric: string;
|
|
535
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
727
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
536
728
|
model?: string | undefined;
|
|
537
729
|
}, {
|
|
538
730
|
type: "llm_grader";
|
|
539
731
|
rubric: string;
|
|
540
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
732
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
733
|
+
strict?: boolean | undefined;
|
|
541
734
|
threshold?: number | undefined;
|
|
542
735
|
model?: string | undefined;
|
|
543
736
|
}>, z.ZodObject<{
|
|
@@ -585,6 +778,24 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
585
778
|
type: "custom";
|
|
586
779
|
evaluator: string;
|
|
587
780
|
config?: Record<string, unknown> | undefined;
|
|
781
|
+
}>, z.ZodObject<{
|
|
782
|
+
type: z.ZodLiteral<"tool_trace">;
|
|
783
|
+
requiredTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
784
|
+
forbiddenTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
785
|
+
ordered: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
786
|
+
maxCalls: z.ZodOptional<z.ZodNumber>;
|
|
787
|
+
}, "strip", z.ZodTypeAny, {
|
|
788
|
+
type: "tool_trace";
|
|
789
|
+
ordered: boolean;
|
|
790
|
+
requiredTools?: string[] | undefined;
|
|
791
|
+
forbiddenTools?: string[] | undefined;
|
|
792
|
+
maxCalls?: number | undefined;
|
|
793
|
+
}, {
|
|
794
|
+
type: "tool_trace";
|
|
795
|
+
requiredTools?: string[] | undefined;
|
|
796
|
+
forbiddenTools?: string[] | undefined;
|
|
797
|
+
ordered?: boolean | undefined;
|
|
798
|
+
maxCalls?: number | undefined;
|
|
588
799
|
}>, z.ZodObject<{
|
|
589
800
|
type: z.ZodLiteral<"similarity">;
|
|
590
801
|
value: z.ZodString;
|
|
@@ -600,14 +811,14 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
600
811
|
type: "similarity";
|
|
601
812
|
threshold: number;
|
|
602
813
|
model?: string | undefined;
|
|
603
|
-
mode?: "
|
|
814
|
+
mode?: "llm" | "embedding" | undefined;
|
|
604
815
|
embeddingModel?: string | undefined;
|
|
605
816
|
}, {
|
|
606
817
|
value: string;
|
|
607
818
|
type: "similarity";
|
|
608
819
|
threshold?: number | undefined;
|
|
609
820
|
model?: string | undefined;
|
|
610
|
-
mode?: "
|
|
821
|
+
mode?: "llm" | "embedding" | undefined;
|
|
611
822
|
embeddingModel?: string | undefined;
|
|
612
823
|
}>, z.ZodObject<{
|
|
613
824
|
type: z.ZodLiteral<"inline">;
|
|
@@ -664,18 +875,22 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
664
875
|
type: z.ZodLiteral<"llm_grader">;
|
|
665
876
|
rubric: z.ZodString;
|
|
666
877
|
model: z.ZodOptional<z.ZodString>;
|
|
667
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
878
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
668
879
|
threshold: z.ZodDefault<z.ZodNumber>;
|
|
880
|
+
/** Require exact, validated JSON from the judge for assurance assessments. */
|
|
881
|
+
strict: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
669
882
|
}, "strip", z.ZodTypeAny, {
|
|
670
883
|
type: "llm_grader";
|
|
884
|
+
strict: boolean;
|
|
671
885
|
threshold: number;
|
|
672
886
|
rubric: string;
|
|
673
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
887
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
674
888
|
model?: string | undefined;
|
|
675
889
|
}, {
|
|
676
890
|
type: "llm_grader";
|
|
677
891
|
rubric: string;
|
|
678
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
892
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
893
|
+
strict?: boolean | undefined;
|
|
679
894
|
threshold?: number | undefined;
|
|
680
895
|
model?: string | undefined;
|
|
681
896
|
}>, z.ZodObject<{
|
|
@@ -723,6 +938,24 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
723
938
|
type: "custom";
|
|
724
939
|
evaluator: string;
|
|
725
940
|
config?: Record<string, unknown> | undefined;
|
|
941
|
+
}>, z.ZodObject<{
|
|
942
|
+
type: z.ZodLiteral<"tool_trace">;
|
|
943
|
+
requiredTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
944
|
+
forbiddenTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
945
|
+
ordered: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
946
|
+
maxCalls: z.ZodOptional<z.ZodNumber>;
|
|
947
|
+
}, "strip", z.ZodTypeAny, {
|
|
948
|
+
type: "tool_trace";
|
|
949
|
+
ordered: boolean;
|
|
950
|
+
requiredTools?: string[] | undefined;
|
|
951
|
+
forbiddenTools?: string[] | undefined;
|
|
952
|
+
maxCalls?: number | undefined;
|
|
953
|
+
}, {
|
|
954
|
+
type: "tool_trace";
|
|
955
|
+
requiredTools?: string[] | undefined;
|
|
956
|
+
forbiddenTools?: string[] | undefined;
|
|
957
|
+
ordered?: boolean | undefined;
|
|
958
|
+
maxCalls?: number | undefined;
|
|
726
959
|
}>, z.ZodObject<{
|
|
727
960
|
type: z.ZodLiteral<"similarity">;
|
|
728
961
|
value: z.ZodString;
|
|
@@ -738,14 +971,14 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
738
971
|
type: "similarity";
|
|
739
972
|
threshold: number;
|
|
740
973
|
model?: string | undefined;
|
|
741
|
-
mode?: "
|
|
974
|
+
mode?: "llm" | "embedding" | undefined;
|
|
742
975
|
embeddingModel?: string | undefined;
|
|
743
976
|
}, {
|
|
744
977
|
value: string;
|
|
745
978
|
type: "similarity";
|
|
746
979
|
threshold?: number | undefined;
|
|
747
980
|
model?: string | undefined;
|
|
748
|
-
mode?: "
|
|
981
|
+
mode?: "llm" | "embedding" | undefined;
|
|
749
982
|
embeddingModel?: string | undefined;
|
|
750
983
|
}>, z.ZodObject<{
|
|
751
984
|
type: z.ZodLiteral<"inline">;
|
|
@@ -777,9 +1010,10 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
777
1010
|
threshold: number;
|
|
778
1011
|
} | {
|
|
779
1012
|
type: "llm_grader";
|
|
1013
|
+
strict: boolean;
|
|
780
1014
|
threshold: number;
|
|
781
1015
|
rubric: string;
|
|
782
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1016
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
783
1017
|
model?: string | undefined;
|
|
784
1018
|
} | {
|
|
785
1019
|
values: string[];
|
|
@@ -796,12 +1030,18 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
796
1030
|
type: "custom";
|
|
797
1031
|
evaluator: string;
|
|
798
1032
|
config?: Record<string, unknown> | undefined;
|
|
1033
|
+
} | {
|
|
1034
|
+
type: "tool_trace";
|
|
1035
|
+
ordered: boolean;
|
|
1036
|
+
requiredTools?: string[] | undefined;
|
|
1037
|
+
forbiddenTools?: string[] | undefined;
|
|
1038
|
+
maxCalls?: number | undefined;
|
|
799
1039
|
} | {
|
|
800
1040
|
value: string;
|
|
801
1041
|
type: "similarity";
|
|
802
1042
|
threshold: number;
|
|
803
1043
|
model?: string | undefined;
|
|
804
|
-
mode?: "
|
|
1044
|
+
mode?: "llm" | "embedding" | undefined;
|
|
805
1045
|
embeddingModel?: string | undefined;
|
|
806
1046
|
} | {
|
|
807
1047
|
type: "inline";
|
|
@@ -826,7 +1066,8 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
826
1066
|
} | {
|
|
827
1067
|
type: "llm_grader";
|
|
828
1068
|
rubric: string;
|
|
829
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1069
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1070
|
+
strict?: boolean | undefined;
|
|
830
1071
|
threshold?: number | undefined;
|
|
831
1072
|
model?: string | undefined;
|
|
832
1073
|
} | {
|
|
@@ -844,12 +1085,18 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
844
1085
|
type: "custom";
|
|
845
1086
|
evaluator: string;
|
|
846
1087
|
config?: Record<string, unknown> | undefined;
|
|
1088
|
+
} | {
|
|
1089
|
+
type: "tool_trace";
|
|
1090
|
+
requiredTools?: string[] | undefined;
|
|
1091
|
+
forbiddenTools?: string[] | undefined;
|
|
1092
|
+
ordered?: boolean | undefined;
|
|
1093
|
+
maxCalls?: number | undefined;
|
|
847
1094
|
} | {
|
|
848
1095
|
value: string;
|
|
849
1096
|
type: "similarity";
|
|
850
1097
|
threshold?: number | undefined;
|
|
851
1098
|
model?: string | undefined;
|
|
852
|
-
mode?: "
|
|
1099
|
+
mode?: "llm" | "embedding" | undefined;
|
|
853
1100
|
embeddingModel?: string | undefined;
|
|
854
1101
|
} | {
|
|
855
1102
|
type: "inline";
|
|
@@ -861,7 +1108,7 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
861
1108
|
metadata: z.ZodDefault<z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>>;
|
|
862
1109
|
timeout: z.ZodOptional<z.ZodNumber>;
|
|
863
1110
|
retries: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
864
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
1111
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
865
1112
|
model: z.ZodOptional<z.ZodString>;
|
|
866
1113
|
variables: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnion<[z.ZodString, z.ZodNumber, z.ZodBoolean]>>>;
|
|
867
1114
|
/** Case-level redaction config (overrides scenario-level) */
|
|
@@ -888,6 +1135,7 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
888
1135
|
replacement?: string | undefined;
|
|
889
1136
|
}>>;
|
|
890
1137
|
}, "strip", z.ZodTypeAny, {
|
|
1138
|
+
metadata: Record<string, unknown>;
|
|
891
1139
|
expected: {
|
|
892
1140
|
value: string;
|
|
893
1141
|
type: "exact";
|
|
@@ -902,9 +1150,10 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
902
1150
|
threshold: number;
|
|
903
1151
|
} | {
|
|
904
1152
|
type: "llm_grader";
|
|
1153
|
+
strict: boolean;
|
|
905
1154
|
threshold: number;
|
|
906
1155
|
rubric: string;
|
|
907
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1156
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
908
1157
|
model?: string | undefined;
|
|
909
1158
|
} | {
|
|
910
1159
|
values: string[];
|
|
@@ -921,12 +1170,18 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
921
1170
|
type: "custom";
|
|
922
1171
|
evaluator: string;
|
|
923
1172
|
config?: Record<string, unknown> | undefined;
|
|
1173
|
+
} | {
|
|
1174
|
+
type: "tool_trace";
|
|
1175
|
+
ordered: boolean;
|
|
1176
|
+
requiredTools?: string[] | undefined;
|
|
1177
|
+
forbiddenTools?: string[] | undefined;
|
|
1178
|
+
maxCalls?: number | undefined;
|
|
924
1179
|
} | {
|
|
925
1180
|
value: string;
|
|
926
1181
|
type: "similarity";
|
|
927
1182
|
threshold: number;
|
|
928
1183
|
model?: string | undefined;
|
|
929
|
-
mode?: "
|
|
1184
|
+
mode?: "llm" | "embedding" | undefined;
|
|
930
1185
|
embeddingModel?: string | undefined;
|
|
931
1186
|
} | {
|
|
932
1187
|
type: "inline";
|
|
@@ -949,9 +1204,10 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
949
1204
|
threshold: number;
|
|
950
1205
|
} | {
|
|
951
1206
|
type: "llm_grader";
|
|
1207
|
+
strict: boolean;
|
|
952
1208
|
threshold: number;
|
|
953
1209
|
rubric: string;
|
|
954
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1210
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
955
1211
|
model?: string | undefined;
|
|
956
1212
|
} | {
|
|
957
1213
|
values: string[];
|
|
@@ -968,12 +1224,18 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
968
1224
|
type: "custom";
|
|
969
1225
|
evaluator: string;
|
|
970
1226
|
config?: Record<string, unknown> | undefined;
|
|
1227
|
+
} | {
|
|
1228
|
+
type: "tool_trace";
|
|
1229
|
+
ordered: boolean;
|
|
1230
|
+
requiredTools?: string[] | undefined;
|
|
1231
|
+
forbiddenTools?: string[] | undefined;
|
|
1232
|
+
maxCalls?: number | undefined;
|
|
971
1233
|
} | {
|
|
972
1234
|
value: string;
|
|
973
1235
|
type: "similarity";
|
|
974
1236
|
threshold: number;
|
|
975
1237
|
model?: string | undefined;
|
|
976
|
-
mode?: "
|
|
1238
|
+
mode?: "llm" | "embedding" | undefined;
|
|
977
1239
|
embeddingModel?: string | undefined;
|
|
978
1240
|
} | {
|
|
979
1241
|
type: "inline";
|
|
@@ -983,14 +1245,23 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
983
1245
|
};
|
|
984
1246
|
id: string;
|
|
985
1247
|
prompt: string | {
|
|
986
|
-
role: "system" | "user" | "assistant";
|
|
1248
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
987
1249
|
content: string;
|
|
1250
|
+
tool_calls?: {
|
|
1251
|
+
function: {
|
|
1252
|
+
name: string;
|
|
1253
|
+
arguments: string;
|
|
1254
|
+
};
|
|
1255
|
+
type: "function";
|
|
1256
|
+
id: string;
|
|
1257
|
+
}[] | undefined;
|
|
1258
|
+
name?: string | undefined;
|
|
1259
|
+
toolCallId?: string | undefined;
|
|
988
1260
|
}[];
|
|
989
1261
|
tags: string[];
|
|
990
|
-
metadata: Record<string, unknown>;
|
|
991
1262
|
retries: number;
|
|
992
1263
|
name?: string | undefined;
|
|
993
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1264
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
994
1265
|
timeout?: number | undefined;
|
|
995
1266
|
model?: string | undefined;
|
|
996
1267
|
description?: string | undefined;
|
|
@@ -1019,7 +1290,8 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
1019
1290
|
} | {
|
|
1020
1291
|
type: "llm_grader";
|
|
1021
1292
|
rubric: string;
|
|
1022
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1293
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1294
|
+
strict?: boolean | undefined;
|
|
1023
1295
|
threshold?: number | undefined;
|
|
1024
1296
|
model?: string | undefined;
|
|
1025
1297
|
} | {
|
|
@@ -1037,12 +1309,18 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
1037
1309
|
type: "custom";
|
|
1038
1310
|
evaluator: string;
|
|
1039
1311
|
config?: Record<string, unknown> | undefined;
|
|
1312
|
+
} | {
|
|
1313
|
+
type: "tool_trace";
|
|
1314
|
+
requiredTools?: string[] | undefined;
|
|
1315
|
+
forbiddenTools?: string[] | undefined;
|
|
1316
|
+
ordered?: boolean | undefined;
|
|
1317
|
+
maxCalls?: number | undefined;
|
|
1040
1318
|
} | {
|
|
1041
1319
|
value: string;
|
|
1042
1320
|
type: "similarity";
|
|
1043
1321
|
threshold?: number | undefined;
|
|
1044
1322
|
model?: string | undefined;
|
|
1045
|
-
mode?: "
|
|
1323
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1046
1324
|
embeddingModel?: string | undefined;
|
|
1047
1325
|
} | {
|
|
1048
1326
|
type: "inline";
|
|
@@ -1066,7 +1344,8 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
1066
1344
|
} | {
|
|
1067
1345
|
type: "llm_grader";
|
|
1068
1346
|
rubric: string;
|
|
1069
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1347
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1348
|
+
strict?: boolean | undefined;
|
|
1070
1349
|
threshold?: number | undefined;
|
|
1071
1350
|
model?: string | undefined;
|
|
1072
1351
|
} | {
|
|
@@ -1084,12 +1363,18 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
1084
1363
|
type: "custom";
|
|
1085
1364
|
evaluator: string;
|
|
1086
1365
|
config?: Record<string, unknown> | undefined;
|
|
1366
|
+
} | {
|
|
1367
|
+
type: "tool_trace";
|
|
1368
|
+
requiredTools?: string[] | undefined;
|
|
1369
|
+
forbiddenTools?: string[] | undefined;
|
|
1370
|
+
ordered?: boolean | undefined;
|
|
1371
|
+
maxCalls?: number | undefined;
|
|
1087
1372
|
} | {
|
|
1088
1373
|
value: string;
|
|
1089
1374
|
type: "similarity";
|
|
1090
1375
|
threshold?: number | undefined;
|
|
1091
1376
|
model?: string | undefined;
|
|
1092
|
-
mode?: "
|
|
1377
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1093
1378
|
embeddingModel?: string | undefined;
|
|
1094
1379
|
} | {
|
|
1095
1380
|
type: "inline";
|
|
@@ -1099,16 +1384,26 @@ export declare const TestCaseSchema: z.ZodObject<{
|
|
|
1099
1384
|
};
|
|
1100
1385
|
id: string;
|
|
1101
1386
|
prompt: string | {
|
|
1102
|
-
role: "system" | "user" | "assistant";
|
|
1387
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
1103
1388
|
content: string;
|
|
1389
|
+
tool_calls?: {
|
|
1390
|
+
function: {
|
|
1391
|
+
name: string;
|
|
1392
|
+
arguments: string;
|
|
1393
|
+
};
|
|
1394
|
+
type: "function";
|
|
1395
|
+
id: string;
|
|
1396
|
+
}[] | undefined;
|
|
1397
|
+
name?: string | undefined;
|
|
1398
|
+
toolCallId?: string | undefined;
|
|
1104
1399
|
}[];
|
|
1105
1400
|
name?: string | undefined;
|
|
1106
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1401
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1402
|
+
metadata?: Record<string, unknown> | undefined;
|
|
1107
1403
|
timeout?: number | undefined;
|
|
1108
1404
|
model?: string | undefined;
|
|
1109
1405
|
description?: string | undefined;
|
|
1110
1406
|
tags?: string[] | undefined;
|
|
1111
|
-
metadata?: Record<string, unknown> | undefined;
|
|
1112
1407
|
retries?: number | undefined;
|
|
1113
1408
|
variables?: Record<string, string | number | boolean> | undefined;
|
|
1114
1409
|
redaction?: {
|
|
@@ -1127,7 +1422,7 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1127
1422
|
name: z.ZodString;
|
|
1128
1423
|
description: z.ZodOptional<z.ZodString>;
|
|
1129
1424
|
version: z.ZodDefault<z.ZodOptional<z.ZodString>>;
|
|
1130
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
1425
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
1131
1426
|
model: z.ZodOptional<z.ZodString>;
|
|
1132
1427
|
providerConfig: z.ZodOptional<z.ZodObject<{
|
|
1133
1428
|
apiKey: z.ZodOptional<z.ZodString>;
|
|
@@ -1142,7 +1437,24 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1142
1437
|
embeddingDeploymentName: z.ZodOptional<z.ZodString>;
|
|
1143
1438
|
modelFamily: z.ZodOptional<z.ZodString>;
|
|
1144
1439
|
underlyingProvider: z.ZodOptional<z.ZodEnum<["openai", "azure", "anthropic", "google", "mistral"]>>;
|
|
1440
|
+
name: z.ZodOptional<z.ZodString>;
|
|
1441
|
+
runnableType: z.ZodOptional<z.ZodEnum<["chain", "agent", "llm", "runnable"]>>;
|
|
1442
|
+
captureTraces: z.ZodOptional<z.ZodBoolean>;
|
|
1443
|
+
captureMessages: z.ZodOptional<z.ZodBoolean>;
|
|
1444
|
+
thinking: z.ZodOptional<z.ZodObject<{
|
|
1445
|
+
type: z.ZodEnum<["enabled", "disabled"]>;
|
|
1446
|
+
}, "strip", z.ZodTypeAny, {
|
|
1447
|
+
type: "enabled" | "disabled";
|
|
1448
|
+
}, {
|
|
1449
|
+
type: "enabled" | "disabled";
|
|
1450
|
+
}>>;
|
|
1451
|
+
enableSearch: z.ZodOptional<z.ZodBoolean>;
|
|
1452
|
+
searchOptions: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
1145
1453
|
}, "strip", z.ZodTypeAny, {
|
|
1454
|
+
thinking?: {
|
|
1455
|
+
type: "enabled" | "disabled";
|
|
1456
|
+
} | undefined;
|
|
1457
|
+
name?: string | undefined;
|
|
1146
1458
|
apiKey?: string | undefined;
|
|
1147
1459
|
baseUrl?: string | undefined;
|
|
1148
1460
|
defaultModel?: string | undefined;
|
|
@@ -1155,7 +1467,16 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1155
1467
|
embeddingDeploymentName?: string | undefined;
|
|
1156
1468
|
modelFamily?: string | undefined;
|
|
1157
1469
|
underlyingProvider?: "openai" | "anthropic" | "google" | "mistral" | "azure" | undefined;
|
|
1470
|
+
runnableType?: "chain" | "agent" | "llm" | "runnable" | undefined;
|
|
1471
|
+
captureTraces?: boolean | undefined;
|
|
1472
|
+
captureMessages?: boolean | undefined;
|
|
1473
|
+
enableSearch?: boolean | undefined;
|
|
1474
|
+
searchOptions?: Record<string, unknown> | undefined;
|
|
1158
1475
|
}, {
|
|
1476
|
+
thinking?: {
|
|
1477
|
+
type: "enabled" | "disabled";
|
|
1478
|
+
} | undefined;
|
|
1479
|
+
name?: string | undefined;
|
|
1159
1480
|
apiKey?: string | undefined;
|
|
1160
1481
|
baseUrl?: string | undefined;
|
|
1161
1482
|
defaultModel?: string | undefined;
|
|
@@ -1168,6 +1489,11 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1168
1489
|
embeddingDeploymentName?: string | undefined;
|
|
1169
1490
|
modelFamily?: string | undefined;
|
|
1170
1491
|
underlyingProvider?: "openai" | "anthropic" | "google" | "mistral" | "azure" | undefined;
|
|
1492
|
+
runnableType?: "chain" | "agent" | "llm" | "runnable" | undefined;
|
|
1493
|
+
captureTraces?: boolean | undefined;
|
|
1494
|
+
captureMessages?: boolean | undefined;
|
|
1495
|
+
enableSearch?: boolean | undefined;
|
|
1496
|
+
searchOptions?: Record<string, unknown> | undefined;
|
|
1171
1497
|
}>>;
|
|
1172
1498
|
seed: z.ZodOptional<z.ZodNumber>;
|
|
1173
1499
|
temperature: z.ZodOptional<z.ZodNumber>;
|
|
@@ -1200,26 +1526,178 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1200
1526
|
setup: z.ZodOptional<z.ZodObject<{
|
|
1201
1527
|
systemPrompt: z.ZodOptional<z.ZodString>;
|
|
1202
1528
|
functions: z.ZodOptional<z.ZodArray<z.ZodUnknown, "many">>;
|
|
1529
|
+
tools: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1530
|
+
type: z.ZodLiteral<"function">;
|
|
1531
|
+
function: z.ZodObject<{
|
|
1532
|
+
name: z.ZodString;
|
|
1533
|
+
description: z.ZodOptional<z.ZodString>;
|
|
1534
|
+
parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
1535
|
+
}, "strip", z.ZodTypeAny, {
|
|
1536
|
+
name: string;
|
|
1537
|
+
parameters: Record<string, unknown>;
|
|
1538
|
+
description?: string | undefined;
|
|
1539
|
+
}, {
|
|
1540
|
+
name: string;
|
|
1541
|
+
parameters: Record<string, unknown>;
|
|
1542
|
+
description?: string | undefined;
|
|
1543
|
+
}>;
|
|
1544
|
+
}, "strict", z.ZodTypeAny, {
|
|
1545
|
+
function: {
|
|
1546
|
+
name: string;
|
|
1547
|
+
parameters: Record<string, unknown>;
|
|
1548
|
+
description?: string | undefined;
|
|
1549
|
+
};
|
|
1550
|
+
type: "function";
|
|
1551
|
+
}, {
|
|
1552
|
+
function: {
|
|
1553
|
+
name: string;
|
|
1554
|
+
parameters: Record<string, unknown>;
|
|
1555
|
+
description?: string | undefined;
|
|
1556
|
+
};
|
|
1557
|
+
type: "function";
|
|
1558
|
+
}>, "many">>;
|
|
1559
|
+
fixtures: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodArray<z.ZodObject<{
|
|
1560
|
+
when: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
1561
|
+
result: z.ZodOptional<z.ZodUnknown>;
|
|
1562
|
+
error: z.ZodOptional<z.ZodString>;
|
|
1563
|
+
}, "strict", z.ZodTypeAny, {
|
|
1564
|
+
when?: Record<string, unknown> | undefined;
|
|
1565
|
+
result?: unknown;
|
|
1566
|
+
error?: string | undefined;
|
|
1567
|
+
}, {
|
|
1568
|
+
when?: Record<string, unknown> | undefined;
|
|
1569
|
+
result?: unknown;
|
|
1570
|
+
error?: string | undefined;
|
|
1571
|
+
}>, "many">>>;
|
|
1572
|
+
toolLoop: z.ZodOptional<z.ZodObject<{
|
|
1573
|
+
enabled: z.ZodDefault<z.ZodBoolean>;
|
|
1574
|
+
maxSteps: z.ZodDefault<z.ZodNumber>;
|
|
1575
|
+
timeoutMs: z.ZodDefault<z.ZodNumber>;
|
|
1576
|
+
maxToolResultBytes: z.ZodDefault<z.ZodNumber>;
|
|
1577
|
+
rejectDuplicateCalls: z.ZodDefault<z.ZodBoolean>;
|
|
1578
|
+
}, "strict", z.ZodTypeAny, {
|
|
1579
|
+
enabled: boolean;
|
|
1580
|
+
maxSteps: number;
|
|
1581
|
+
timeoutMs: number;
|
|
1582
|
+
maxToolResultBytes: number;
|
|
1583
|
+
rejectDuplicateCalls: boolean;
|
|
1584
|
+
}, {
|
|
1585
|
+
enabled?: boolean | undefined;
|
|
1586
|
+
maxSteps?: number | undefined;
|
|
1587
|
+
timeoutMs?: number | undefined;
|
|
1588
|
+
maxToolResultBytes?: number | undefined;
|
|
1589
|
+
rejectDuplicateCalls?: boolean | undefined;
|
|
1590
|
+
}>>;
|
|
1203
1591
|
}, "strip", z.ZodTypeAny, {
|
|
1204
1592
|
systemPrompt?: string | undefined;
|
|
1205
1593
|
functions?: unknown[] | undefined;
|
|
1594
|
+
tools?: {
|
|
1595
|
+
function: {
|
|
1596
|
+
name: string;
|
|
1597
|
+
parameters: Record<string, unknown>;
|
|
1598
|
+
description?: string | undefined;
|
|
1599
|
+
};
|
|
1600
|
+
type: "function";
|
|
1601
|
+
}[] | undefined;
|
|
1602
|
+
fixtures?: Record<string, {
|
|
1603
|
+
when?: Record<string, unknown> | undefined;
|
|
1604
|
+
result?: unknown;
|
|
1605
|
+
error?: string | undefined;
|
|
1606
|
+
}[]> | undefined;
|
|
1607
|
+
toolLoop?: {
|
|
1608
|
+
enabled: boolean;
|
|
1609
|
+
maxSteps: number;
|
|
1610
|
+
timeoutMs: number;
|
|
1611
|
+
maxToolResultBytes: number;
|
|
1612
|
+
rejectDuplicateCalls: boolean;
|
|
1613
|
+
} | undefined;
|
|
1206
1614
|
}, {
|
|
1207
1615
|
systemPrompt?: string | undefined;
|
|
1208
1616
|
functions?: unknown[] | undefined;
|
|
1617
|
+
tools?: {
|
|
1618
|
+
function: {
|
|
1619
|
+
name: string;
|
|
1620
|
+
parameters: Record<string, unknown>;
|
|
1621
|
+
description?: string | undefined;
|
|
1622
|
+
};
|
|
1623
|
+
type: "function";
|
|
1624
|
+
}[] | undefined;
|
|
1625
|
+
fixtures?: Record<string, {
|
|
1626
|
+
when?: Record<string, unknown> | undefined;
|
|
1627
|
+
result?: unknown;
|
|
1628
|
+
error?: string | undefined;
|
|
1629
|
+
}[]> | undefined;
|
|
1630
|
+
toolLoop?: {
|
|
1631
|
+
enabled?: boolean | undefined;
|
|
1632
|
+
maxSteps?: number | undefined;
|
|
1633
|
+
timeoutMs?: number | undefined;
|
|
1634
|
+
maxToolResultBytes?: number | undefined;
|
|
1635
|
+
rejectDuplicateCalls?: boolean | undefined;
|
|
1636
|
+
} | undefined;
|
|
1209
1637
|
}>>;
|
|
1210
1638
|
cases: z.ZodArray<z.ZodObject<{
|
|
1211
1639
|
id: z.ZodString;
|
|
1212
1640
|
name: z.ZodOptional<z.ZodString>;
|
|
1213
1641
|
description: z.ZodOptional<z.ZodString>;
|
|
1214
1642
|
prompt: z.ZodUnion<[z.ZodString, z.ZodArray<z.ZodObject<{
|
|
1215
|
-
role: z.ZodEnum<["system", "user", "assistant"]>;
|
|
1643
|
+
role: z.ZodEnum<["system", "user", "assistant", "tool"]>;
|
|
1216
1644
|
content: z.ZodString;
|
|
1645
|
+
name: z.ZodOptional<z.ZodString>;
|
|
1646
|
+
toolCallId: z.ZodOptional<z.ZodString>;
|
|
1647
|
+
tool_calls: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1648
|
+
id: z.ZodString;
|
|
1649
|
+
type: z.ZodLiteral<"function">;
|
|
1650
|
+
function: z.ZodObject<{
|
|
1651
|
+
name: z.ZodString;
|
|
1652
|
+
arguments: z.ZodString;
|
|
1653
|
+
}, "strip", z.ZodTypeAny, {
|
|
1654
|
+
name: string;
|
|
1655
|
+
arguments: string;
|
|
1656
|
+
}, {
|
|
1657
|
+
name: string;
|
|
1658
|
+
arguments: string;
|
|
1659
|
+
}>;
|
|
1660
|
+
}, "strip", z.ZodTypeAny, {
|
|
1661
|
+
function: {
|
|
1662
|
+
name: string;
|
|
1663
|
+
arguments: string;
|
|
1664
|
+
};
|
|
1665
|
+
type: "function";
|
|
1666
|
+
id: string;
|
|
1667
|
+
}, {
|
|
1668
|
+
function: {
|
|
1669
|
+
name: string;
|
|
1670
|
+
arguments: string;
|
|
1671
|
+
};
|
|
1672
|
+
type: "function";
|
|
1673
|
+
id: string;
|
|
1674
|
+
}>, "many">>;
|
|
1217
1675
|
}, "strip", z.ZodTypeAny, {
|
|
1218
|
-
role: "system" | "user" | "assistant";
|
|
1676
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
1219
1677
|
content: string;
|
|
1678
|
+
tool_calls?: {
|
|
1679
|
+
function: {
|
|
1680
|
+
name: string;
|
|
1681
|
+
arguments: string;
|
|
1682
|
+
};
|
|
1683
|
+
type: "function";
|
|
1684
|
+
id: string;
|
|
1685
|
+
}[] | undefined;
|
|
1686
|
+
name?: string | undefined;
|
|
1687
|
+
toolCallId?: string | undefined;
|
|
1220
1688
|
}, {
|
|
1221
|
-
role: "system" | "user" | "assistant";
|
|
1689
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
1222
1690
|
content: string;
|
|
1691
|
+
tool_calls?: {
|
|
1692
|
+
function: {
|
|
1693
|
+
name: string;
|
|
1694
|
+
arguments: string;
|
|
1695
|
+
};
|
|
1696
|
+
type: "function";
|
|
1697
|
+
id: string;
|
|
1698
|
+
}[] | undefined;
|
|
1699
|
+
name?: string | undefined;
|
|
1700
|
+
toolCallId?: string | undefined;
|
|
1223
1701
|
}>, "many">]>;
|
|
1224
1702
|
expected: z.ZodUnion<[z.ZodDiscriminatedUnion<"type", [z.ZodObject<{
|
|
1225
1703
|
type: z.ZodLiteral<"exact">;
|
|
@@ -1261,18 +1739,22 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1261
1739
|
type: z.ZodLiteral<"llm_grader">;
|
|
1262
1740
|
rubric: z.ZodString;
|
|
1263
1741
|
model: z.ZodOptional<z.ZodString>;
|
|
1264
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
1742
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
1265
1743
|
threshold: z.ZodDefault<z.ZodNumber>;
|
|
1744
|
+
/** Require exact, validated JSON from the judge for assurance assessments. */
|
|
1745
|
+
strict: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
1266
1746
|
}, "strip", z.ZodTypeAny, {
|
|
1267
1747
|
type: "llm_grader";
|
|
1748
|
+
strict: boolean;
|
|
1268
1749
|
threshold: number;
|
|
1269
1750
|
rubric: string;
|
|
1270
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1751
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1271
1752
|
model?: string | undefined;
|
|
1272
1753
|
}, {
|
|
1273
1754
|
type: "llm_grader";
|
|
1274
1755
|
rubric: string;
|
|
1275
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1756
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1757
|
+
strict?: boolean | undefined;
|
|
1276
1758
|
threshold?: number | undefined;
|
|
1277
1759
|
model?: string | undefined;
|
|
1278
1760
|
}>, z.ZodObject<{
|
|
@@ -1320,6 +1802,24 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1320
1802
|
type: "custom";
|
|
1321
1803
|
evaluator: string;
|
|
1322
1804
|
config?: Record<string, unknown> | undefined;
|
|
1805
|
+
}>, z.ZodObject<{
|
|
1806
|
+
type: z.ZodLiteral<"tool_trace">;
|
|
1807
|
+
requiredTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
1808
|
+
forbiddenTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
1809
|
+
ordered: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
1810
|
+
maxCalls: z.ZodOptional<z.ZodNumber>;
|
|
1811
|
+
}, "strip", z.ZodTypeAny, {
|
|
1812
|
+
type: "tool_trace";
|
|
1813
|
+
ordered: boolean;
|
|
1814
|
+
requiredTools?: string[] | undefined;
|
|
1815
|
+
forbiddenTools?: string[] | undefined;
|
|
1816
|
+
maxCalls?: number | undefined;
|
|
1817
|
+
}, {
|
|
1818
|
+
type: "tool_trace";
|
|
1819
|
+
requiredTools?: string[] | undefined;
|
|
1820
|
+
forbiddenTools?: string[] | undefined;
|
|
1821
|
+
ordered?: boolean | undefined;
|
|
1822
|
+
maxCalls?: number | undefined;
|
|
1323
1823
|
}>, z.ZodObject<{
|
|
1324
1824
|
type: z.ZodLiteral<"similarity">;
|
|
1325
1825
|
value: z.ZodString;
|
|
@@ -1335,14 +1835,14 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1335
1835
|
type: "similarity";
|
|
1336
1836
|
threshold: number;
|
|
1337
1837
|
model?: string | undefined;
|
|
1338
|
-
mode?: "
|
|
1838
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1339
1839
|
embeddingModel?: string | undefined;
|
|
1340
1840
|
}, {
|
|
1341
1841
|
value: string;
|
|
1342
1842
|
type: "similarity";
|
|
1343
1843
|
threshold?: number | undefined;
|
|
1344
1844
|
model?: string | undefined;
|
|
1345
|
-
mode?: "
|
|
1845
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1346
1846
|
embeddingModel?: string | undefined;
|
|
1347
1847
|
}>, z.ZodObject<{
|
|
1348
1848
|
type: z.ZodLiteral<"inline">;
|
|
@@ -1399,18 +1899,22 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1399
1899
|
type: z.ZodLiteral<"llm_grader">;
|
|
1400
1900
|
rubric: z.ZodString;
|
|
1401
1901
|
model: z.ZodOptional<z.ZodString>;
|
|
1402
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
1902
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
1403
1903
|
threshold: z.ZodDefault<z.ZodNumber>;
|
|
1904
|
+
/** Require exact, validated JSON from the judge for assurance assessments. */
|
|
1905
|
+
strict: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
1404
1906
|
}, "strip", z.ZodTypeAny, {
|
|
1405
1907
|
type: "llm_grader";
|
|
1908
|
+
strict: boolean;
|
|
1406
1909
|
threshold: number;
|
|
1407
1910
|
rubric: string;
|
|
1408
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1911
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1409
1912
|
model?: string | undefined;
|
|
1410
1913
|
}, {
|
|
1411
1914
|
type: "llm_grader";
|
|
1412
1915
|
rubric: string;
|
|
1413
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1916
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1917
|
+
strict?: boolean | undefined;
|
|
1414
1918
|
threshold?: number | undefined;
|
|
1415
1919
|
model?: string | undefined;
|
|
1416
1920
|
}>, z.ZodObject<{
|
|
@@ -1458,6 +1962,24 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1458
1962
|
type: "custom";
|
|
1459
1963
|
evaluator: string;
|
|
1460
1964
|
config?: Record<string, unknown> | undefined;
|
|
1965
|
+
}>, z.ZodObject<{
|
|
1966
|
+
type: z.ZodLiteral<"tool_trace">;
|
|
1967
|
+
requiredTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
1968
|
+
forbiddenTools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
1969
|
+
ordered: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
1970
|
+
maxCalls: z.ZodOptional<z.ZodNumber>;
|
|
1971
|
+
}, "strip", z.ZodTypeAny, {
|
|
1972
|
+
type: "tool_trace";
|
|
1973
|
+
ordered: boolean;
|
|
1974
|
+
requiredTools?: string[] | undefined;
|
|
1975
|
+
forbiddenTools?: string[] | undefined;
|
|
1976
|
+
maxCalls?: number | undefined;
|
|
1977
|
+
}, {
|
|
1978
|
+
type: "tool_trace";
|
|
1979
|
+
requiredTools?: string[] | undefined;
|
|
1980
|
+
forbiddenTools?: string[] | undefined;
|
|
1981
|
+
ordered?: boolean | undefined;
|
|
1982
|
+
maxCalls?: number | undefined;
|
|
1461
1983
|
}>, z.ZodObject<{
|
|
1462
1984
|
type: z.ZodLiteral<"similarity">;
|
|
1463
1985
|
value: z.ZodString;
|
|
@@ -1473,14 +1995,14 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1473
1995
|
type: "similarity";
|
|
1474
1996
|
threshold: number;
|
|
1475
1997
|
model?: string | undefined;
|
|
1476
|
-
mode?: "
|
|
1998
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1477
1999
|
embeddingModel?: string | undefined;
|
|
1478
2000
|
}, {
|
|
1479
2001
|
value: string;
|
|
1480
2002
|
type: "similarity";
|
|
1481
2003
|
threshold?: number | undefined;
|
|
1482
2004
|
model?: string | undefined;
|
|
1483
|
-
mode?: "
|
|
2005
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1484
2006
|
embeddingModel?: string | undefined;
|
|
1485
2007
|
}>, z.ZodObject<{
|
|
1486
2008
|
type: z.ZodLiteral<"inline">;
|
|
@@ -1512,9 +2034,10 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1512
2034
|
threshold: number;
|
|
1513
2035
|
} | {
|
|
1514
2036
|
type: "llm_grader";
|
|
2037
|
+
strict: boolean;
|
|
1515
2038
|
threshold: number;
|
|
1516
2039
|
rubric: string;
|
|
1517
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2040
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1518
2041
|
model?: string | undefined;
|
|
1519
2042
|
} | {
|
|
1520
2043
|
values: string[];
|
|
@@ -1531,12 +2054,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1531
2054
|
type: "custom";
|
|
1532
2055
|
evaluator: string;
|
|
1533
2056
|
config?: Record<string, unknown> | undefined;
|
|
2057
|
+
} | {
|
|
2058
|
+
type: "tool_trace";
|
|
2059
|
+
ordered: boolean;
|
|
2060
|
+
requiredTools?: string[] | undefined;
|
|
2061
|
+
forbiddenTools?: string[] | undefined;
|
|
2062
|
+
maxCalls?: number | undefined;
|
|
1534
2063
|
} | {
|
|
1535
2064
|
value: string;
|
|
1536
2065
|
type: "similarity";
|
|
1537
2066
|
threshold: number;
|
|
1538
2067
|
model?: string | undefined;
|
|
1539
|
-
mode?: "
|
|
2068
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1540
2069
|
embeddingModel?: string | undefined;
|
|
1541
2070
|
} | {
|
|
1542
2071
|
type: "inline";
|
|
@@ -1561,7 +2090,8 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1561
2090
|
} | {
|
|
1562
2091
|
type: "llm_grader";
|
|
1563
2092
|
rubric: string;
|
|
1564
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2093
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2094
|
+
strict?: boolean | undefined;
|
|
1565
2095
|
threshold?: number | undefined;
|
|
1566
2096
|
model?: string | undefined;
|
|
1567
2097
|
} | {
|
|
@@ -1579,12 +2109,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1579
2109
|
type: "custom";
|
|
1580
2110
|
evaluator: string;
|
|
1581
2111
|
config?: Record<string, unknown> | undefined;
|
|
2112
|
+
} | {
|
|
2113
|
+
type: "tool_trace";
|
|
2114
|
+
requiredTools?: string[] | undefined;
|
|
2115
|
+
forbiddenTools?: string[] | undefined;
|
|
2116
|
+
ordered?: boolean | undefined;
|
|
2117
|
+
maxCalls?: number | undefined;
|
|
1582
2118
|
} | {
|
|
1583
2119
|
value: string;
|
|
1584
2120
|
type: "similarity";
|
|
1585
2121
|
threshold?: number | undefined;
|
|
1586
2122
|
model?: string | undefined;
|
|
1587
|
-
mode?: "
|
|
2123
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1588
2124
|
embeddingModel?: string | undefined;
|
|
1589
2125
|
} | {
|
|
1590
2126
|
type: "inline";
|
|
@@ -1596,7 +2132,7 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1596
2132
|
metadata: z.ZodDefault<z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>>;
|
|
1597
2133
|
timeout: z.ZodOptional<z.ZodNumber>;
|
|
1598
2134
|
retries: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
1599
|
-
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "custom"]>>;
|
|
2135
|
+
provider: z.ZodOptional<z.ZodEnum<["openai", "azure-openai", "vercel-ai", "anthropic", "google", "mistral", "cohere", "huggingface", "ollama", "langchain", "deepagents", "ling", "custom"]>>;
|
|
1600
2136
|
model: z.ZodOptional<z.ZodString>;
|
|
1601
2137
|
variables: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnion<[z.ZodString, z.ZodNumber, z.ZodBoolean]>>>;
|
|
1602
2138
|
/** Case-level redaction config (overrides scenario-level) */
|
|
@@ -1623,6 +2159,7 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1623
2159
|
replacement?: string | undefined;
|
|
1624
2160
|
}>>;
|
|
1625
2161
|
}, "strip", z.ZodTypeAny, {
|
|
2162
|
+
metadata: Record<string, unknown>;
|
|
1626
2163
|
expected: {
|
|
1627
2164
|
value: string;
|
|
1628
2165
|
type: "exact";
|
|
@@ -1637,9 +2174,10 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1637
2174
|
threshold: number;
|
|
1638
2175
|
} | {
|
|
1639
2176
|
type: "llm_grader";
|
|
2177
|
+
strict: boolean;
|
|
1640
2178
|
threshold: number;
|
|
1641
2179
|
rubric: string;
|
|
1642
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2180
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1643
2181
|
model?: string | undefined;
|
|
1644
2182
|
} | {
|
|
1645
2183
|
values: string[];
|
|
@@ -1656,12 +2194,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1656
2194
|
type: "custom";
|
|
1657
2195
|
evaluator: string;
|
|
1658
2196
|
config?: Record<string, unknown> | undefined;
|
|
2197
|
+
} | {
|
|
2198
|
+
type: "tool_trace";
|
|
2199
|
+
ordered: boolean;
|
|
2200
|
+
requiredTools?: string[] | undefined;
|
|
2201
|
+
forbiddenTools?: string[] | undefined;
|
|
2202
|
+
maxCalls?: number | undefined;
|
|
1659
2203
|
} | {
|
|
1660
2204
|
value: string;
|
|
1661
2205
|
type: "similarity";
|
|
1662
2206
|
threshold: number;
|
|
1663
2207
|
model?: string | undefined;
|
|
1664
|
-
mode?: "
|
|
2208
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1665
2209
|
embeddingModel?: string | undefined;
|
|
1666
2210
|
} | {
|
|
1667
2211
|
type: "inline";
|
|
@@ -1684,9 +2228,10 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1684
2228
|
threshold: number;
|
|
1685
2229
|
} | {
|
|
1686
2230
|
type: "llm_grader";
|
|
2231
|
+
strict: boolean;
|
|
1687
2232
|
threshold: number;
|
|
1688
2233
|
rubric: string;
|
|
1689
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2234
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1690
2235
|
model?: string | undefined;
|
|
1691
2236
|
} | {
|
|
1692
2237
|
values: string[];
|
|
@@ -1703,12 +2248,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1703
2248
|
type: "custom";
|
|
1704
2249
|
evaluator: string;
|
|
1705
2250
|
config?: Record<string, unknown> | undefined;
|
|
2251
|
+
} | {
|
|
2252
|
+
type: "tool_trace";
|
|
2253
|
+
ordered: boolean;
|
|
2254
|
+
requiredTools?: string[] | undefined;
|
|
2255
|
+
forbiddenTools?: string[] | undefined;
|
|
2256
|
+
maxCalls?: number | undefined;
|
|
1706
2257
|
} | {
|
|
1707
2258
|
value: string;
|
|
1708
2259
|
type: "similarity";
|
|
1709
2260
|
threshold: number;
|
|
1710
2261
|
model?: string | undefined;
|
|
1711
|
-
mode?: "
|
|
2262
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1712
2263
|
embeddingModel?: string | undefined;
|
|
1713
2264
|
} | {
|
|
1714
2265
|
type: "inline";
|
|
@@ -1718,14 +2269,23 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1718
2269
|
};
|
|
1719
2270
|
id: string;
|
|
1720
2271
|
prompt: string | {
|
|
1721
|
-
role: "system" | "user" | "assistant";
|
|
2272
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
1722
2273
|
content: string;
|
|
2274
|
+
tool_calls?: {
|
|
2275
|
+
function: {
|
|
2276
|
+
name: string;
|
|
2277
|
+
arguments: string;
|
|
2278
|
+
};
|
|
2279
|
+
type: "function";
|
|
2280
|
+
id: string;
|
|
2281
|
+
}[] | undefined;
|
|
2282
|
+
name?: string | undefined;
|
|
2283
|
+
toolCallId?: string | undefined;
|
|
1723
2284
|
}[];
|
|
1724
2285
|
tags: string[];
|
|
1725
|
-
metadata: Record<string, unknown>;
|
|
1726
2286
|
retries: number;
|
|
1727
2287
|
name?: string | undefined;
|
|
1728
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2288
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1729
2289
|
timeout?: number | undefined;
|
|
1730
2290
|
model?: string | undefined;
|
|
1731
2291
|
description?: string | undefined;
|
|
@@ -1754,7 +2314,8 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1754
2314
|
} | {
|
|
1755
2315
|
type: "llm_grader";
|
|
1756
2316
|
rubric: string;
|
|
1757
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2317
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2318
|
+
strict?: boolean | undefined;
|
|
1758
2319
|
threshold?: number | undefined;
|
|
1759
2320
|
model?: string | undefined;
|
|
1760
2321
|
} | {
|
|
@@ -1772,12 +2333,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1772
2333
|
type: "custom";
|
|
1773
2334
|
evaluator: string;
|
|
1774
2335
|
config?: Record<string, unknown> | undefined;
|
|
2336
|
+
} | {
|
|
2337
|
+
type: "tool_trace";
|
|
2338
|
+
requiredTools?: string[] | undefined;
|
|
2339
|
+
forbiddenTools?: string[] | undefined;
|
|
2340
|
+
ordered?: boolean | undefined;
|
|
2341
|
+
maxCalls?: number | undefined;
|
|
1775
2342
|
} | {
|
|
1776
2343
|
value: string;
|
|
1777
2344
|
type: "similarity";
|
|
1778
2345
|
threshold?: number | undefined;
|
|
1779
2346
|
model?: string | undefined;
|
|
1780
|
-
mode?: "
|
|
2347
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1781
2348
|
embeddingModel?: string | undefined;
|
|
1782
2349
|
} | {
|
|
1783
2350
|
type: "inline";
|
|
@@ -1801,7 +2368,8 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1801
2368
|
} | {
|
|
1802
2369
|
type: "llm_grader";
|
|
1803
2370
|
rubric: string;
|
|
1804
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2371
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2372
|
+
strict?: boolean | undefined;
|
|
1805
2373
|
threshold?: number | undefined;
|
|
1806
2374
|
model?: string | undefined;
|
|
1807
2375
|
} | {
|
|
@@ -1819,12 +2387,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1819
2387
|
type: "custom";
|
|
1820
2388
|
evaluator: string;
|
|
1821
2389
|
config?: Record<string, unknown> | undefined;
|
|
2390
|
+
} | {
|
|
2391
|
+
type: "tool_trace";
|
|
2392
|
+
requiredTools?: string[] | undefined;
|
|
2393
|
+
forbiddenTools?: string[] | undefined;
|
|
2394
|
+
ordered?: boolean | undefined;
|
|
2395
|
+
maxCalls?: number | undefined;
|
|
1822
2396
|
} | {
|
|
1823
2397
|
value: string;
|
|
1824
2398
|
type: "similarity";
|
|
1825
2399
|
threshold?: number | undefined;
|
|
1826
2400
|
model?: string | undefined;
|
|
1827
|
-
mode?: "
|
|
2401
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1828
2402
|
embeddingModel?: string | undefined;
|
|
1829
2403
|
} | {
|
|
1830
2404
|
type: "inline";
|
|
@@ -1834,16 +2408,26 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1834
2408
|
};
|
|
1835
2409
|
id: string;
|
|
1836
2410
|
prompt: string | {
|
|
1837
|
-
role: "system" | "user" | "assistant";
|
|
2411
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
1838
2412
|
content: string;
|
|
2413
|
+
tool_calls?: {
|
|
2414
|
+
function: {
|
|
2415
|
+
name: string;
|
|
2416
|
+
arguments: string;
|
|
2417
|
+
};
|
|
2418
|
+
type: "function";
|
|
2419
|
+
id: string;
|
|
2420
|
+
}[] | undefined;
|
|
2421
|
+
name?: string | undefined;
|
|
2422
|
+
toolCallId?: string | undefined;
|
|
1839
2423
|
}[];
|
|
1840
2424
|
name?: string | undefined;
|
|
1841
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2425
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2426
|
+
metadata?: Record<string, unknown> | undefined;
|
|
1842
2427
|
timeout?: number | undefined;
|
|
1843
2428
|
model?: string | undefined;
|
|
1844
2429
|
description?: string | undefined;
|
|
1845
2430
|
tags?: string[] | undefined;
|
|
1846
|
-
metadata?: Record<string, unknown> | undefined;
|
|
1847
2431
|
retries?: number | undefined;
|
|
1848
2432
|
variables?: Record<string, string | number | boolean> | undefined;
|
|
1849
2433
|
redaction?: {
|
|
@@ -1867,6 +2451,7 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1867
2451
|
tags: string[];
|
|
1868
2452
|
version: string;
|
|
1869
2453
|
cases: {
|
|
2454
|
+
metadata: Record<string, unknown>;
|
|
1870
2455
|
expected: {
|
|
1871
2456
|
value: string;
|
|
1872
2457
|
type: "exact";
|
|
@@ -1881,9 +2466,10 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1881
2466
|
threshold: number;
|
|
1882
2467
|
} | {
|
|
1883
2468
|
type: "llm_grader";
|
|
2469
|
+
strict: boolean;
|
|
1884
2470
|
threshold: number;
|
|
1885
2471
|
rubric: string;
|
|
1886
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2472
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1887
2473
|
model?: string | undefined;
|
|
1888
2474
|
} | {
|
|
1889
2475
|
values: string[];
|
|
@@ -1900,12 +2486,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1900
2486
|
type: "custom";
|
|
1901
2487
|
evaluator: string;
|
|
1902
2488
|
config?: Record<string, unknown> | undefined;
|
|
2489
|
+
} | {
|
|
2490
|
+
type: "tool_trace";
|
|
2491
|
+
ordered: boolean;
|
|
2492
|
+
requiredTools?: string[] | undefined;
|
|
2493
|
+
forbiddenTools?: string[] | undefined;
|
|
2494
|
+
maxCalls?: number | undefined;
|
|
1903
2495
|
} | {
|
|
1904
2496
|
value: string;
|
|
1905
2497
|
type: "similarity";
|
|
1906
2498
|
threshold: number;
|
|
1907
2499
|
model?: string | undefined;
|
|
1908
|
-
mode?: "
|
|
2500
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1909
2501
|
embeddingModel?: string | undefined;
|
|
1910
2502
|
} | {
|
|
1911
2503
|
type: "inline";
|
|
@@ -1928,9 +2520,10 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1928
2520
|
threshold: number;
|
|
1929
2521
|
} | {
|
|
1930
2522
|
type: "llm_grader";
|
|
2523
|
+
strict: boolean;
|
|
1931
2524
|
threshold: number;
|
|
1932
2525
|
rubric: string;
|
|
1933
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2526
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1934
2527
|
model?: string | undefined;
|
|
1935
2528
|
} | {
|
|
1936
2529
|
values: string[];
|
|
@@ -1947,12 +2540,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1947
2540
|
type: "custom";
|
|
1948
2541
|
evaluator: string;
|
|
1949
2542
|
config?: Record<string, unknown> | undefined;
|
|
2543
|
+
} | {
|
|
2544
|
+
type: "tool_trace";
|
|
2545
|
+
ordered: boolean;
|
|
2546
|
+
requiredTools?: string[] | undefined;
|
|
2547
|
+
forbiddenTools?: string[] | undefined;
|
|
2548
|
+
maxCalls?: number | undefined;
|
|
1950
2549
|
} | {
|
|
1951
2550
|
value: string;
|
|
1952
2551
|
type: "similarity";
|
|
1953
2552
|
threshold: number;
|
|
1954
2553
|
model?: string | undefined;
|
|
1955
|
-
mode?: "
|
|
2554
|
+
mode?: "llm" | "embedding" | undefined;
|
|
1956
2555
|
embeddingModel?: string | undefined;
|
|
1957
2556
|
} | {
|
|
1958
2557
|
type: "inline";
|
|
@@ -1962,14 +2561,23 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1962
2561
|
};
|
|
1963
2562
|
id: string;
|
|
1964
2563
|
prompt: string | {
|
|
1965
|
-
role: "system" | "user" | "assistant";
|
|
2564
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
1966
2565
|
content: string;
|
|
2566
|
+
tool_calls?: {
|
|
2567
|
+
function: {
|
|
2568
|
+
name: string;
|
|
2569
|
+
arguments: string;
|
|
2570
|
+
};
|
|
2571
|
+
type: "function";
|
|
2572
|
+
id: string;
|
|
2573
|
+
}[] | undefined;
|
|
2574
|
+
name?: string | undefined;
|
|
2575
|
+
toolCallId?: string | undefined;
|
|
1967
2576
|
}[];
|
|
1968
2577
|
tags: string[];
|
|
1969
|
-
metadata: Record<string, unknown>;
|
|
1970
2578
|
retries: number;
|
|
1971
2579
|
name?: string | undefined;
|
|
1972
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2580
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1973
2581
|
timeout?: number | undefined;
|
|
1974
2582
|
model?: string | undefined;
|
|
1975
2583
|
description?: string | undefined;
|
|
@@ -1983,7 +2591,7 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1983
2591
|
patterns?: string[] | undefined;
|
|
1984
2592
|
} | undefined;
|
|
1985
2593
|
}[];
|
|
1986
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2594
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
1987
2595
|
model?: string | undefined;
|
|
1988
2596
|
description?: string | undefined;
|
|
1989
2597
|
variables?: Record<string, string | number | boolean> | undefined;
|
|
@@ -1996,6 +2604,10 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
1996
2604
|
patterns?: string[] | undefined;
|
|
1997
2605
|
} | undefined;
|
|
1998
2606
|
providerConfig?: {
|
|
2607
|
+
thinking?: {
|
|
2608
|
+
type: "enabled" | "disabled";
|
|
2609
|
+
} | undefined;
|
|
2610
|
+
name?: string | undefined;
|
|
1999
2611
|
apiKey?: string | undefined;
|
|
2000
2612
|
baseUrl?: string | undefined;
|
|
2001
2613
|
defaultModel?: string | undefined;
|
|
@@ -2008,6 +2620,11 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2008
2620
|
embeddingDeploymentName?: string | undefined;
|
|
2009
2621
|
modelFamily?: string | undefined;
|
|
2010
2622
|
underlyingProvider?: "openai" | "anthropic" | "google" | "mistral" | "azure" | undefined;
|
|
2623
|
+
runnableType?: "chain" | "agent" | "llm" | "runnable" | undefined;
|
|
2624
|
+
captureTraces?: boolean | undefined;
|
|
2625
|
+
captureMessages?: boolean | undefined;
|
|
2626
|
+
enableSearch?: boolean | undefined;
|
|
2627
|
+
searchOptions?: Record<string, unknown> | undefined;
|
|
2011
2628
|
} | undefined;
|
|
2012
2629
|
seed?: number | undefined;
|
|
2013
2630
|
temperature?: number | undefined;
|
|
@@ -2015,6 +2632,26 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2015
2632
|
setup?: {
|
|
2016
2633
|
systemPrompt?: string | undefined;
|
|
2017
2634
|
functions?: unknown[] | undefined;
|
|
2635
|
+
tools?: {
|
|
2636
|
+
function: {
|
|
2637
|
+
name: string;
|
|
2638
|
+
parameters: Record<string, unknown>;
|
|
2639
|
+
description?: string | undefined;
|
|
2640
|
+
};
|
|
2641
|
+
type: "function";
|
|
2642
|
+
}[] | undefined;
|
|
2643
|
+
fixtures?: Record<string, {
|
|
2644
|
+
when?: Record<string, unknown> | undefined;
|
|
2645
|
+
result?: unknown;
|
|
2646
|
+
error?: string | undefined;
|
|
2647
|
+
}[]> | undefined;
|
|
2648
|
+
toolLoop?: {
|
|
2649
|
+
enabled: boolean;
|
|
2650
|
+
maxSteps: number;
|
|
2651
|
+
timeoutMs: number;
|
|
2652
|
+
maxToolResultBytes: number;
|
|
2653
|
+
rejectDuplicateCalls: boolean;
|
|
2654
|
+
} | undefined;
|
|
2018
2655
|
} | undefined;
|
|
2019
2656
|
teardown?: {
|
|
2020
2657
|
cleanup?: boolean | undefined;
|
|
@@ -2037,7 +2674,8 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2037
2674
|
} | {
|
|
2038
2675
|
type: "llm_grader";
|
|
2039
2676
|
rubric: string;
|
|
2040
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2677
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2678
|
+
strict?: boolean | undefined;
|
|
2041
2679
|
threshold?: number | undefined;
|
|
2042
2680
|
model?: string | undefined;
|
|
2043
2681
|
} | {
|
|
@@ -2055,12 +2693,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2055
2693
|
type: "custom";
|
|
2056
2694
|
evaluator: string;
|
|
2057
2695
|
config?: Record<string, unknown> | undefined;
|
|
2696
|
+
} | {
|
|
2697
|
+
type: "tool_trace";
|
|
2698
|
+
requiredTools?: string[] | undefined;
|
|
2699
|
+
forbiddenTools?: string[] | undefined;
|
|
2700
|
+
ordered?: boolean | undefined;
|
|
2701
|
+
maxCalls?: number | undefined;
|
|
2058
2702
|
} | {
|
|
2059
2703
|
value: string;
|
|
2060
2704
|
type: "similarity";
|
|
2061
2705
|
threshold?: number | undefined;
|
|
2062
2706
|
model?: string | undefined;
|
|
2063
|
-
mode?: "
|
|
2707
|
+
mode?: "llm" | "embedding" | undefined;
|
|
2064
2708
|
embeddingModel?: string | undefined;
|
|
2065
2709
|
} | {
|
|
2066
2710
|
type: "inline";
|
|
@@ -2084,7 +2728,8 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2084
2728
|
} | {
|
|
2085
2729
|
type: "llm_grader";
|
|
2086
2730
|
rubric: string;
|
|
2087
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2731
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2732
|
+
strict?: boolean | undefined;
|
|
2088
2733
|
threshold?: number | undefined;
|
|
2089
2734
|
model?: string | undefined;
|
|
2090
2735
|
} | {
|
|
@@ -2102,12 +2747,18 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2102
2747
|
type: "custom";
|
|
2103
2748
|
evaluator: string;
|
|
2104
2749
|
config?: Record<string, unknown> | undefined;
|
|
2750
|
+
} | {
|
|
2751
|
+
type: "tool_trace";
|
|
2752
|
+
requiredTools?: string[] | undefined;
|
|
2753
|
+
forbiddenTools?: string[] | undefined;
|
|
2754
|
+
ordered?: boolean | undefined;
|
|
2755
|
+
maxCalls?: number | undefined;
|
|
2105
2756
|
} | {
|
|
2106
2757
|
value: string;
|
|
2107
2758
|
type: "similarity";
|
|
2108
2759
|
threshold?: number | undefined;
|
|
2109
2760
|
model?: string | undefined;
|
|
2110
|
-
mode?: "
|
|
2761
|
+
mode?: "llm" | "embedding" | undefined;
|
|
2111
2762
|
embeddingModel?: string | undefined;
|
|
2112
2763
|
} | {
|
|
2113
2764
|
type: "inline";
|
|
@@ -2117,16 +2768,26 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2117
2768
|
};
|
|
2118
2769
|
id: string;
|
|
2119
2770
|
prompt: string | {
|
|
2120
|
-
role: "system" | "user" | "assistant";
|
|
2771
|
+
role: "system" | "user" | "assistant" | "tool";
|
|
2121
2772
|
content: string;
|
|
2773
|
+
tool_calls?: {
|
|
2774
|
+
function: {
|
|
2775
|
+
name: string;
|
|
2776
|
+
arguments: string;
|
|
2777
|
+
};
|
|
2778
|
+
type: "function";
|
|
2779
|
+
id: string;
|
|
2780
|
+
}[] | undefined;
|
|
2781
|
+
name?: string | undefined;
|
|
2782
|
+
toolCallId?: string | undefined;
|
|
2122
2783
|
}[];
|
|
2123
2784
|
name?: string | undefined;
|
|
2124
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2785
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2786
|
+
metadata?: Record<string, unknown> | undefined;
|
|
2125
2787
|
timeout?: number | undefined;
|
|
2126
2788
|
model?: string | undefined;
|
|
2127
2789
|
description?: string | undefined;
|
|
2128
2790
|
tags?: string[] | undefined;
|
|
2129
|
-
metadata?: Record<string, unknown> | undefined;
|
|
2130
2791
|
retries?: number | undefined;
|
|
2131
2792
|
variables?: Record<string, string | number | boolean> | undefined;
|
|
2132
2793
|
redaction?: {
|
|
@@ -2138,7 +2799,7 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2138
2799
|
replacement?: string | undefined;
|
|
2139
2800
|
} | undefined;
|
|
2140
2801
|
}[];
|
|
2141
|
-
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2802
|
+
provider?: "openai" | "azure-openai" | "vercel-ai" | "anthropic" | "langchain" | "deepagents" | "ling" | "google" | "mistral" | "cohere" | "huggingface" | "ollama" | "custom" | undefined;
|
|
2142
2803
|
model?: string | undefined;
|
|
2143
2804
|
description?: string | undefined;
|
|
2144
2805
|
tags?: string[] | undefined;
|
|
@@ -2153,6 +2814,10 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2153
2814
|
} | undefined;
|
|
2154
2815
|
version?: string | undefined;
|
|
2155
2816
|
providerConfig?: {
|
|
2817
|
+
thinking?: {
|
|
2818
|
+
type: "enabled" | "disabled";
|
|
2819
|
+
} | undefined;
|
|
2820
|
+
name?: string | undefined;
|
|
2156
2821
|
apiKey?: string | undefined;
|
|
2157
2822
|
baseUrl?: string | undefined;
|
|
2158
2823
|
defaultModel?: string | undefined;
|
|
@@ -2165,6 +2830,11 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2165
2830
|
embeddingDeploymentName?: string | undefined;
|
|
2166
2831
|
modelFamily?: string | undefined;
|
|
2167
2832
|
underlyingProvider?: "openai" | "anthropic" | "google" | "mistral" | "azure" | undefined;
|
|
2833
|
+
runnableType?: "chain" | "agent" | "llm" | "runnable" | undefined;
|
|
2834
|
+
captureTraces?: boolean | undefined;
|
|
2835
|
+
captureMessages?: boolean | undefined;
|
|
2836
|
+
enableSearch?: boolean | undefined;
|
|
2837
|
+
searchOptions?: Record<string, unknown> | undefined;
|
|
2168
2838
|
} | undefined;
|
|
2169
2839
|
seed?: number | undefined;
|
|
2170
2840
|
temperature?: number | undefined;
|
|
@@ -2172,6 +2842,26 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
2172
2842
|
setup?: {
|
|
2173
2843
|
systemPrompt?: string | undefined;
|
|
2174
2844
|
functions?: unknown[] | undefined;
|
|
2845
|
+
tools?: {
|
|
2846
|
+
function: {
|
|
2847
|
+
name: string;
|
|
2848
|
+
parameters: Record<string, unknown>;
|
|
2849
|
+
description?: string | undefined;
|
|
2850
|
+
};
|
|
2851
|
+
type: "function";
|
|
2852
|
+
}[] | undefined;
|
|
2853
|
+
fixtures?: Record<string, {
|
|
2854
|
+
when?: Record<string, unknown> | undefined;
|
|
2855
|
+
result?: unknown;
|
|
2856
|
+
error?: string | undefined;
|
|
2857
|
+
}[]> | undefined;
|
|
2858
|
+
toolLoop?: {
|
|
2859
|
+
enabled?: boolean | undefined;
|
|
2860
|
+
maxSteps?: number | undefined;
|
|
2861
|
+
timeoutMs?: number | undefined;
|
|
2862
|
+
maxToolResultBytes?: number | undefined;
|
|
2863
|
+
rejectDuplicateCalls?: boolean | undefined;
|
|
2864
|
+
} | undefined;
|
|
2175
2865
|
} | undefined;
|
|
2176
2866
|
teardown?: {
|
|
2177
2867
|
cleanup?: boolean | undefined;
|