promptfoo 0.71.1 → 0.72.1

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.
Files changed (121) hide show
  1. package/LICENSE +1 -1
  2. package/dist/package.json +2 -2
  3. package/dist/src/commands/generate/redteam.d.ts +43 -13
  4. package/dist/src/commands/generate/redteam.d.ts.map +1 -1
  5. package/dist/src/commands/generate/redteam.js +63 -22
  6. package/dist/src/commands/generate/redteam.js.map +1 -1
  7. package/dist/src/commands/redteam.d.ts +0 -7
  8. package/dist/src/commands/redteam.d.ts.map +1 -1
  9. package/dist/src/commands/redteam.js +72 -89
  10. package/dist/src/commands/redteam.js.map +1 -1
  11. package/dist/src/commands/share.d.ts.map +1 -1
  12. package/dist/src/commands/share.js +4 -1
  13. package/dist/src/commands/share.js.map +1 -1
  14. package/dist/src/constants.d.ts +1 -0
  15. package/dist/src/constants.d.ts.map +1 -1
  16. package/dist/src/constants.js +6 -1
  17. package/dist/src/constants.js.map +1 -1
  18. package/dist/src/database.d.ts +1129 -3
  19. package/dist/src/database.d.ts.map +1 -1
  20. package/dist/src/onboarding.d.ts.map +1 -1
  21. package/dist/src/onboarding.js +23 -36
  22. package/dist/src/onboarding.js.map +1 -1
  23. package/dist/src/providers/bedrock.js +1 -1
  24. package/dist/src/providers/bedrock.js.map +1 -1
  25. package/dist/src/providers/manualInput.d.ts +15 -0
  26. package/dist/src/providers/manualInput.d.ts.map +1 -0
  27. package/dist/src/providers/manualInput.js +33 -0
  28. package/dist/src/providers/manualInput.js.map +1 -0
  29. package/dist/src/providers/mistral.d.ts.map +1 -1
  30. package/dist/src/providers/mistral.js +7 -0
  31. package/dist/src/providers/mistral.js.map +1 -1
  32. package/dist/src/providers.d.ts.map +1 -1
  33. package/dist/src/providers.js +12 -9
  34. package/dist/src/providers.js.map +1 -1
  35. package/dist/src/redteam/constants.d.ts +2 -2
  36. package/dist/src/redteam/constants.d.ts.map +1 -1
  37. package/dist/src/redteam/constants.js +1 -1
  38. package/dist/src/redteam/constants.js.map +1 -1
  39. package/dist/src/redteam/index.d.ts +2 -6
  40. package/dist/src/redteam/index.d.ts.map +1 -1
  41. package/dist/src/redteam/index.js +7 -16
  42. package/dist/src/redteam/index.js.map +1 -1
  43. package/dist/src/redteam/methods/injections.d.ts.map +1 -1
  44. package/dist/src/redteam/methods/injections.js +10 -0
  45. package/dist/src/redteam/methods/injections.js.map +1 -1
  46. package/dist/src/redteam/plugins/harmful.d.ts.map +1 -1
  47. package/dist/src/redteam/plugins/harmful.js +20 -17
  48. package/dist/src/redteam/plugins/harmful.js.map +1 -1
  49. package/dist/src/redteam/providers/iterative.js +1 -1
  50. package/dist/src/redteam/providers/iterative.js.map +1 -1
  51. package/dist/src/redteam/providers/iterativeImage.js +1 -1
  52. package/dist/src/redteam/providers/iterativeImage.js.map +1 -1
  53. package/dist/src/types.d.ts +30599 -174
  54. package/dist/src/types.d.ts.map +1 -1
  55. package/dist/src/types.js +242 -1
  56. package/dist/src/types.js.map +1 -1
  57. package/dist/src/updates.d.ts.map +1 -1
  58. package/dist/src/updates.js +2 -4
  59. package/dist/src/updates.js.map +1 -1
  60. package/dist/src/util/index.d.ts +521 -3
  61. package/dist/src/util/index.d.ts.map +1 -1
  62. package/dist/src/util/index.js +11 -10
  63. package/dist/src/util/index.js.map +1 -1
  64. package/dist/src/web/nextui/404/index.html +1 -1
  65. package/dist/src/web/nextui/404.html +1 -1
  66. package/dist/src/web/nextui/_next/static/chunks/{258-9bad5f8846b93785.js → 258-4acb452fe85cff6f.js} +2 -2
  67. package/dist/src/web/nextui/_next/static/chunks/{540-ab57db4fdf86cc8a.js → 448-4da3bf74eae4996a.js} +7 -7
  68. package/dist/src/web/nextui/_next/static/chunks/53-fae6e50ace6c83a1.js +1 -0
  69. package/dist/src/web/nextui/_next/static/chunks/{106-40e3e363c957c313.js → 578-8efe57f906f7573c.js} +6 -6
  70. package/dist/src/web/nextui/_next/static/chunks/737-538f50dc31cc8c49.js +1 -0
  71. package/dist/src/web/nextui/_next/static/chunks/{531-efd0c50e833e81f1.js → 905-0da344cee75690e7.js} +4 -4
  72. package/dist/src/web/nextui/_next/static/chunks/932-75585b3be8ce494d.js +1 -0
  73. package/dist/src/web/nextui/_next/static/chunks/973-ffe3726c956d08ef.js +2 -0
  74. package/dist/src/web/nextui/_next/static/chunks/app/auth/login/page-2e2c0c725127ea2d.js +1 -0
  75. package/dist/src/web/nextui/_next/static/chunks/app/auth/signup/page-ccdb1f6890601666.js +1 -0
  76. package/dist/src/web/nextui/_next/static/chunks/app/datasets/{page-4f93aacd25866d60.js → page-642db4f4a8e8ba40.js} +1 -1
  77. package/dist/src/web/nextui/_next/static/chunks/app/eval/[id]/{page-7bc08c50ac222ddd.js → page-0a39d3450aa09dda.js} +1 -1
  78. package/dist/src/web/nextui/_next/static/chunks/app/eval/page-32eaa14d2384c5b0.js +1 -0
  79. package/dist/src/web/nextui/_next/static/chunks/app/layout-4282b1d33566e258.js +1 -0
  80. package/dist/src/web/nextui/_next/static/chunks/app/{page-e07a0ddbf3d6e21c.js → page-5650318e57510b6c.js} +1 -1
  81. package/dist/src/web/nextui/_next/static/chunks/app/progress/page-462526776efd6fd6.js +1 -0
  82. package/dist/src/web/nextui/_next/static/chunks/app/prompts/{page-50e27c24c9e255bd.js → page-0ae78bc0bf7c56bc.js} +1 -1
  83. package/dist/src/web/nextui/_next/static/chunks/app/report/{page-6716acff9825da98.js → page-e4f7851d41eb0f92.js} +1 -1
  84. package/dist/src/web/nextui/_next/static/chunks/app/setup/{page-59d712c1f56ecac6.js → page-6cde750ec428cd75.js} +1 -1
  85. package/dist/src/web/nextui/_next/static/chunks/{webpack-8a9bc9ee0defb756.js → webpack-c9f728822666f852.js} +1 -1
  86. package/dist/src/web/nextui/_next/static/css/5bd2f45de1f3ba83.css +1 -0
  87. package/dist/src/web/nextui/auth/login/index.html +1 -1
  88. package/dist/src/web/nextui/auth/login/index.txt +6 -6
  89. package/dist/src/web/nextui/auth/signup/index.html +1 -1
  90. package/dist/src/web/nextui/auth/signup/index.txt +6 -6
  91. package/dist/src/web/nextui/datasets/index.html +1 -1
  92. package/dist/src/web/nextui/datasets/index.txt +6 -6
  93. package/dist/src/web/nextui/eval/index.html +1 -1
  94. package/dist/src/web/nextui/eval/index.txt +6 -6
  95. package/dist/src/web/nextui/index.html +1 -1
  96. package/dist/src/web/nextui/index.txt +5 -5
  97. package/dist/src/web/nextui/progress/index.html +1 -1
  98. package/dist/src/web/nextui/progress/index.txt +6 -6
  99. package/dist/src/web/nextui/prompts/index.html +1 -1
  100. package/dist/src/web/nextui/prompts/index.txt +6 -6
  101. package/dist/src/web/nextui/report/index.html +1 -1
  102. package/dist/src/web/nextui/report/index.txt +6 -6
  103. package/dist/src/web/nextui/setup/index.html +1 -27
  104. package/dist/src/web/nextui/setup/index.txt +7 -7
  105. package/dist/src/web/server.d.ts.map +1 -1
  106. package/dist/src/web/server.js +2 -1
  107. package/dist/src/web/server.js.map +1 -1
  108. package/package.json +2 -2
  109. package/dist/src/web/nextui/_next/static/chunks/215-f97c02cbf00861f3.js +0 -1
  110. package/dist/src/web/nextui/_next/static/chunks/611-d73a08bebf1d6f3c.js +0 -1
  111. package/dist/src/web/nextui/_next/static/chunks/770-8ec8b39ce6f615c4.js +0 -2
  112. package/dist/src/web/nextui/_next/static/chunks/858-9393537662061ee7.js +0 -1
  113. package/dist/src/web/nextui/_next/static/chunks/873-177b84c58d904b5a.js +0 -1
  114. package/dist/src/web/nextui/_next/static/chunks/app/auth/login/page-3bd92b90dc19429b.js +0 -1
  115. package/dist/src/web/nextui/_next/static/chunks/app/auth/signup/page-557ff1f8c491e68a.js +0 -1
  116. package/dist/src/web/nextui/_next/static/chunks/app/eval/page-b26513300f6f0273.js +0 -1
  117. package/dist/src/web/nextui/_next/static/chunks/app/layout-1b9c6b87155ccc6f.js +0 -1
  118. package/dist/src/web/nextui/_next/static/chunks/app/progress/page-e8235b15bd64fb03.js +0 -1
  119. package/dist/src/web/nextui/_next/static/css/106779eb64615639.css +0 -1
  120. /package/dist/src/web/nextui/_next/static/{FqU2doca_2rkP7LW_2ZLN → 8E4q0Tmu22G1wENZyXEq0}/_buildManifest.js +0 -0
  121. /package/dist/src/web/nextui/_next/static/{FqU2doca_2rkP7LW_2ZLN → 8E4q0Tmu22G1wENZyXEq0}/_ssgManifest.js +0 -0
@@ -1,4 +1,4 @@
1
- import type { EvaluateSummary, UnifiedConfig } from './types';
1
+ import type { EvaluateSummary } from './types';
2
2
  export declare const prompts: import("drizzle-orm/sqlite-core").SQLiteTableWithColumns<{
3
3
  name: "prompts";
4
4
  schema: undefined;
@@ -143,7 +143,976 @@ export declare const evals: import("drizzle-orm/sqlite-core").SQLiteTableWithCol
143
143
  tableName: "evals";
144
144
  dataType: "json";
145
145
  columnType: "SQLiteTextJson";
146
- data: Partial<UnifiedConfig>;
146
+ data: Partial<{
147
+ prompts: string | Record<string, string> | (string | {
148
+ label: string;
149
+ raw: string;
150
+ function?: ((args_0: {
151
+ vars: Record<string, any>;
152
+ provider?: {
153
+ id: (...args: unknown[]) => string;
154
+ callApi: (((args_0: string, args_1: {
155
+ vars: Record<string, string | {}>;
156
+ logger?: any;
157
+ fetchWithCache?: any;
158
+ getCache?: any;
159
+ } | undefined, args_2: {
160
+ includeLogProbs?: boolean | undefined;
161
+ originalProvider?: any;
162
+ } | undefined, ...args_3: unknown[]) => Promise<{
163
+ output?: any;
164
+ cached?: boolean | undefined;
165
+ cost?: number | undefined;
166
+ error?: string | undefined;
167
+ logProbs?: number[] | undefined;
168
+ metadata?: import("zod").objectOutputType<{
169
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
170
+ }, import("zod").ZodAny, "strip"> | undefined;
171
+ tokenUsage?: {
172
+ prompt?: number | undefined;
173
+ cached?: number | undefined;
174
+ completion?: number | undefined;
175
+ total?: number | undefined;
176
+ } | undefined;
177
+ }>) & {
178
+ label?: string | undefined;
179
+ }) | (((args_0: string, args_1: {
180
+ includeLogProbs?: boolean | undefined;
181
+ originalProvider?: any;
182
+ } | undefined, ...args_2: unknown[]) => Promise<{
183
+ output?: any;
184
+ cached?: boolean | undefined;
185
+ cost?: number | undefined;
186
+ error?: string | undefined;
187
+ logProbs?: number[] | undefined;
188
+ metadata?: import("zod").objectOutputType<{
189
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
190
+ }, import("zod").ZodAny, "strip"> | undefined;
191
+ tokenUsage?: {
192
+ prompt?: number | undefined;
193
+ cached?: number | undefined;
194
+ completion?: number | undefined;
195
+ total?: number | undefined;
196
+ } | undefined;
197
+ }>) & {
198
+ label?: string | undefined;
199
+ });
200
+ delay?: number | undefined;
201
+ label?: string | undefined;
202
+ transform?: string | undefined;
203
+ callEmbeddingApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
204
+ error?: string | undefined;
205
+ tokenUsage?: {
206
+ prompt?: number | undefined;
207
+ cached?: number | undefined;
208
+ completion?: number | undefined;
209
+ total?: number | undefined;
210
+ } | undefined;
211
+ embedding?: number[] | undefined;
212
+ }>) | undefined;
213
+ callClassificationApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
214
+ error?: string | undefined;
215
+ classification?: Record<string, number> | undefined;
216
+ }>) | undefined;
217
+ } | undefined;
218
+ }, ...args_1: unknown[]) => Promise<any>) | undefined;
219
+ id?: string | undefined;
220
+ display?: string | undefined;
221
+ })[];
222
+ providers: string | (((args_0: string, args_1: {
223
+ vars: Record<string, string | {}>;
224
+ logger?: any;
225
+ fetchWithCache?: any;
226
+ getCache?: any;
227
+ } | undefined, args_2: {
228
+ includeLogProbs?: boolean | undefined;
229
+ originalProvider?: any;
230
+ } | undefined, ...args_3: unknown[]) => Promise<{
231
+ output?: any;
232
+ cached?: boolean | undefined;
233
+ cost?: number | undefined;
234
+ error?: string | undefined;
235
+ logProbs?: number[] | undefined;
236
+ metadata?: import("zod").objectOutputType<{
237
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
238
+ }, import("zod").ZodAny, "strip"> | undefined;
239
+ tokenUsage?: {
240
+ prompt?: number | undefined;
241
+ cached?: number | undefined;
242
+ completion?: number | undefined;
243
+ total?: number | undefined;
244
+ } | undefined;
245
+ }>) & {
246
+ label?: string | undefined;
247
+ }) | (((args_0: string, args_1: {
248
+ includeLogProbs?: boolean | undefined;
249
+ originalProvider?: any;
250
+ } | undefined, ...args_2: unknown[]) => Promise<{
251
+ output?: any;
252
+ cached?: boolean | undefined;
253
+ cost?: number | undefined;
254
+ error?: string | undefined;
255
+ logProbs?: number[] | undefined;
256
+ metadata?: import("zod").objectOutputType<{
257
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
258
+ }, import("zod").ZodAny, "strip"> | undefined;
259
+ tokenUsage?: {
260
+ prompt?: number | undefined;
261
+ cached?: number | undefined;
262
+ completion?: number | undefined;
263
+ total?: number | undefined;
264
+ } | undefined;
265
+ }>) & {
266
+ label?: string | undefined;
267
+ }) | (string | {
268
+ prompts?: string[] | undefined;
269
+ delay?: number | undefined;
270
+ config?: any;
271
+ id?: string | undefined;
272
+ label?: string | undefined;
273
+ transform?: string | undefined;
274
+ env?: {
275
+ ANTHROPIC_API_KEY?: string | undefined;
276
+ BAM_API_KEY?: string | undefined;
277
+ BAM_API_HOST?: string | undefined;
278
+ AZURE_OPENAI_API_HOST?: string | undefined;
279
+ AZURE_OPENAI_API_KEY?: string | undefined;
280
+ AZURE_OPENAI_API_BASE_URL?: string | undefined;
281
+ AZURE_OPENAI_BASE_URL?: string | undefined;
282
+ AWS_BEDROCK_REGION?: string | undefined;
283
+ COHERE_API_KEY?: string | undefined;
284
+ OPENAI_API_KEY?: string | undefined;
285
+ OPENAI_API_HOST?: string | undefined;
286
+ OPENAI_API_BASE_URL?: string | undefined;
287
+ OPENAI_BASE_URL?: string | undefined;
288
+ OPENAI_ORGANIZATION?: string | undefined;
289
+ REPLICATE_API_KEY?: string | undefined;
290
+ REPLICATE_API_TOKEN?: string | undefined;
291
+ LOCALAI_BASE_URL?: string | undefined;
292
+ MISTRAL_API_HOST?: string | undefined;
293
+ MISTRAL_API_BASE_URL?: string | undefined;
294
+ PALM_API_KEY?: string | undefined;
295
+ PALM_API_HOST?: string | undefined;
296
+ GOOGLE_API_KEY?: string | undefined;
297
+ GOOGLE_API_HOST?: string | undefined;
298
+ VERTEX_API_KEY?: string | undefined;
299
+ VERTEX_API_HOST?: string | undefined;
300
+ VERTEX_PROJECT_ID?: string | undefined;
301
+ VERTEX_REGION?: string | undefined;
302
+ VERTEX_PUBLISHER?: string | undefined;
303
+ MISTRAL_API_KEY?: string | undefined;
304
+ CLOUDFLARE_API_KEY?: string | undefined;
305
+ CLOUDFLARE_ACCOUNT_ID?: string | undefined;
306
+ } | undefined;
307
+ } | (((args_0: string, args_1: {
308
+ vars: Record<string, string | {}>;
309
+ logger?: any;
310
+ fetchWithCache?: any;
311
+ getCache?: any;
312
+ } | undefined, args_2: {
313
+ includeLogProbs?: boolean | undefined;
314
+ originalProvider?: any;
315
+ } | undefined, ...args_3: unknown[]) => Promise<{
316
+ output?: any;
317
+ cached?: boolean | undefined;
318
+ cost?: number | undefined;
319
+ error?: string | undefined;
320
+ logProbs?: number[] | undefined;
321
+ metadata?: import("zod").objectOutputType<{
322
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
323
+ }, import("zod").ZodAny, "strip"> | undefined;
324
+ tokenUsage?: {
325
+ prompt?: number | undefined;
326
+ cached?: number | undefined;
327
+ completion?: number | undefined;
328
+ total?: number | undefined;
329
+ } | undefined;
330
+ }>) & {
331
+ label?: string | undefined;
332
+ }) | (((args_0: string, args_1: {
333
+ includeLogProbs?: boolean | undefined;
334
+ originalProvider?: any;
335
+ } | undefined, ...args_2: unknown[]) => Promise<{
336
+ output?: any;
337
+ cached?: boolean | undefined;
338
+ cost?: number | undefined;
339
+ error?: string | undefined;
340
+ logProbs?: number[] | undefined;
341
+ metadata?: import("zod").objectOutputType<{
342
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
343
+ }, import("zod").ZodAny, "strip"> | undefined;
344
+ tokenUsage?: {
345
+ prompt?: number | undefined;
346
+ cached?: number | undefined;
347
+ completion?: number | undefined;
348
+ total?: number | undefined;
349
+ } | undefined;
350
+ }>) & {
351
+ label?: string | undefined;
352
+ }) | Record<string, {
353
+ prompts?: string[] | undefined;
354
+ delay?: number | undefined;
355
+ config?: any;
356
+ id?: string | undefined;
357
+ label?: string | undefined;
358
+ transform?: string | undefined;
359
+ env?: {
360
+ ANTHROPIC_API_KEY?: string | undefined;
361
+ BAM_API_KEY?: string | undefined;
362
+ BAM_API_HOST?: string | undefined;
363
+ AZURE_OPENAI_API_HOST?: string | undefined;
364
+ AZURE_OPENAI_API_KEY?: string | undefined;
365
+ AZURE_OPENAI_API_BASE_URL?: string | undefined;
366
+ AZURE_OPENAI_BASE_URL?: string | undefined;
367
+ AWS_BEDROCK_REGION?: string | undefined;
368
+ COHERE_API_KEY?: string | undefined;
369
+ OPENAI_API_KEY?: string | undefined;
370
+ OPENAI_API_HOST?: string | undefined;
371
+ OPENAI_API_BASE_URL?: string | undefined;
372
+ OPENAI_BASE_URL?: string | undefined;
373
+ OPENAI_ORGANIZATION?: string | undefined;
374
+ REPLICATE_API_KEY?: string | undefined;
375
+ REPLICATE_API_TOKEN?: string | undefined;
376
+ LOCALAI_BASE_URL?: string | undefined;
377
+ MISTRAL_API_HOST?: string | undefined;
378
+ MISTRAL_API_BASE_URL?: string | undefined;
379
+ PALM_API_KEY?: string | undefined;
380
+ PALM_API_HOST?: string | undefined;
381
+ GOOGLE_API_KEY?: string | undefined;
382
+ GOOGLE_API_HOST?: string | undefined;
383
+ VERTEX_API_KEY?: string | undefined;
384
+ VERTEX_API_HOST?: string | undefined;
385
+ VERTEX_PROJECT_ID?: string | undefined;
386
+ VERTEX_REGION?: string | undefined;
387
+ VERTEX_PUBLISHER?: string | undefined;
388
+ MISTRAL_API_KEY?: string | undefined;
389
+ CLOUDFLARE_API_KEY?: string | undefined;
390
+ CLOUDFLARE_ACCOUNT_ID?: string | undefined;
391
+ } | undefined;
392
+ }>)[];
393
+ tests: string | (string | {
394
+ options?: ({
395
+ prefix?: string | undefined;
396
+ suffix?: string | undefined;
397
+ } & {
398
+ transform?: string | undefined;
399
+ postprocess?: string | undefined;
400
+ storeOutputAs?: string | undefined;
401
+ } & {
402
+ rubricPrompt?: string | string[] | undefined;
403
+ provider?: any;
404
+ factuality?: {
405
+ subset?: number | undefined;
406
+ superset?: number | undefined;
407
+ agree?: number | undefined;
408
+ disagree?: number | undefined;
409
+ differButFactual?: number | undefined;
410
+ } | undefined;
411
+ } & {
412
+ disableVarExpansion?: boolean | undefined;
413
+ disableConversationVar?: boolean | undefined;
414
+ }) | undefined;
415
+ vars?: Record<string, string | string[] | {}> | undefined;
416
+ provider?: string | {
417
+ prompts?: string[] | undefined;
418
+ delay?: number | undefined;
419
+ config?: any;
420
+ id?: string | undefined;
421
+ label?: string | undefined;
422
+ transform?: string | undefined;
423
+ env?: {
424
+ ANTHROPIC_API_KEY?: string | undefined;
425
+ BAM_API_KEY?: string | undefined;
426
+ BAM_API_HOST?: string | undefined;
427
+ AZURE_OPENAI_API_HOST?: string | undefined;
428
+ AZURE_OPENAI_API_KEY?: string | undefined;
429
+ AZURE_OPENAI_API_BASE_URL?: string | undefined;
430
+ AZURE_OPENAI_BASE_URL?: string | undefined;
431
+ AWS_BEDROCK_REGION?: string | undefined;
432
+ COHERE_API_KEY?: string | undefined;
433
+ OPENAI_API_KEY?: string | undefined;
434
+ OPENAI_API_HOST?: string | undefined;
435
+ OPENAI_API_BASE_URL?: string | undefined;
436
+ OPENAI_BASE_URL?: string | undefined;
437
+ OPENAI_ORGANIZATION?: string | undefined;
438
+ REPLICATE_API_KEY?: string | undefined;
439
+ REPLICATE_API_TOKEN?: string | undefined;
440
+ LOCALAI_BASE_URL?: string | undefined;
441
+ MISTRAL_API_HOST?: string | undefined;
442
+ MISTRAL_API_BASE_URL?: string | undefined;
443
+ PALM_API_KEY?: string | undefined;
444
+ PALM_API_HOST?: string | undefined;
445
+ GOOGLE_API_KEY?: string | undefined;
446
+ GOOGLE_API_HOST?: string | undefined;
447
+ VERTEX_API_KEY?: string | undefined;
448
+ VERTEX_API_HOST?: string | undefined;
449
+ VERTEX_PROJECT_ID?: string | undefined;
450
+ VERTEX_REGION?: string | undefined;
451
+ VERTEX_PUBLISHER?: string | undefined;
452
+ MISTRAL_API_KEY?: string | undefined;
453
+ CLOUDFLARE_API_KEY?: string | undefined;
454
+ CLOUDFLARE_ACCOUNT_ID?: string | undefined;
455
+ } | undefined;
456
+ } | {
457
+ id: (...args: unknown[]) => string;
458
+ callApi: (((args_0: string, args_1: {
459
+ vars: Record<string, string | {}>;
460
+ logger?: any;
461
+ fetchWithCache?: any;
462
+ getCache?: any;
463
+ } | undefined, args_2: {
464
+ includeLogProbs?: boolean | undefined;
465
+ originalProvider?: any;
466
+ } | undefined, ...args_3: unknown[]) => Promise<{
467
+ output?: any;
468
+ cached?: boolean | undefined;
469
+ cost?: number | undefined;
470
+ error?: string | undefined;
471
+ logProbs?: number[] | undefined;
472
+ metadata?: import("zod").objectOutputType<{
473
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
474
+ }, import("zod").ZodAny, "strip"> | undefined;
475
+ tokenUsage?: {
476
+ prompt?: number | undefined;
477
+ cached?: number | undefined;
478
+ completion?: number | undefined;
479
+ total?: number | undefined;
480
+ } | undefined;
481
+ }>) & {
482
+ label?: string | undefined;
483
+ }) | (((args_0: string, args_1: {
484
+ includeLogProbs?: boolean | undefined;
485
+ originalProvider?: any;
486
+ } | undefined, ...args_2: unknown[]) => Promise<{
487
+ output?: any;
488
+ cached?: boolean | undefined;
489
+ cost?: number | undefined;
490
+ error?: string | undefined;
491
+ logProbs?: number[] | undefined;
492
+ metadata?: import("zod").objectOutputType<{
493
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
494
+ }, import("zod").ZodAny, "strip"> | undefined;
495
+ tokenUsage?: {
496
+ prompt?: number | undefined;
497
+ cached?: number | undefined;
498
+ completion?: number | undefined;
499
+ total?: number | undefined;
500
+ } | undefined;
501
+ }>) & {
502
+ label?: string | undefined;
503
+ });
504
+ delay?: number | undefined;
505
+ label?: string | undefined;
506
+ transform?: string | undefined;
507
+ callEmbeddingApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
508
+ error?: string | undefined;
509
+ tokenUsage?: {
510
+ prompt?: number | undefined;
511
+ cached?: number | undefined;
512
+ completion?: number | undefined;
513
+ total?: number | undefined;
514
+ } | undefined;
515
+ embedding?: number[] | undefined;
516
+ }>) | undefined;
517
+ callClassificationApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
518
+ error?: string | undefined;
519
+ classification?: Record<string, number> | undefined;
520
+ }>) | undefined;
521
+ } | undefined;
522
+ description?: string | undefined;
523
+ providerOutput?: string | {} | undefined;
524
+ assert?: ({
525
+ type: import("./types").AssertionType;
526
+ value?: import("./types").AssertionValue | undefined;
527
+ transform?: string | undefined;
528
+ rubricPrompt?: string | string[] | undefined;
529
+ provider?: any;
530
+ threshold?: number | undefined;
531
+ weight?: number | undefined;
532
+ metric?: string | undefined;
533
+ } | {
534
+ type: "assert-set";
535
+ assert: {
536
+ type: import("./types").AssertionType;
537
+ value?: import("./types").AssertionValue | undefined;
538
+ transform?: string | undefined;
539
+ rubricPrompt?: string | string[] | undefined;
540
+ provider?: any;
541
+ threshold?: number | undefined;
542
+ weight?: number | undefined;
543
+ metric?: string | undefined;
544
+ }[];
545
+ threshold?: number | undefined;
546
+ weight?: number | undefined;
547
+ metric?: string | undefined;
548
+ })[] | undefined;
549
+ threshold?: number | undefined;
550
+ })[];
551
+ evaluateOptions: {
552
+ maxConcurrency?: number | undefined;
553
+ repeat?: number | undefined;
554
+ delay?: number | undefined;
555
+ cache?: boolean | undefined;
556
+ interactiveProviders?: boolean | undefined;
557
+ generateSuggestions?: boolean | undefined;
558
+ showProgressBar?: boolean | undefined;
559
+ progressCallback?: ((args_0: number, args_1: number, args_2: number, args_3: import("./types").RunEvalOptions) => void) | undefined;
560
+ eventSource?: string | undefined;
561
+ };
562
+ commandLineOptions: {
563
+ prompts?: string[] | undefined;
564
+ providers?: string[] | undefined;
565
+ output?: string[] | undefined;
566
+ maxConcurrency?: string | undefined;
567
+ repeat?: string | undefined;
568
+ delay?: string | undefined;
569
+ vars?: string | undefined;
570
+ tests?: string | undefined;
571
+ config?: string[] | undefined;
572
+ assertions?: string | undefined;
573
+ modelOutputs?: string | undefined;
574
+ verbose?: boolean | undefined;
575
+ grader?: string | undefined;
576
+ tableCellMaxLength?: string | undefined;
577
+ write?: boolean | undefined;
578
+ cache?: boolean | undefined;
579
+ table?: boolean | undefined;
580
+ share?: boolean | undefined;
581
+ progressBar?: boolean | undefined;
582
+ watch?: boolean | undefined;
583
+ interactiveProviders?: boolean | undefined;
584
+ filterFailing?: string | undefined;
585
+ filterFirstN?: string | undefined;
586
+ filterPattern?: string | undefined;
587
+ filterProviders?: string | undefined;
588
+ var?: Record<string, string> | undefined;
589
+ generateSuggestions?: boolean | undefined;
590
+ promptPrefix?: string | undefined;
591
+ promptSuffix?: string | undefined;
592
+ envFile?: string | undefined;
593
+ };
594
+ env?: {
595
+ ANTHROPIC_API_KEY?: string | undefined;
596
+ BAM_API_KEY?: string | undefined;
597
+ BAM_API_HOST?: string | undefined;
598
+ AZURE_OPENAI_API_HOST?: string | undefined;
599
+ AZURE_OPENAI_API_KEY?: string | undefined;
600
+ AZURE_OPENAI_API_BASE_URL?: string | undefined;
601
+ AZURE_OPENAI_BASE_URL?: string | undefined;
602
+ AWS_BEDROCK_REGION?: string | undefined;
603
+ COHERE_API_KEY?: string | undefined;
604
+ OPENAI_API_KEY?: string | undefined;
605
+ OPENAI_API_HOST?: string | undefined;
606
+ OPENAI_API_BASE_URL?: string | undefined;
607
+ OPENAI_BASE_URL?: string | undefined;
608
+ OPENAI_ORGANIZATION?: string | undefined;
609
+ REPLICATE_API_KEY?: string | undefined;
610
+ REPLICATE_API_TOKEN?: string | undefined;
611
+ LOCALAI_BASE_URL?: string | undefined;
612
+ MISTRAL_API_HOST?: string | undefined;
613
+ MISTRAL_API_BASE_URL?: string | undefined;
614
+ PALM_API_KEY?: string | undefined;
615
+ PALM_API_HOST?: string | undefined;
616
+ GOOGLE_API_KEY?: string | undefined;
617
+ GOOGLE_API_HOST?: string | undefined;
618
+ VERTEX_API_KEY?: string | undefined;
619
+ VERTEX_API_HOST?: string | undefined;
620
+ VERTEX_PROJECT_ID?: string | undefined;
621
+ VERTEX_REGION?: string | undefined;
622
+ VERTEX_PUBLISHER?: string | undefined;
623
+ MISTRAL_API_KEY?: string | undefined;
624
+ CLOUDFLARE_API_KEY?: string | undefined;
625
+ CLOUDFLARE_ACCOUNT_ID?: string | undefined;
626
+ } | undefined;
627
+ metadata?: Record<string, any> | undefined;
628
+ description?: string | undefined;
629
+ scenarios?: {
630
+ tests: {
631
+ options?: ({
632
+ prefix?: string | undefined;
633
+ suffix?: string | undefined;
634
+ } & {
635
+ transform?: string | undefined;
636
+ postprocess?: string | undefined;
637
+ storeOutputAs?: string | undefined;
638
+ } & {
639
+ rubricPrompt?: string | string[] | undefined;
640
+ provider?: any;
641
+ factuality?: {
642
+ subset?: number | undefined;
643
+ superset?: number | undefined;
644
+ agree?: number | undefined;
645
+ disagree?: number | undefined;
646
+ differButFactual?: number | undefined;
647
+ } | undefined;
648
+ } & {
649
+ disableVarExpansion?: boolean | undefined;
650
+ disableConversationVar?: boolean | undefined;
651
+ }) | undefined;
652
+ vars?: Record<string, string | string[] | {}> | undefined;
653
+ provider?: string | {
654
+ prompts?: string[] | undefined;
655
+ delay?: number | undefined;
656
+ config?: any;
657
+ id?: string | undefined;
658
+ label?: string | undefined;
659
+ transform?: string | undefined;
660
+ env?: {
661
+ ANTHROPIC_API_KEY?: string | undefined;
662
+ BAM_API_KEY?: string | undefined;
663
+ BAM_API_HOST?: string | undefined;
664
+ AZURE_OPENAI_API_HOST?: string | undefined;
665
+ AZURE_OPENAI_API_KEY?: string | undefined;
666
+ AZURE_OPENAI_API_BASE_URL?: string | undefined;
667
+ AZURE_OPENAI_BASE_URL?: string | undefined;
668
+ AWS_BEDROCK_REGION?: string | undefined;
669
+ COHERE_API_KEY?: string | undefined;
670
+ OPENAI_API_KEY?: string | undefined;
671
+ OPENAI_API_HOST?: string | undefined;
672
+ OPENAI_API_BASE_URL?: string | undefined;
673
+ OPENAI_BASE_URL?: string | undefined;
674
+ OPENAI_ORGANIZATION?: string | undefined;
675
+ REPLICATE_API_KEY?: string | undefined;
676
+ REPLICATE_API_TOKEN?: string | undefined;
677
+ LOCALAI_BASE_URL?: string | undefined;
678
+ MISTRAL_API_HOST?: string | undefined;
679
+ MISTRAL_API_BASE_URL?: string | undefined;
680
+ PALM_API_KEY?: string | undefined;
681
+ PALM_API_HOST?: string | undefined;
682
+ GOOGLE_API_KEY?: string | undefined;
683
+ GOOGLE_API_HOST?: string | undefined;
684
+ VERTEX_API_KEY?: string | undefined;
685
+ VERTEX_API_HOST?: string | undefined;
686
+ VERTEX_PROJECT_ID?: string | undefined;
687
+ VERTEX_REGION?: string | undefined;
688
+ VERTEX_PUBLISHER?: string | undefined;
689
+ MISTRAL_API_KEY?: string | undefined;
690
+ CLOUDFLARE_API_KEY?: string | undefined;
691
+ CLOUDFLARE_ACCOUNT_ID?: string | undefined;
692
+ } | undefined;
693
+ } | {
694
+ id: (...args: unknown[]) => string;
695
+ callApi: (((args_0: string, args_1: {
696
+ vars: Record<string, string | {}>;
697
+ logger?: any;
698
+ fetchWithCache?: any;
699
+ getCache?: any;
700
+ } | undefined, args_2: {
701
+ includeLogProbs?: boolean | undefined;
702
+ originalProvider?: any;
703
+ } | undefined, ...args_3: unknown[]) => Promise<{
704
+ output?: any;
705
+ cached?: boolean | undefined;
706
+ cost?: number | undefined;
707
+ error?: string | undefined;
708
+ logProbs?: number[] | undefined;
709
+ metadata?: import("zod").objectOutputType<{
710
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
711
+ }, import("zod").ZodAny, "strip"> | undefined;
712
+ tokenUsage?: {
713
+ prompt?: number | undefined;
714
+ cached?: number | undefined;
715
+ completion?: number | undefined;
716
+ total?: number | undefined;
717
+ } | undefined;
718
+ }>) & {
719
+ label?: string | undefined;
720
+ }) | (((args_0: string, args_1: {
721
+ includeLogProbs?: boolean | undefined;
722
+ originalProvider?: any;
723
+ } | undefined, ...args_2: unknown[]) => Promise<{
724
+ output?: any;
725
+ cached?: boolean | undefined;
726
+ cost?: number | undefined;
727
+ error?: string | undefined;
728
+ logProbs?: number[] | undefined;
729
+ metadata?: import("zod").objectOutputType<{
730
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
731
+ }, import("zod").ZodAny, "strip"> | undefined;
732
+ tokenUsage?: {
733
+ prompt?: number | undefined;
734
+ cached?: number | undefined;
735
+ completion?: number | undefined;
736
+ total?: number | undefined;
737
+ } | undefined;
738
+ }>) & {
739
+ label?: string | undefined;
740
+ });
741
+ delay?: number | undefined;
742
+ label?: string | undefined;
743
+ transform?: string | undefined;
744
+ callEmbeddingApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
745
+ error?: string | undefined;
746
+ tokenUsage?: {
747
+ prompt?: number | undefined;
748
+ cached?: number | undefined;
749
+ completion?: number | undefined;
750
+ total?: number | undefined;
751
+ } | undefined;
752
+ embedding?: number[] | undefined;
753
+ }>) | undefined;
754
+ callClassificationApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
755
+ error?: string | undefined;
756
+ classification?: Record<string, number> | undefined;
757
+ }>) | undefined;
758
+ } | undefined;
759
+ description?: string | undefined;
760
+ providerOutput?: string | {} | undefined;
761
+ assert?: ({
762
+ type: import("./types").AssertionType;
763
+ value?: import("./types").AssertionValue | undefined;
764
+ transform?: string | undefined;
765
+ rubricPrompt?: string | string[] | undefined;
766
+ provider?: any;
767
+ threshold?: number | undefined;
768
+ weight?: number | undefined;
769
+ metric?: string | undefined;
770
+ } | {
771
+ type: "assert-set";
772
+ assert: {
773
+ type: import("./types").AssertionType;
774
+ value?: import("./types").AssertionValue | undefined;
775
+ transform?: string | undefined;
776
+ rubricPrompt?: string | string[] | undefined;
777
+ provider?: any;
778
+ threshold?: number | undefined;
779
+ weight?: number | undefined;
780
+ metric?: string | undefined;
781
+ }[];
782
+ threshold?: number | undefined;
783
+ weight?: number | undefined;
784
+ metric?: string | undefined;
785
+ })[] | undefined;
786
+ threshold?: number | undefined;
787
+ }[];
788
+ config: {
789
+ options?: ({
790
+ prefix?: string | undefined;
791
+ suffix?: string | undefined;
792
+ } & {
793
+ transform?: string | undefined;
794
+ postprocess?: string | undefined;
795
+ storeOutputAs?: string | undefined;
796
+ } & {
797
+ rubricPrompt?: string | string[] | undefined;
798
+ provider?: any;
799
+ factuality?: {
800
+ subset?: number | undefined;
801
+ superset?: number | undefined;
802
+ agree?: number | undefined;
803
+ disagree?: number | undefined;
804
+ differButFactual?: number | undefined;
805
+ } | undefined;
806
+ } & {
807
+ disableVarExpansion?: boolean | undefined;
808
+ disableConversationVar?: boolean | undefined;
809
+ }) | undefined;
810
+ vars?: Record<string, string | string[] | {}> | undefined;
811
+ provider?: string | {
812
+ prompts?: string[] | undefined;
813
+ delay?: number | undefined;
814
+ config?: any;
815
+ id?: string | undefined;
816
+ label?: string | undefined;
817
+ transform?: string | undefined;
818
+ env?: {
819
+ ANTHROPIC_API_KEY?: string | undefined;
820
+ BAM_API_KEY?: string | undefined;
821
+ BAM_API_HOST?: string | undefined;
822
+ AZURE_OPENAI_API_HOST?: string | undefined;
823
+ AZURE_OPENAI_API_KEY?: string | undefined;
824
+ AZURE_OPENAI_API_BASE_URL?: string | undefined;
825
+ AZURE_OPENAI_BASE_URL?: string | undefined;
826
+ AWS_BEDROCK_REGION?: string | undefined;
827
+ COHERE_API_KEY?: string | undefined;
828
+ OPENAI_API_KEY?: string | undefined;
829
+ OPENAI_API_HOST?: string | undefined;
830
+ OPENAI_API_BASE_URL?: string | undefined;
831
+ OPENAI_BASE_URL?: string | undefined;
832
+ OPENAI_ORGANIZATION?: string | undefined;
833
+ REPLICATE_API_KEY?: string | undefined;
834
+ REPLICATE_API_TOKEN?: string | undefined;
835
+ LOCALAI_BASE_URL?: string | undefined;
836
+ MISTRAL_API_HOST?: string | undefined;
837
+ MISTRAL_API_BASE_URL?: string | undefined;
838
+ PALM_API_KEY?: string | undefined;
839
+ PALM_API_HOST?: string | undefined;
840
+ GOOGLE_API_KEY?: string | undefined;
841
+ GOOGLE_API_HOST?: string | undefined;
842
+ VERTEX_API_KEY?: string | undefined;
843
+ VERTEX_API_HOST?: string | undefined;
844
+ VERTEX_PROJECT_ID?: string | undefined;
845
+ VERTEX_REGION?: string | undefined;
846
+ VERTEX_PUBLISHER?: string | undefined;
847
+ MISTRAL_API_KEY?: string | undefined;
848
+ CLOUDFLARE_API_KEY?: string | undefined;
849
+ CLOUDFLARE_ACCOUNT_ID?: string | undefined;
850
+ } | undefined;
851
+ } | {
852
+ id: (...args: unknown[]) => string;
853
+ callApi: (((args_0: string, args_1: {
854
+ vars: Record<string, string | {}>;
855
+ logger?: any;
856
+ fetchWithCache?: any;
857
+ getCache?: any;
858
+ } | undefined, args_2: {
859
+ includeLogProbs?: boolean | undefined;
860
+ originalProvider?: any;
861
+ } | undefined, ...args_3: unknown[]) => Promise<{
862
+ output?: any;
863
+ cached?: boolean | undefined;
864
+ cost?: number | undefined;
865
+ error?: string | undefined;
866
+ logProbs?: number[] | undefined;
867
+ metadata?: import("zod").objectOutputType<{
868
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
869
+ }, import("zod").ZodAny, "strip"> | undefined;
870
+ tokenUsage?: {
871
+ prompt?: number | undefined;
872
+ cached?: number | undefined;
873
+ completion?: number | undefined;
874
+ total?: number | undefined;
875
+ } | undefined;
876
+ }>) & {
877
+ label?: string | undefined;
878
+ }) | (((args_0: string, args_1: {
879
+ includeLogProbs?: boolean | undefined;
880
+ originalProvider?: any;
881
+ } | undefined, ...args_2: unknown[]) => Promise<{
882
+ output?: any;
883
+ cached?: boolean | undefined;
884
+ cost?: number | undefined;
885
+ error?: string | undefined;
886
+ logProbs?: number[] | undefined;
887
+ metadata?: import("zod").objectOutputType<{
888
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
889
+ }, import("zod").ZodAny, "strip"> | undefined;
890
+ tokenUsage?: {
891
+ prompt?: number | undefined;
892
+ cached?: number | undefined;
893
+ completion?: number | undefined;
894
+ total?: number | undefined;
895
+ } | undefined;
896
+ }>) & {
897
+ label?: string | undefined;
898
+ });
899
+ delay?: number | undefined;
900
+ label?: string | undefined;
901
+ transform?: string | undefined;
902
+ callEmbeddingApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
903
+ error?: string | undefined;
904
+ tokenUsage?: {
905
+ prompt?: number | undefined;
906
+ cached?: number | undefined;
907
+ completion?: number | undefined;
908
+ total?: number | undefined;
909
+ } | undefined;
910
+ embedding?: number[] | undefined;
911
+ }>) | undefined;
912
+ callClassificationApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
913
+ error?: string | undefined;
914
+ classification?: Record<string, number> | undefined;
915
+ }>) | undefined;
916
+ } | undefined;
917
+ description?: string | undefined;
918
+ providerOutput?: string | {} | undefined;
919
+ assert?: ({
920
+ type: import("./types").AssertionType;
921
+ value?: import("./types").AssertionValue | undefined;
922
+ transform?: string | undefined;
923
+ rubricPrompt?: string | string[] | undefined;
924
+ provider?: any;
925
+ threshold?: number | undefined;
926
+ weight?: number | undefined;
927
+ metric?: string | undefined;
928
+ } | {
929
+ type: "assert-set";
930
+ assert: {
931
+ type: import("./types").AssertionType;
932
+ value?: import("./types").AssertionValue | undefined;
933
+ transform?: string | undefined;
934
+ rubricPrompt?: string | string[] | undefined;
935
+ provider?: any;
936
+ threshold?: number | undefined;
937
+ weight?: number | undefined;
938
+ metric?: string | undefined;
939
+ }[];
940
+ threshold?: number | undefined;
941
+ weight?: number | undefined;
942
+ metric?: string | undefined;
943
+ })[] | undefined;
944
+ threshold?: number | undefined;
945
+ }[];
946
+ description?: string | undefined;
947
+ }[] | undefined;
948
+ defaultTest?: {
949
+ options?: ({
950
+ prefix?: string | undefined;
951
+ suffix?: string | undefined;
952
+ } & {
953
+ transform?: string | undefined;
954
+ postprocess?: string | undefined;
955
+ storeOutputAs?: string | undefined;
956
+ } & {
957
+ rubricPrompt?: string | string[] | undefined;
958
+ provider?: any;
959
+ factuality?: {
960
+ subset?: number | undefined;
961
+ superset?: number | undefined;
962
+ agree?: number | undefined;
963
+ disagree?: number | undefined;
964
+ differButFactual?: number | undefined;
965
+ } | undefined;
966
+ } & {
967
+ disableVarExpansion?: boolean | undefined;
968
+ disableConversationVar?: boolean | undefined;
969
+ }) | undefined;
970
+ vars?: Record<string, string | string[] | {}> | undefined;
971
+ provider?: string | {
972
+ prompts?: string[] | undefined;
973
+ delay?: number | undefined;
974
+ config?: any;
975
+ id?: string | undefined;
976
+ label?: string | undefined;
977
+ transform?: string | undefined;
978
+ env?: {
979
+ ANTHROPIC_API_KEY?: string | undefined;
980
+ BAM_API_KEY?: string | undefined;
981
+ BAM_API_HOST?: string | undefined;
982
+ AZURE_OPENAI_API_HOST?: string | undefined;
983
+ AZURE_OPENAI_API_KEY?: string | undefined;
984
+ AZURE_OPENAI_API_BASE_URL?: string | undefined;
985
+ AZURE_OPENAI_BASE_URL?: string | undefined;
986
+ AWS_BEDROCK_REGION?: string | undefined;
987
+ COHERE_API_KEY?: string | undefined;
988
+ OPENAI_API_KEY?: string | undefined;
989
+ OPENAI_API_HOST?: string | undefined;
990
+ OPENAI_API_BASE_URL?: string | undefined;
991
+ OPENAI_BASE_URL?: string | undefined;
992
+ OPENAI_ORGANIZATION?: string | undefined;
993
+ REPLICATE_API_KEY?: string | undefined;
994
+ REPLICATE_API_TOKEN?: string | undefined;
995
+ LOCALAI_BASE_URL?: string | undefined;
996
+ MISTRAL_API_HOST?: string | undefined;
997
+ MISTRAL_API_BASE_URL?: string | undefined;
998
+ PALM_API_KEY?: string | undefined;
999
+ PALM_API_HOST?: string | undefined;
1000
+ GOOGLE_API_KEY?: string | undefined;
1001
+ GOOGLE_API_HOST?: string | undefined;
1002
+ VERTEX_API_KEY?: string | undefined;
1003
+ VERTEX_API_HOST?: string | undefined;
1004
+ VERTEX_PROJECT_ID?: string | undefined;
1005
+ VERTEX_REGION?: string | undefined;
1006
+ VERTEX_PUBLISHER?: string | undefined;
1007
+ MISTRAL_API_KEY?: string | undefined;
1008
+ CLOUDFLARE_API_KEY?: string | undefined;
1009
+ CLOUDFLARE_ACCOUNT_ID?: string | undefined;
1010
+ } | undefined;
1011
+ } | {
1012
+ id: (...args: unknown[]) => string;
1013
+ callApi: (((args_0: string, args_1: {
1014
+ vars: Record<string, string | {}>;
1015
+ logger?: any;
1016
+ fetchWithCache?: any;
1017
+ getCache?: any;
1018
+ } | undefined, args_2: {
1019
+ includeLogProbs?: boolean | undefined;
1020
+ originalProvider?: any;
1021
+ } | undefined, ...args_3: unknown[]) => Promise<{
1022
+ output?: any;
1023
+ cached?: boolean | undefined;
1024
+ cost?: number | undefined;
1025
+ error?: string | undefined;
1026
+ logProbs?: number[] | undefined;
1027
+ metadata?: import("zod").objectOutputType<{
1028
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
1029
+ }, import("zod").ZodAny, "strip"> | undefined;
1030
+ tokenUsage?: {
1031
+ prompt?: number | undefined;
1032
+ cached?: number | undefined;
1033
+ completion?: number | undefined;
1034
+ total?: number | undefined;
1035
+ } | undefined;
1036
+ }>) & {
1037
+ label?: string | undefined;
1038
+ }) | (((args_0: string, args_1: {
1039
+ includeLogProbs?: boolean | undefined;
1040
+ originalProvider?: any;
1041
+ } | undefined, ...args_2: unknown[]) => Promise<{
1042
+ output?: any;
1043
+ cached?: boolean | undefined;
1044
+ cost?: number | undefined;
1045
+ error?: string | undefined;
1046
+ logProbs?: number[] | undefined;
1047
+ metadata?: import("zod").objectOutputType<{
1048
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
1049
+ }, import("zod").ZodAny, "strip"> | undefined;
1050
+ tokenUsage?: {
1051
+ prompt?: number | undefined;
1052
+ cached?: number | undefined;
1053
+ completion?: number | undefined;
1054
+ total?: number | undefined;
1055
+ } | undefined;
1056
+ }>) & {
1057
+ label?: string | undefined;
1058
+ });
1059
+ delay?: number | undefined;
1060
+ label?: string | undefined;
1061
+ transform?: string | undefined;
1062
+ callEmbeddingApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
1063
+ error?: string | undefined;
1064
+ tokenUsage?: {
1065
+ prompt?: number | undefined;
1066
+ cached?: number | undefined;
1067
+ completion?: number | undefined;
1068
+ total?: number | undefined;
1069
+ } | undefined;
1070
+ embedding?: number[] | undefined;
1071
+ }>) | undefined;
1072
+ callClassificationApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
1073
+ error?: string | undefined;
1074
+ classification?: Record<string, number> | undefined;
1075
+ }>) | undefined;
1076
+ } | undefined;
1077
+ providerOutput?: string | {} | undefined;
1078
+ assert?: ({
1079
+ type: import("./types").AssertionType;
1080
+ value?: import("./types").AssertionValue | undefined;
1081
+ transform?: string | undefined;
1082
+ rubricPrompt?: string | string[] | undefined;
1083
+ provider?: any;
1084
+ threshold?: number | undefined;
1085
+ weight?: number | undefined;
1086
+ metric?: string | undefined;
1087
+ } | {
1088
+ type: "assert-set";
1089
+ assert: {
1090
+ type: import("./types").AssertionType;
1091
+ value?: import("./types").AssertionValue | undefined;
1092
+ transform?: string | undefined;
1093
+ rubricPrompt?: string | string[] | undefined;
1094
+ provider?: any;
1095
+ threshold?: number | undefined;
1096
+ weight?: number | undefined;
1097
+ metric?: string | undefined;
1098
+ }[];
1099
+ threshold?: number | undefined;
1100
+ weight?: number | undefined;
1101
+ metric?: string | undefined;
1102
+ })[] | undefined;
1103
+ threshold?: number | undefined;
1104
+ } | undefined;
1105
+ nunjucksFilters?: Record<string, string> | undefined;
1106
+ derivedMetrics?: {
1107
+ value: string | ((args_0: Record<string, number>, args_1: import("./types").RunEvalOptions, ...args_2: unknown[]) => number);
1108
+ name: string;
1109
+ }[] | undefined;
1110
+ outputPath?: string | string[] | undefined;
1111
+ sharing?: boolean | {
1112
+ apiBaseUrl?: string | undefined;
1113
+ appBaseUrl?: string | undefined;
1114
+ } | undefined;
1115
+ }>;
147
1116
  driverParam: string;
148
1117
  notNull: true;
149
1118
  hasDefault: false;
@@ -224,7 +1193,164 @@ export declare const datasets: import("drizzle-orm/sqlite-core").SQLiteTableWith
224
1193
  tableName: "datasets";
225
1194
  dataType: "json";
226
1195
  columnType: "SQLiteTextJson";
227
- data: string | (string | import("./types").TestCase<Record<string, string | object | string[]>>)[];
1196
+ data: string | (string | {
1197
+ options?: ({
1198
+ prefix?: string | undefined;
1199
+ suffix?: string | undefined;
1200
+ } & {
1201
+ transform?: string | undefined;
1202
+ postprocess?: string | undefined;
1203
+ storeOutputAs?: string | undefined;
1204
+ } & {
1205
+ rubricPrompt?: string | string[] | undefined;
1206
+ provider?: any;
1207
+ factuality?: {
1208
+ subset?: number | undefined;
1209
+ superset?: number | undefined;
1210
+ agree?: number | undefined;
1211
+ disagree?: number | undefined;
1212
+ differButFactual?: number | undefined;
1213
+ } | undefined;
1214
+ } & {
1215
+ disableVarExpansion?: boolean | undefined;
1216
+ disableConversationVar?: boolean | undefined;
1217
+ }) | undefined;
1218
+ vars?: Record<string, string | string[] | {}> | undefined;
1219
+ provider?: string | {
1220
+ prompts?: string[] | undefined;
1221
+ delay?: number | undefined;
1222
+ config?: any;
1223
+ id?: string | undefined;
1224
+ label?: string | undefined;
1225
+ transform?: string | undefined;
1226
+ env?: {
1227
+ ANTHROPIC_API_KEY?: string | undefined;
1228
+ BAM_API_KEY?: string | undefined;
1229
+ BAM_API_HOST?: string | undefined;
1230
+ AZURE_OPENAI_API_HOST?: string | undefined;
1231
+ AZURE_OPENAI_API_KEY?: string | undefined;
1232
+ AZURE_OPENAI_API_BASE_URL?: string | undefined;
1233
+ AZURE_OPENAI_BASE_URL?: string | undefined;
1234
+ AWS_BEDROCK_REGION?: string | undefined;
1235
+ COHERE_API_KEY?: string | undefined;
1236
+ OPENAI_API_KEY?: string | undefined;
1237
+ OPENAI_API_HOST?: string | undefined;
1238
+ OPENAI_API_BASE_URL?: string | undefined;
1239
+ OPENAI_BASE_URL?: string | undefined;
1240
+ OPENAI_ORGANIZATION?: string | undefined;
1241
+ REPLICATE_API_KEY?: string | undefined;
1242
+ REPLICATE_API_TOKEN?: string | undefined;
1243
+ LOCALAI_BASE_URL?: string | undefined;
1244
+ MISTRAL_API_HOST?: string | undefined;
1245
+ MISTRAL_API_BASE_URL?: string | undefined;
1246
+ PALM_API_KEY?: string | undefined;
1247
+ PALM_API_HOST?: string | undefined;
1248
+ GOOGLE_API_KEY?: string | undefined;
1249
+ GOOGLE_API_HOST?: string | undefined;
1250
+ VERTEX_API_KEY?: string | undefined;
1251
+ VERTEX_API_HOST?: string | undefined;
1252
+ VERTEX_PROJECT_ID?: string | undefined;
1253
+ VERTEX_REGION?: string | undefined;
1254
+ VERTEX_PUBLISHER?: string | undefined;
1255
+ MISTRAL_API_KEY?: string | undefined;
1256
+ CLOUDFLARE_API_KEY?: string | undefined;
1257
+ CLOUDFLARE_ACCOUNT_ID?: string | undefined;
1258
+ } | undefined;
1259
+ } | {
1260
+ id: (...args: unknown[]) => string;
1261
+ callApi: (((args_0: string, args_1: {
1262
+ vars: Record<string, string | {}>;
1263
+ logger?: any;
1264
+ fetchWithCache?: any;
1265
+ getCache?: any;
1266
+ } | undefined, args_2: {
1267
+ includeLogProbs?: boolean | undefined;
1268
+ originalProvider?: any;
1269
+ } | undefined, ...args_3: unknown[]) => Promise<{
1270
+ output?: any;
1271
+ cached?: boolean | undefined;
1272
+ cost?: number | undefined;
1273
+ error?: string | undefined;
1274
+ logProbs?: number[] | undefined;
1275
+ metadata?: import("zod").objectOutputType<{
1276
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
1277
+ }, import("zod").ZodAny, "strip"> | undefined;
1278
+ tokenUsage?: {
1279
+ prompt?: number | undefined;
1280
+ cached?: number | undefined;
1281
+ completion?: number | undefined;
1282
+ total?: number | undefined;
1283
+ } | undefined;
1284
+ }>) & {
1285
+ label?: string | undefined;
1286
+ }) | (((args_0: string, args_1: {
1287
+ includeLogProbs?: boolean | undefined;
1288
+ originalProvider?: any;
1289
+ } | undefined, ...args_2: unknown[]) => Promise<{
1290
+ output?: any;
1291
+ cached?: boolean | undefined;
1292
+ cost?: number | undefined;
1293
+ error?: string | undefined;
1294
+ logProbs?: number[] | undefined;
1295
+ metadata?: import("zod").objectOutputType<{
1296
+ redteamFinalPrompt: import("zod").ZodOptional<import("zod").ZodString>;
1297
+ }, import("zod").ZodAny, "strip"> | undefined;
1298
+ tokenUsage?: {
1299
+ prompt?: number | undefined;
1300
+ cached?: number | undefined;
1301
+ completion?: number | undefined;
1302
+ total?: number | undefined;
1303
+ } | undefined;
1304
+ }>) & {
1305
+ label?: string | undefined;
1306
+ });
1307
+ delay?: number | undefined;
1308
+ label?: string | undefined;
1309
+ transform?: string | undefined;
1310
+ callEmbeddingApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
1311
+ error?: string | undefined;
1312
+ tokenUsage?: {
1313
+ prompt?: number | undefined;
1314
+ cached?: number | undefined;
1315
+ completion?: number | undefined;
1316
+ total?: number | undefined;
1317
+ } | undefined;
1318
+ embedding?: number[] | undefined;
1319
+ }>) | undefined;
1320
+ callClassificationApi?: ((args_0: string, ...args_1: unknown[]) => Promise<{
1321
+ error?: string | undefined;
1322
+ classification?: Record<string, number> | undefined;
1323
+ }>) | undefined;
1324
+ } | undefined;
1325
+ description?: string | undefined;
1326
+ providerOutput?: string | {} | undefined;
1327
+ assert?: ({
1328
+ type: import("./types").AssertionType;
1329
+ value?: import("./types").AssertionValue | undefined;
1330
+ transform?: string | undefined;
1331
+ rubricPrompt?: string | string[] | undefined;
1332
+ provider?: any;
1333
+ threshold?: number | undefined;
1334
+ weight?: number | undefined;
1335
+ metric?: string | undefined;
1336
+ } | {
1337
+ type: "assert-set";
1338
+ assert: {
1339
+ type: import("./types").AssertionType;
1340
+ value?: import("./types").AssertionValue | undefined;
1341
+ transform?: string | undefined;
1342
+ rubricPrompt?: string | string[] | undefined;
1343
+ provider?: any;
1344
+ threshold?: number | undefined;
1345
+ weight?: number | undefined;
1346
+ metric?: string | undefined;
1347
+ }[];
1348
+ threshold?: number | undefined;
1349
+ weight?: number | undefined;
1350
+ metric?: string | undefined;
1351
+ })[] | undefined;
1352
+ threshold?: number | undefined;
1353
+ })[];
228
1354
  driverParam: string;
229
1355
  notNull: false;
230
1356
  hasDefault: false;