@anvia/core 0.4.0 → 0.4.2

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 (50) hide show
  1. package/dist/agent/index.d.ts +4 -4
  2. package/dist/agent/index.js +6 -5
  3. package/dist/{agent-0UeJ9Rad.d.ts → agent-B-ls5y_g.d.ts} +2 -15
  4. package/dist/{chunk-I2HOMD3R.js → chunk-35GF7P43.js} +3 -3
  5. package/dist/{chunk-TILE6Z2N.js → chunk-DUP7FMAF.js} +2 -2
  6. package/dist/{chunk-A7VDIZQN.js → chunk-EFGX3EX5.js} +5 -14
  7. package/dist/chunk-EFGX3EX5.js.map +1 -0
  8. package/dist/{chunk-PP4VIN3Y.js → chunk-LGETU3RG.js} +3 -3
  9. package/dist/{chunk-6U6PZ5MD.js → chunk-MOICCK3J.js} +6 -15
  10. package/dist/chunk-MOICCK3J.js.map +1 -0
  11. package/dist/{chunk-TP32W7XT.js → chunk-N7NMSGZI.js} +597 -555
  12. package/dist/chunk-N7NMSGZI.js.map +1 -0
  13. package/dist/chunk-OIMLU4SF.js +20 -0
  14. package/dist/chunk-OIMLU4SF.js.map +1 -0
  15. package/dist/{chunk-65QV627O.js → chunk-P425B6GR.js} +31 -2
  16. package/dist/chunk-P425B6GR.js.map +1 -0
  17. package/dist/embeddings/index.d.ts +4 -24
  18. package/dist/embeddings/index.js +2 -1
  19. package/dist/evals/index.d.ts +23 -19
  20. package/dist/evals/index.js +92 -90
  21. package/dist/evals/index.js.map +1 -1
  22. package/dist/extractor/index.d.ts +3 -3
  23. package/dist/extractor/index.js +7 -6
  24. package/dist/index.d.ts +3 -3
  25. package/dist/index.js +7 -6
  26. package/dist/internal/agent.d.ts +3 -3
  27. package/dist/internal/agent.js +5 -4
  28. package/dist/loaders/index.d.ts +20 -12
  29. package/dist/loaders/index.js +107 -96
  30. package/dist/loaders/index.js.map +1 -1
  31. package/dist/mcp/index.js +10 -1
  32. package/dist/mcp/index.js.map +1 -1
  33. package/dist/{middleware-BQ7fkEEe.d.ts → middleware-CGiEIaBx.d.ts} +1 -1
  34. package/dist/pipeline/index.d.ts +5 -3
  35. package/dist/pipeline/index.js +156 -160
  36. package/dist/pipeline/index.js.map +1 -1
  37. package/dist/skills/index.js +5 -4
  38. package/dist/tool/index.d.ts +2 -2
  39. package/dist/tool/index.js +4 -3
  40. package/dist/types-IB2e9u5M.d.ts +25 -0
  41. package/dist/vector-store/index.d.ts +4 -1
  42. package/dist/vector-store/index.js +3 -2
  43. package/package.json +1 -1
  44. package/dist/chunk-65QV627O.js.map +0 -1
  45. package/dist/chunk-6U6PZ5MD.js.map +0 -1
  46. package/dist/chunk-A7VDIZQN.js.map +0 -1
  47. package/dist/chunk-TP32W7XT.js.map +0 -1
  48. /package/dist/{chunk-I2HOMD3R.js.map → chunk-35GF7P43.js.map} +0 -0
  49. /package/dist/{chunk-TILE6Z2N.js.map → chunk-DUP7FMAF.js.map} +0 -0
  50. /package/dist/{chunk-PP4VIN3Y.js.map → chunk-LGETU3RG.js.map} +0 -0
@@ -1,21 +1,62 @@
1
1
  import {
2
2
  ExtractorBuilder
3
- } from "../chunk-PP4VIN3Y.js";
4
- import "../chunk-6U6PZ5MD.js";
5
- import "../chunk-TP32W7XT.js";
3
+ } from "../chunk-LGETU3RG.js";
4
+ import "../chunk-MOICCK3J.js";
5
+ import "../chunk-N7NMSGZI.js";
6
6
  import "../chunk-XUUY2L2D.js";
7
7
  import "../chunk-XXT2UCAR.js";
8
8
  import "../chunk-YK4WAAS4.js";
9
- import "../chunk-I2HOMD3R.js";
10
- import "../chunk-65QV627O.js";
9
+ import "../chunk-35GF7P43.js";
10
+ import "../chunk-P425B6GR.js";
11
+ import "../chunk-6GJDBBDC.js";
11
12
  import {
12
13
  cosineSimilarity,
13
14
  embedText
14
- } from "../chunk-A7VDIZQN.js";
15
- import "../chunk-6GJDBBDC.js";
15
+ } from "../chunk-EFGX3EX5.js";
16
+ import {
17
+ mapWithConcurrency
18
+ } from "../chunk-OIMLU4SF.js";
16
19
 
17
- // src/evals/index.ts
20
+ // src/evals/agent-target.ts
21
+ function agentEvalTarget(agent, options = {}) {
22
+ return async (input, testCase) => {
23
+ const prompt = options.prompt?.(input, testCase) ?? String(input);
24
+ const response = await agent.prompt(prompt).send();
25
+ return options.output === void 0 ? response : options.output(response, testCase);
26
+ };
27
+ }
28
+
29
+ // src/evals/metrics.ts
18
30
  import { z } from "zod";
31
+
32
+ // src/evals/format.ts
33
+ function defaultOutputValue(output) {
34
+ if (typeof output === "object" && output !== null && "output" in output && typeof output.output === "string") {
35
+ return output.output;
36
+ }
37
+ return output;
38
+ }
39
+ function stableComparable(value) {
40
+ if (typeof value === "string") {
41
+ return value;
42
+ }
43
+ return JSON.stringify(value);
44
+ }
45
+ function formatValue(value) {
46
+ if (typeof value === "string") {
47
+ return value;
48
+ }
49
+ try {
50
+ return JSON.stringify(value);
51
+ } catch {
52
+ return String(value);
53
+ }
54
+ }
55
+ function errorMessage(error) {
56
+ return error instanceof Error ? error.message : String(error);
57
+ }
58
+
59
+ // src/evals/outcome.ts
19
60
  var EvalOutcome = {
20
61
  pass(score, options = {}) {
21
62
  return {
@@ -43,21 +84,35 @@ var EvalOutcome = {
43
84
  };
44
85
  }
45
86
  };
46
- async function runEvalSuite(options) {
47
- const startedAt = Date.now();
48
- const results = await mapWithConcurrency(
49
- options.cases,
50
- Math.max(1, Math.trunc(options.concurrency ?? 1)),
51
- (testCase) => runEvalCase(options, testCase)
52
- );
53
- const counts = countOutcomes(results);
54
- return {
55
- name: options.name,
56
- results,
57
- ...counts,
58
- durationMs: Date.now() - startedAt
59
- };
87
+
88
+ // src/evals/selectors.ts
89
+ async function resolveActual(selector, args) {
90
+ return selector === void 0 ? defaultOutputValue(args.output) : selector(args);
91
+ }
92
+ async function resolveActualText(selector, args) {
93
+ const value = selector === void 0 ? defaultOutputValue(args.output) : await selector(args);
94
+ return typeof value === "string" ? value : JSON.stringify(value);
95
+ }
96
+ async function resolveExpected(selectorOrValue, args) {
97
+ if (selectorOrValue === void 0) {
98
+ return args.case.expected;
99
+ }
100
+ return typeof selectorOrValue === "function" ? selectorOrValue(args) : selectorOrValue;
101
+ }
102
+ async function resolveJudgePrompt(selector, args) {
103
+ if (selector !== void 0) {
104
+ return selector(args);
105
+ }
106
+ return [
107
+ `Suite: ${args.suiteName}`,
108
+ `Case: ${args.case.id}`,
109
+ `Input: ${formatValue(args.case.input)}`,
110
+ `Expected: ${formatValue(args.case.expected)}`,
111
+ `Output: ${formatValue(defaultOutputValue(args.output))}`
112
+ ].join("\n\n");
60
113
  }
114
+
115
+ // src/evals/metrics.ts
61
116
  function exactMatch(options = {}) {
62
117
  return {
63
118
  name: options.name ?? "exact_match",
@@ -157,11 +212,21 @@ Return a score between 0 and 1 and brief feedback.`
157
212
  }
158
213
  };
159
214
  }
160
- function agentEvalTarget(agent, options = {}) {
161
- return async (input, testCase) => {
162
- const prompt = options.prompt?.(input, testCase) ?? String(input);
163
- const response = await agent.prompt(prompt).send();
164
- return options.output === void 0 ? response : options.output(response, testCase);
215
+
216
+ // src/evals/runner.ts
217
+ async function runEvalSuite(options) {
218
+ const startedAt = Date.now();
219
+ const results = await mapWithConcurrency(
220
+ options.cases,
221
+ Math.max(1, Math.trunc(options.concurrency ?? 1)),
222
+ (testCase) => runEvalCase(options, testCase)
223
+ );
224
+ const counts = countOutcomes(results);
225
+ return {
226
+ name: options.name,
227
+ results,
228
+ ...counts,
229
+ durationMs: Date.now() - startedAt
165
230
  };
166
231
  }
167
232
  async function runEvalCase(options, testCase) {
@@ -235,69 +300,6 @@ function countOutcomes(results) {
235
300
  }
236
301
  return { passed, failed, invalid };
237
302
  }
238
- async function resolveActual(selector, args) {
239
- return selector === void 0 ? defaultOutputValue(args.output) : selector(args);
240
- }
241
- async function resolveActualText(selector, args) {
242
- const value = selector === void 0 ? defaultOutputValue(args.output) : await selector(args);
243
- return typeof value === "string" ? value : JSON.stringify(value);
244
- }
245
- async function resolveExpected(selectorOrValue, args) {
246
- if (selectorOrValue === void 0) {
247
- return args.case.expected;
248
- }
249
- return typeof selectorOrValue === "function" ? selectorOrValue(args) : selectorOrValue;
250
- }
251
- async function resolveJudgePrompt(selector, args) {
252
- if (selector !== void 0) {
253
- return selector(args);
254
- }
255
- return [
256
- `Suite: ${args.suiteName}`,
257
- `Case: ${args.case.id}`,
258
- `Input: ${formatValue(args.case.input)}`,
259
- `Expected: ${formatValue(args.case.expected)}`,
260
- `Output: ${formatValue(defaultOutputValue(args.output))}`
261
- ].join("\n\n");
262
- }
263
- function defaultOutputValue(output) {
264
- if (typeof output === "object" && output !== null && "output" in output && typeof output.output === "string") {
265
- return output.output;
266
- }
267
- return output;
268
- }
269
- function stableComparable(value) {
270
- if (typeof value === "string") {
271
- return value;
272
- }
273
- return JSON.stringify(value);
274
- }
275
- function formatValue(value) {
276
- if (typeof value === "string") {
277
- return value;
278
- }
279
- try {
280
- return JSON.stringify(value);
281
- } catch {
282
- return String(value);
283
- }
284
- }
285
- function errorMessage(error) {
286
- return error instanceof Error ? error.message : String(error);
287
- }
288
- async function mapWithConcurrency(inputs, concurrency, mapper) {
289
- const results = new Array(inputs.length);
290
- let next = 0;
291
- async function worker() {
292
- while (next < inputs.length) {
293
- const index = next;
294
- next += 1;
295
- results[index] = await mapper(inputs[index]);
296
- }
297
- }
298
- await Promise.all(Array.from({ length: Math.min(concurrency, inputs.length) }, () => worker()));
299
- return results;
300
- }
301
303
  export {
302
304
  EvalOutcome,
303
305
  agentEvalTarget,
@@ -1 +1 @@
1
- {"version":3,"sources":["../../src/evals/index.ts"],"sourcesContent":["import { z } from \"zod\";\nimport type { Agent } from \"../agent/agent\";\nimport type { PromptResponse } from \"../agent/request\";\nimport type { CompletionModel, JsonValue, Message } from \"../completion\";\nimport { cosineSimilarity, type EmbeddingModel, embedText } from \"../embeddings\";\nimport { ExtractorBuilder } from \"../extractor\";\nimport type { ZodSchema } from \"../schema\";\n\nexport type EvalMetadata = Record<string, JsonValue | undefined>;\n\nexport type EvalCase<Input, Expected = unknown> = {\n id: string;\n input: Input;\n expected?: Expected | undefined;\n metadata?: EvalMetadata | undefined;\n};\n\nexport type EvalTarget<Input, Output, Expected = unknown> = (\n input: Input,\n testCase: EvalCase<Input, Expected>,\n) => Output | Promise<Output>;\n\nexport type EvalOutcomeStatus = \"pass\" | \"fail\" | \"invalid\";\n\nexport type EvalOutcome<Score = unknown> =\n | {\n outcome: \"pass\";\n score?: Score | undefined;\n comment?: string | undefined;\n metadata?: EvalMetadata | undefined;\n }\n | {\n outcome: \"fail\";\n score?: Score | undefined;\n comment?: string | undefined;\n metadata?: EvalMetadata | undefined;\n }\n | {\n outcome: \"invalid\";\n reason: string;\n score?: Score | undefined;\n comment?: string | undefined;\n metadata?: EvalMetadata | undefined;\n };\n\nexport const EvalOutcome = {\n pass<Score>(\n score?: Score,\n options: { comment?: string | undefined; metadata?: EvalMetadata | undefined } = {},\n ): EvalOutcome<Score> {\n return {\n outcome: \"pass\",\n ...(score === undefined ? {} : { score }),\n ...(options.comment === undefined ? {} : { comment: options.comment }),\n ...(options.metadata === undefined ? {} : { metadata: options.metadata }),\n };\n },\n\n fail<Score>(\n score?: Score,\n options: { comment?: string | undefined; metadata?: EvalMetadata | undefined } = {},\n ): EvalOutcome<Score> {\n return {\n outcome: \"fail\",\n ...(score === undefined ? {} : { score }),\n ...(options.comment === undefined ? {} : { comment: options.comment }),\n ...(options.metadata === undefined ? {} : { metadata: options.metadata }),\n };\n },\n\n invalid<Score = never>(\n reason: string,\n options: {\n score?: Score | undefined;\n comment?: string | undefined;\n metadata?: EvalMetadata | undefined;\n } = {},\n ): EvalOutcome<Score> {\n return {\n outcome: \"invalid\",\n reason,\n ...(options.score === undefined ? {} : { score: options.score }),\n ...(options.comment === undefined ? {} : { comment: options.comment }),\n ...(options.metadata === undefined ? {} : { metadata: options.metadata }),\n };\n },\n};\n\nexport type EvalMetricArgs<Input, Output, Expected = unknown> = {\n suiteName: string;\n case: EvalCase<Input, Expected>;\n output: Output;\n};\n\nexport type EvalMetric<Input, Output, Score = unknown, Expected = unknown> = {\n name: string;\n evaluate(\n args: EvalMetricArgs<Input, Output, Expected>,\n ): EvalOutcome<Score> | Promise<EvalOutcome<Score>>;\n};\n\nexport type EvalMetricResult<Score = unknown> = {\n metricName: string;\n outcome: EvalOutcome<Score>;\n reporterErrors: unknown[];\n};\n\nexport type EvalCaseResult<Input, Output, Expected = unknown> = {\n case: EvalCase<Input, Expected>;\n output?: Output | undefined;\n targetError?: unknown;\n metrics: EvalMetricResult[];\n};\n\nexport type EvalSuiteResult<Input, Output, Expected = unknown> = {\n name: string;\n results: Array<EvalCaseResult<Input, Output, Expected>>;\n passed: number;\n failed: number;\n invalid: number;\n durationMs: number;\n};\n\nexport type EvalReportArgs<Input, Output, Score = unknown, Expected = unknown> = {\n suiteName: string;\n case: EvalCase<Input, Expected>;\n output?: Output | undefined;\n targetError?: unknown;\n metric: EvalMetric<Input, Output, Score, Expected>;\n outcome: EvalOutcome<Score>;\n};\n\nexport type EvalReporter<Input = unknown, Output = unknown, Expected = unknown> = {\n report(args: EvalReportArgs<Input, Output, unknown, Expected>): void | Promise<void>;\n};\n\nexport type RunEvalSuiteOptions<Input, Output, Expected = unknown> = {\n name: string;\n cases: Array<EvalCase<Input, Expected>>;\n target: EvalTarget<Input, Output, Expected>;\n metrics: Array<EvalMetric<NoInfer<Input>, NoInfer<Output>, unknown, NoInfer<Expected>>>;\n concurrency?: number | undefined;\n reporters?: Array<EvalReporter<NoInfer<Input>, NoInfer<Output>, NoInfer<Expected>>> | undefined;\n failOnReporterError?: boolean | undefined;\n};\n\nexport async function runEvalSuite<Input, Output, Expected = unknown>(\n options: RunEvalSuiteOptions<Input, Output, Expected>,\n): Promise<EvalSuiteResult<Input, Output, Expected>> {\n const startedAt = Date.now();\n const results = await mapWithConcurrency(\n options.cases,\n Math.max(1, Math.trunc(options.concurrency ?? 1)),\n (testCase) => runEvalCase(options, testCase),\n );\n const counts = countOutcomes(results);\n return {\n name: options.name,\n results,\n ...counts,\n durationMs: Date.now() - startedAt,\n };\n}\n\nexport type ValueSelector<Input, Output, Expected, Value> = (\n args: EvalMetricArgs<Input, Output, Expected>,\n) => Value | Promise<Value>;\n\nexport type SelectorOrValue<Input, Output, Expected, Value> =\n | Value\n | ValueSelector<Input, Output, Expected, Value>;\n\nexport type ExactMatchOptions<Input, Output, Expected = unknown> = {\n name?: string | undefined;\n actual?: ValueSelector<Input, Output, Expected, unknown> | undefined;\n expected?: SelectorOrValue<Input, Output, Expected, unknown> | undefined;\n};\n\nexport function exactMatch<Input, Output, Expected = unknown>(\n options: ExactMatchOptions<Input, Output, Expected> = {},\n): EvalMetric<Input, Output, boolean, Expected> {\n return {\n name: options.name ?? \"exact_match\",\n async evaluate(args) {\n const actual = await resolveActual(options.actual, args);\n const expected = await resolveExpected(options.expected, args);\n if (expected === undefined) {\n return EvalOutcome.invalid(\"No expected value provided for exact match.\");\n }\n const passed = stableComparable(actual) === stableComparable(expected);\n return passed\n ? EvalOutcome.pass(true)\n : EvalOutcome.fail(false, { comment: `Expected ${formatValue(expected)}.` });\n },\n };\n}\n\nexport type ContainsOptions<Input, Output, Expected = unknown> = {\n name?: string | undefined;\n actual?: ValueSelector<Input, Output, Expected, string> | undefined;\n expected?: SelectorOrValue<Input, Output, Expected, string | RegExp> | undefined;\n};\n\nexport function contains<Input, Output, Expected = unknown>(\n options: ContainsOptions<Input, Output, Expected> = {},\n): EvalMetric<Input, Output, boolean, Expected> {\n return {\n name: options.name ?? \"contains\",\n async evaluate(args) {\n const actual = await resolveActualText(options.actual, args);\n const expected = await resolveExpected(options.expected, args);\n if (expected === undefined) {\n return EvalOutcome.invalid(\"No expected value provided for contains.\");\n }\n if (typeof expected !== \"string\" && !(expected instanceof RegExp)) {\n return EvalOutcome.invalid(\"Contains expected value must be a string or RegExp.\");\n }\n const passed = expected instanceof RegExp ? expected.test(actual) : actual.includes(expected);\n return passed\n ? EvalOutcome.pass(true)\n : EvalOutcome.fail(false, { comment: `Output did not contain ${String(expected)}.` });\n },\n };\n}\n\nexport type SemanticSimilarityOptions<Input, Output, Expected = unknown> = {\n name?: string | undefined;\n model: EmbeddingModel;\n threshold: number;\n actual?: ValueSelector<Input, Output, Expected, string> | undefined;\n expected?: SelectorOrValue<Input, Output, Expected, string> | undefined;\n};\n\nexport function semanticSimilarity<Input, Output, Expected = unknown>(\n options: SemanticSimilarityOptions<Input, Output, Expected>,\n): EvalMetric<Input, Output, number, Expected> {\n return {\n name: options.name ?? \"semantic_similarity\",\n async evaluate(args) {\n const actual = await resolveActualText(options.actual, args);\n const expected = await resolveExpected(options.expected, args);\n if (expected === undefined) {\n return EvalOutcome.invalid(\"No expected value provided for semantic similarity.\");\n }\n if (typeof expected !== \"string\") {\n return EvalOutcome.invalid(\"Semantic similarity expected value must be a string.\");\n }\n const [actualEmbedding, expectedEmbedding] = await Promise.all([\n embedText(options.model, actual),\n embedText(options.model, expected),\n ]);\n const score = cosineSimilarity(actualEmbedding.vector, expectedEmbedding.vector);\n return score >= options.threshold\n ? EvalOutcome.pass(score)\n : EvalOutcome.fail(score, { comment: `Similarity below threshold ${options.threshold}.` });\n },\n };\n}\n\nexport type LlmJudgeOptions<Input, Output, SchemaOutput, Expected = unknown> = {\n name?: string | undefined;\n model: CompletionModel;\n schema: ZodSchema<SchemaOutput>;\n passes(value: SchemaOutput): boolean;\n instructions?: string | undefined;\n retries?: number | undefined;\n prompt?: ValueSelector<Input, Output, Expected, string> | undefined;\n};\n\nexport function llmJudge<Input, Output, SchemaOutput, Expected = unknown>(\n options: LlmJudgeOptions<Input, Output, SchemaOutput, Expected>,\n): EvalMetric<Input, Output, SchemaOutput, Expected> {\n const extractor = new ExtractorBuilder(options.model, options.schema)\n .instructions(\n options.instructions ??\n \"Judge the eval case by the requested schema. Submit the judgment using the schema.\",\n )\n .retries(options.retries ?? 0)\n .build();\n\n return {\n name: options.name ?? \"llm_judge\",\n async evaluate(args) {\n try {\n const judgment = await extractor.extract(await resolveJudgePrompt(options.prompt, args));\n return options.passes(judgment) ? EvalOutcome.pass(judgment) : EvalOutcome.fail(judgment);\n } catch (error) {\n return EvalOutcome.invalid(errorMessage(error));\n }\n },\n };\n}\n\nexport type LlmScoreMetricScore = {\n score: number;\n feedback: string;\n};\n\nexport type LlmScoreOptions<Input, Output, Expected = unknown> = {\n name?: string | undefined;\n model: CompletionModel;\n threshold: number;\n criteria: string | string[];\n instructions?: string | undefined;\n retries?: number | undefined;\n prompt?: ValueSelector<Input, Output, Expected, string> | undefined;\n};\n\nexport function llmScore<Input, Output, Expected = unknown>(\n options: LlmScoreOptions<Input, Output, Expected>,\n): EvalMetric<Input, Output, LlmScoreMetricScore, Expected> {\n const criteria = Array.isArray(options.criteria) ? options.criteria.join(\"\\n\") : options.criteria;\n const extractor = new ExtractorBuilder(\n options.model,\n z.object({\n score: z.number(),\n feedback: z.string(),\n }),\n )\n .instructions(\n options.instructions ??\n `Score the eval case against these criteria:\\n${criteria}\\n\\nReturn a score between 0 and 1 and brief feedback.`,\n )\n .retries(options.retries ?? 0)\n .build();\n\n return {\n name: options.name ?? \"llm_score\",\n async evaluate(args) {\n try {\n const score = await extractor.extract(await resolveJudgePrompt(options.prompt, args));\n if (score.score < 0 || score.score > 1) {\n return EvalOutcome.invalid(`Score ${score.score} outside valid range [0, 1].`, {\n score,\n });\n }\n return score.score >= options.threshold\n ? EvalOutcome.pass(score, { comment: score.feedback })\n : EvalOutcome.fail(score, { comment: score.feedback });\n } catch (error) {\n return EvalOutcome.invalid(errorMessage(error));\n }\n },\n };\n}\n\nexport type AgentEvalTargetOptions<Input, Output = PromptResponse> = {\n prompt?: ((input: Input, testCase: EvalCase<Input>) => string | Message) | undefined;\n output?: ((response: PromptResponse, testCase: EvalCase<Input>) => Output) | undefined;\n};\n\nexport function agentEvalTarget<Input>(\n agent: Agent,\n options?: AgentEvalTargetOptions<Input, PromptResponse>,\n): EvalTarget<Input, PromptResponse>;\nexport function agentEvalTarget<Input, Output>(\n agent: Agent,\n options: AgentEvalTargetOptions<Input, Output>,\n): EvalTarget<Input, Output>;\nexport function agentEvalTarget<Input, Output>(\n agent: Agent,\n options: AgentEvalTargetOptions<Input, Output | PromptResponse> = {},\n): EvalTarget<Input, Output | PromptResponse> {\n return async (input, testCase) => {\n const prompt = options.prompt?.(input, testCase) ?? String(input);\n const response = await agent.prompt(prompt).send();\n return options.output === undefined ? response : options.output(response, testCase);\n };\n}\n\nasync function runEvalCase<Input, Output, Expected>(\n options: RunEvalSuiteOptions<Input, Output, Expected>,\n testCase: EvalCase<Input, Expected>,\n): Promise<EvalCaseResult<Input, Output, Expected>> {\n let output: Output | undefined;\n let targetError: unknown;\n try {\n output = await options.target(testCase.input, testCase);\n } catch (error) {\n targetError = error;\n }\n\n const metrics: EvalMetricResult[] = [];\n for (const metric of options.metrics) {\n const outcome =\n targetError === undefined\n ? await safeEvaluate(options.name, testCase, output as Output, metric)\n : EvalOutcome.invalid(`Target failed: ${errorMessage(targetError)}`);\n const reporterErrors = await reportOutcome({\n suiteName: options.name,\n testCase,\n output,\n targetError,\n metric,\n outcome,\n reporters: options.reporters ?? [],\n failOnReporterError: options.failOnReporterError === true,\n });\n metrics.push({ metricName: metric.name, outcome, reporterErrors });\n }\n\n return {\n case: testCase,\n ...(output === undefined ? {} : { output }),\n ...(targetError === undefined ? {} : { targetError }),\n metrics,\n };\n}\n\nasync function safeEvaluate<Input, Output, Expected>(\n suiteName: string,\n testCase: EvalCase<Input, Expected>,\n output: Output,\n metric: EvalMetric<Input, Output, unknown, Expected>,\n): Promise<EvalOutcome> {\n try {\n return await metric.evaluate({ suiteName, case: testCase, output });\n } catch (error) {\n return EvalOutcome.invalid(errorMessage(error));\n }\n}\n\nasync function reportOutcome<Input, Output, Expected>(args: {\n suiteName: string;\n testCase: EvalCase<Input, Expected>;\n output: Output | undefined;\n targetError: unknown;\n metric: EvalMetric<Input, Output, unknown, Expected>;\n outcome: EvalOutcome;\n reporters: Array<EvalReporter<Input, Output, Expected>>;\n failOnReporterError: boolean;\n}): Promise<unknown[]> {\n const errors: unknown[] = [];\n for (const reporter of args.reporters) {\n try {\n await reporter.report({\n suiteName: args.suiteName,\n case: args.testCase,\n output: args.output,\n targetError: args.targetError,\n metric: args.metric,\n outcome: args.outcome,\n });\n } catch (error) {\n if (args.failOnReporterError) {\n throw error;\n }\n errors.push(error);\n }\n }\n return errors;\n}\n\nfunction countOutcomes(results: Array<EvalCaseResult<unknown, unknown, unknown>>): {\n passed: number;\n failed: number;\n invalid: number;\n} {\n let passed = 0;\n let failed = 0;\n let invalid = 0;\n for (const result of results) {\n for (const metric of result.metrics) {\n if (metric.outcome.outcome === \"pass\") passed += 1;\n if (metric.outcome.outcome === \"fail\") failed += 1;\n if (metric.outcome.outcome === \"invalid\") invalid += 1;\n }\n }\n return { passed, failed, invalid };\n}\n\nasync function resolveActual<Input, Output, Expected>(\n selector: ValueSelector<Input, Output, Expected, unknown> | undefined,\n args: EvalMetricArgs<Input, Output, Expected>,\n): Promise<unknown> {\n return selector === undefined ? defaultOutputValue(args.output) : selector(args);\n}\n\nasync function resolveActualText<Input, Output, Expected>(\n selector: ValueSelector<Input, Output, Expected, string> | undefined,\n args: EvalMetricArgs<Input, Output, Expected>,\n): Promise<string> {\n const value = selector === undefined ? defaultOutputValue(args.output) : await selector(args);\n return typeof value === \"string\" ? value : JSON.stringify(value);\n}\n\nasync function resolveExpected<Input, Output, Expected, Value>(\n selectorOrValue: SelectorOrValue<Input, Output, Expected, Value> | undefined,\n args: EvalMetricArgs<Input, Output, Expected>,\n): Promise<Value | Expected | undefined> {\n if (selectorOrValue === undefined) {\n return args.case.expected;\n }\n return typeof selectorOrValue === \"function\"\n ? (selectorOrValue as ValueSelector<Input, Output, Expected, Value>)(args)\n : selectorOrValue;\n}\n\nasync function resolveJudgePrompt<Input, Output, Expected>(\n selector: ValueSelector<Input, Output, Expected, string> | undefined,\n args: EvalMetricArgs<Input, Output, Expected>,\n): Promise<string> {\n if (selector !== undefined) {\n return selector(args);\n }\n return [\n `Suite: ${args.suiteName}`,\n `Case: ${args.case.id}`,\n `Input: ${formatValue(args.case.input)}`,\n `Expected: ${formatValue(args.case.expected)}`,\n `Output: ${formatValue(defaultOutputValue(args.output))}`,\n ].join(\"\\n\\n\");\n}\n\nfunction defaultOutputValue(output: unknown): unknown {\n if (\n typeof output === \"object\" &&\n output !== null &&\n \"output\" in output &&\n typeof (output as { output?: unknown }).output === \"string\"\n ) {\n return (output as { output: string }).output;\n }\n return output;\n}\n\nfunction stableComparable(value: unknown): string {\n if (typeof value === \"string\") {\n return value;\n }\n return JSON.stringify(value);\n}\n\nfunction formatValue(value: unknown): string {\n if (typeof value === \"string\") {\n return value;\n }\n try {\n return JSON.stringify(value);\n } catch {\n return String(value);\n }\n}\n\nfunction errorMessage(error: unknown): string {\n return error instanceof Error ? error.message : String(error);\n}\n\nasync function mapWithConcurrency<Input, Output>(\n inputs: Input[],\n concurrency: number,\n mapper: (input: Input) => Promise<Output>,\n): Promise<Output[]> {\n const results = new Array<Output>(inputs.length);\n let next = 0;\n\n async function worker(): Promise<void> {\n while (next < inputs.length) {\n const index = next;\n next += 1;\n results[index] = await mapper(inputs[index] as Input);\n }\n }\n\n await Promise.all(Array.from({ length: Math.min(concurrency, inputs.length) }, () => worker()));\n return results;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAAA,SAAS,SAAS;AA6CX,IAAM,cAAc;AAAA,EACzB,KACE,OACA,UAAiF,CAAC,GAC9D;AACpB,WAAO;AAAA,MACL,SAAS;AAAA,MACT,GAAI,UAAU,SAAY,CAAC,IAAI,EAAE,MAAM;AAAA,MACvC,GAAI,QAAQ,YAAY,SAAY,CAAC,IAAI,EAAE,SAAS,QAAQ,QAAQ;AAAA,MACpE,GAAI,QAAQ,aAAa,SAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;AAAA,IACzE;AAAA,EACF;AAAA,EAEA,KACE,OACA,UAAiF,CAAC,GAC9D;AACpB,WAAO;AAAA,MACL,SAAS;AAAA,MACT,GAAI,UAAU,SAAY,CAAC,IAAI,EAAE,MAAM;AAAA,MACvC,GAAI,QAAQ,YAAY,SAAY,CAAC,IAAI,EAAE,SAAS,QAAQ,QAAQ;AAAA,MACpE,GAAI,QAAQ,aAAa,SAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;AAAA,IACzE;AAAA,EACF;AAAA,EAEA,QACE,QACA,UAII,CAAC,GACe;AACpB,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA,GAAI,QAAQ,UAAU,SAAY,CAAC,IAAI,EAAE,OAAO,QAAQ,MAAM;AAAA,MAC9D,GAAI,QAAQ,YAAY,SAAY,CAAC,IAAI,EAAE,SAAS,QAAQ,QAAQ;AAAA,MACpE,GAAI,QAAQ,aAAa,SAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;AAAA,IACzE;AAAA,EACF;AACF;AA4DA,eAAsB,aACpB,SACmD;AACnD,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,UAAU,MAAM;AAAA,IACpB,QAAQ;AAAA,IACR,KAAK,IAAI,GAAG,KAAK,MAAM,QAAQ,eAAe,CAAC,CAAC;AAAA,IAChD,CAAC,aAAa,YAAY,SAAS,QAAQ;AAAA,EAC7C;AACA,QAAM,SAAS,cAAc,OAAO;AACpC,SAAO;AAAA,IACL,MAAM,QAAQ;AAAA,IACd;AAAA,IACA,GAAG;AAAA,IACH,YAAY,KAAK,IAAI,IAAI;AAAA,EAC3B;AACF;AAgBO,SAAS,WACd,UAAsD,CAAC,GACT;AAC9C,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,YAAM,SAAS,MAAM,cAAc,QAAQ,QAAQ,IAAI;AACvD,YAAM,WAAW,MAAM,gBAAgB,QAAQ,UAAU,IAAI;AAC7D,UAAI,aAAa,QAAW;AAC1B,eAAO,YAAY,QAAQ,6CAA6C;AAAA,MAC1E;AACA,YAAM,SAAS,iBAAiB,MAAM,MAAM,iBAAiB,QAAQ;AACrE,aAAO,SACH,YAAY,KAAK,IAAI,IACrB,YAAY,KAAK,OAAO,EAAE,SAAS,YAAY,YAAY,QAAQ,CAAC,IAAI,CAAC;AAAA,IAC/E;AAAA,EACF;AACF;AAQO,SAAS,SACd,UAAoD,CAAC,GACP;AAC9C,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,YAAM,SAAS,MAAM,kBAAkB,QAAQ,QAAQ,IAAI;AAC3D,YAAM,WAAW,MAAM,gBAAgB,QAAQ,UAAU,IAAI;AAC7D,UAAI,aAAa,QAAW;AAC1B,eAAO,YAAY,QAAQ,0CAA0C;AAAA,MACvE;AACA,UAAI,OAAO,aAAa,YAAY,EAAE,oBAAoB,SAAS;AACjE,eAAO,YAAY,QAAQ,qDAAqD;AAAA,MAClF;AACA,YAAM,SAAS,oBAAoB,SAAS,SAAS,KAAK,MAAM,IAAI,OAAO,SAAS,QAAQ;AAC5F,aAAO,SACH,YAAY,KAAK,IAAI,IACrB,YAAY,KAAK,OAAO,EAAE,SAAS,0BAA0B,OAAO,QAAQ,CAAC,IAAI,CAAC;AAAA,IACxF;AAAA,EACF;AACF;AAUO,SAAS,mBACd,SAC6C;AAC7C,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,YAAM,SAAS,MAAM,kBAAkB,QAAQ,QAAQ,IAAI;AAC3D,YAAM,WAAW,MAAM,gBAAgB,QAAQ,UAAU,IAAI;AAC7D,UAAI,aAAa,QAAW;AAC1B,eAAO,YAAY,QAAQ,qDAAqD;AAAA,MAClF;AACA,UAAI,OAAO,aAAa,UAAU;AAChC,eAAO,YAAY,QAAQ,sDAAsD;AAAA,MACnF;AACA,YAAM,CAAC,iBAAiB,iBAAiB,IAAI,MAAM,QAAQ,IAAI;AAAA,QAC7D,UAAU,QAAQ,OAAO,MAAM;AAAA,QAC/B,UAAU,QAAQ,OAAO,QAAQ;AAAA,MACnC,CAAC;AACD,YAAM,QAAQ,iBAAiB,gBAAgB,QAAQ,kBAAkB,MAAM;AAC/E,aAAO,SAAS,QAAQ,YACpB,YAAY,KAAK,KAAK,IACtB,YAAY,KAAK,OAAO,EAAE,SAAS,8BAA8B,QAAQ,SAAS,IAAI,CAAC;AAAA,IAC7F;AAAA,EACF;AACF;AAYO,SAAS,SACd,SACmD;AACnD,QAAM,YAAY,IAAI,iBAAiB,QAAQ,OAAO,QAAQ,MAAM,EACjE;AAAA,IACC,QAAQ,gBACN;AAAA,EACJ,EACC,QAAQ,QAAQ,WAAW,CAAC,EAC5B,MAAM;AAET,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,UAAI;AACF,cAAM,WAAW,MAAM,UAAU,QAAQ,MAAM,mBAAmB,QAAQ,QAAQ,IAAI,CAAC;AACvF,eAAO,QAAQ,OAAO,QAAQ,IAAI,YAAY,KAAK,QAAQ,IAAI,YAAY,KAAK,QAAQ;AAAA,MAC1F,SAAS,OAAO;AACd,eAAO,YAAY,QAAQ,aAAa,KAAK,CAAC;AAAA,MAChD;AAAA,IACF;AAAA,EACF;AACF;AAiBO,SAAS,SACd,SAC0D;AAC1D,QAAM,WAAW,MAAM,QAAQ,QAAQ,QAAQ,IAAI,QAAQ,SAAS,KAAK,IAAI,IAAI,QAAQ;AACzF,QAAM,YAAY,IAAI;AAAA,IACpB,QAAQ;AAAA,IACR,EAAE,OAAO;AAAA,MACP,OAAO,EAAE,OAAO;AAAA,MAChB,UAAU,EAAE,OAAO;AAAA,IACrB,CAAC;AAAA,EACH,EACG;AAAA,IACC,QAAQ,gBACN;AAAA,EAAgD,QAAQ;AAAA;AAAA;AAAA,EAC5D,EACC,QAAQ,QAAQ,WAAW,CAAC,EAC5B,MAAM;AAET,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,UAAI;AACF,cAAM,QAAQ,MAAM,UAAU,QAAQ,MAAM,mBAAmB,QAAQ,QAAQ,IAAI,CAAC;AACpF,YAAI,MAAM,QAAQ,KAAK,MAAM,QAAQ,GAAG;AACtC,iBAAO,YAAY,QAAQ,SAAS,MAAM,KAAK,gCAAgC;AAAA,YAC7E;AAAA,UACF,CAAC;AAAA,QACH;AACA,eAAO,MAAM,SAAS,QAAQ,YAC1B,YAAY,KAAK,OAAO,EAAE,SAAS,MAAM,SAAS,CAAC,IACnD,YAAY,KAAK,OAAO,EAAE,SAAS,MAAM,SAAS,CAAC;AAAA,MACzD,SAAS,OAAO;AACd,eAAO,YAAY,QAAQ,aAAa,KAAK,CAAC;AAAA,MAChD;AAAA,IACF;AAAA,EACF;AACF;AAeO,SAAS,gBACd,OACA,UAAkE,CAAC,GACvB;AAC5C,SAAO,OAAO,OAAO,aAAa;AAChC,UAAM,SAAS,QAAQ,SAAS,OAAO,QAAQ,KAAK,OAAO,KAAK;AAChE,UAAM,WAAW,MAAM,MAAM,OAAO,MAAM,EAAE,KAAK;AACjD,WAAO,QAAQ,WAAW,SAAY,WAAW,QAAQ,OAAO,UAAU,QAAQ;AAAA,EACpF;AACF;AAEA,eAAe,YACb,SACA,UACkD;AAClD,MAAI;AACJ,MAAI;AACJ,MAAI;AACF,aAAS,MAAM,QAAQ,OAAO,SAAS,OAAO,QAAQ;AAAA,EACxD,SAAS,OAAO;AACd,kBAAc;AAAA,EAChB;AAEA,QAAM,UAA8B,CAAC;AACrC,aAAW,UAAU,QAAQ,SAAS;AACpC,UAAM,UACJ,gBAAgB,SACZ,MAAM,aAAa,QAAQ,MAAM,UAAU,QAAkB,MAAM,IACnE,YAAY,QAAQ,kBAAkB,aAAa,WAAW,CAAC,EAAE;AACvE,UAAM,iBAAiB,MAAM,cAAc;AAAA,MACzC,WAAW,QAAQ;AAAA,MACnB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,WAAW,QAAQ,aAAa,CAAC;AAAA,MACjC,qBAAqB,QAAQ,wBAAwB;AAAA,IACvD,CAAC;AACD,YAAQ,KAAK,EAAE,YAAY,OAAO,MAAM,SAAS,eAAe,CAAC;AAAA,EACnE;AAEA,SAAO;AAAA,IACL,MAAM;AAAA,IACN,GAAI,WAAW,SAAY,CAAC,IAAI,EAAE,OAAO;AAAA,IACzC,GAAI,gBAAgB,SAAY,CAAC,IAAI,EAAE,YAAY;AAAA,IACnD;AAAA,EACF;AACF;AAEA,eAAe,aACb,WACA,UACA,QACA,QACsB;AACtB,MAAI;AACF,WAAO,MAAM,OAAO,SAAS,EAAE,WAAW,MAAM,UAAU,OAAO,CAAC;AAAA,EACpE,SAAS,OAAO;AACd,WAAO,YAAY,QAAQ,aAAa,KAAK,CAAC;AAAA,EAChD;AACF;AAEA,eAAe,cAAuC,MAS/B;AACrB,QAAM,SAAoB,CAAC;AAC3B,aAAW,YAAY,KAAK,WAAW;AACrC,QAAI;AACF,YAAM,SAAS,OAAO;AAAA,QACpB,WAAW,KAAK;AAAA,QAChB,MAAM,KAAK;AAAA,QACX,QAAQ,KAAK;AAAA,QACb,aAAa,KAAK;AAAA,QAClB,QAAQ,KAAK;AAAA,QACb,SAAS,KAAK;AAAA,MAChB,CAAC;AAAA,IACH,SAAS,OAAO;AACd,UAAI,KAAK,qBAAqB;AAC5B,cAAM;AAAA,MACR;AACA,aAAO,KAAK,KAAK;AAAA,IACnB;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,cAAc,SAIrB;AACA,MAAI,SAAS;AACb,MAAI,SAAS;AACb,MAAI,UAAU;AACd,aAAW,UAAU,SAAS;AAC5B,eAAW,UAAU,OAAO,SAAS;AACnC,UAAI,OAAO,QAAQ,YAAY,OAAQ,WAAU;AACjD,UAAI,OAAO,QAAQ,YAAY,OAAQ,WAAU;AACjD,UAAI,OAAO,QAAQ,YAAY,UAAW,YAAW;AAAA,IACvD;AAAA,EACF;AACA,SAAO,EAAE,QAAQ,QAAQ,QAAQ;AACnC;AAEA,eAAe,cACb,UACA,MACkB;AAClB,SAAO,aAAa,SAAY,mBAAmB,KAAK,MAAM,IAAI,SAAS,IAAI;AACjF;AAEA,eAAe,kBACb,UACA,MACiB;AACjB,QAAM,QAAQ,aAAa,SAAY,mBAAmB,KAAK,MAAM,IAAI,MAAM,SAAS,IAAI;AAC5F,SAAO,OAAO,UAAU,WAAW,QAAQ,KAAK,UAAU,KAAK;AACjE;AAEA,eAAe,gBACb,iBACA,MACuC;AACvC,MAAI,oBAAoB,QAAW;AACjC,WAAO,KAAK,KAAK;AAAA,EACnB;AACA,SAAO,OAAO,oBAAoB,aAC7B,gBAAkE,IAAI,IACvE;AACN;AAEA,eAAe,mBACb,UACA,MACiB;AACjB,MAAI,aAAa,QAAW;AAC1B,WAAO,SAAS,IAAI;AAAA,EACtB;AACA,SAAO;AAAA,IACL,UAAU,KAAK,SAAS;AAAA,IACxB,SAAS,KAAK,KAAK,EAAE;AAAA,IACrB,UAAU,YAAY,KAAK,KAAK,KAAK,CAAC;AAAA,IACtC,aAAa,YAAY,KAAK,KAAK,QAAQ,CAAC;AAAA,IAC5C,WAAW,YAAY,mBAAmB,KAAK,MAAM,CAAC,CAAC;AAAA,EACzD,EAAE,KAAK,MAAM;AACf;AAEA,SAAS,mBAAmB,QAA0B;AACpD,MACE,OAAO,WAAW,YAClB,WAAW,QACX,YAAY,UACZ,OAAQ,OAAgC,WAAW,UACnD;AACA,WAAQ,OAA8B;AAAA,EACxC;AACA,SAAO;AACT;AAEA,SAAS,iBAAiB,OAAwB;AAChD,MAAI,OAAO,UAAU,UAAU;AAC7B,WAAO;AAAA,EACT;AACA,SAAO,KAAK,UAAU,KAAK;AAC7B;AAEA,SAAS,YAAY,OAAwB;AAC3C,MAAI,OAAO,UAAU,UAAU;AAC7B,WAAO;AAAA,EACT;AACA,MAAI;AACF,WAAO,KAAK,UAAU,KAAK;AAAA,EAC7B,QAAQ;AACN,WAAO,OAAO,KAAK;AAAA,EACrB;AACF;AAEA,SAAS,aAAa,OAAwB;AAC5C,SAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC9D;AAEA,eAAe,mBACb,QACA,aACA,QACmB;AACnB,QAAM,UAAU,IAAI,MAAc,OAAO,MAAM;AAC/C,MAAI,OAAO;AAEX,iBAAe,SAAwB;AACrC,WAAO,OAAO,OAAO,QAAQ;AAC3B,YAAM,QAAQ;AACd,cAAQ;AACR,cAAQ,KAAK,IAAI,MAAM,OAAO,OAAO,KAAK,CAAU;AAAA,IACtD;AAAA,EACF;AAEA,QAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,aAAa,OAAO,MAAM,EAAE,GAAG,MAAM,OAAO,CAAC,CAAC;AAC9F,SAAO;AACT;","names":[]}
1
+ {"version":3,"sources":["../../src/evals/agent-target.ts","../../src/evals/metrics.ts","../../src/evals/format.ts","../../src/evals/outcome.ts","../../src/evals/selectors.ts","../../src/evals/runner.ts"],"sourcesContent":["import type { Agent } from \"../agent/agent\";\nimport type { PromptResponse } from \"../agent/request\";\nimport type { Message } from \"../completion\";\nimport type { EvalCase, EvalTarget } from \"./types\";\n\nexport type AgentEvalTargetOptions<Input, Output = PromptResponse> = {\n prompt?: ((input: Input, testCase: EvalCase<Input>) => string | Message) | undefined;\n output?: ((response: PromptResponse, testCase: EvalCase<Input>) => Output) | undefined;\n};\n\nexport function agentEvalTarget<Input>(\n agent: Agent,\n options?: AgentEvalTargetOptions<Input, PromptResponse>,\n): EvalTarget<Input, PromptResponse>;\nexport function agentEvalTarget<Input, Output>(\n agent: Agent,\n options: AgentEvalTargetOptions<Input, Output>,\n): EvalTarget<Input, Output>;\nexport function agentEvalTarget<Input, Output>(\n agent: Agent,\n options: AgentEvalTargetOptions<Input, Output | PromptResponse> = {},\n): EvalTarget<Input, Output | PromptResponse> {\n return async (input, testCase) => {\n const prompt = options.prompt?.(input, testCase) ?? String(input);\n const response = await agent.prompt(prompt).send();\n return options.output === undefined ? response : options.output(response, testCase);\n };\n}\n","import { z } from \"zod\";\nimport type { CompletionModel } from \"../completion\";\nimport { cosineSimilarity, type EmbeddingModel, embedText } from \"../embeddings\";\nimport { ExtractorBuilder } from \"../extractor\";\nimport type { ZodSchema } from \"../schema\";\nimport { errorMessage, formatValue, stableComparable } from \"./format\";\nimport { EvalOutcome } from \"./outcome\";\nimport { resolveActual, resolveActualText, resolveExpected, resolveJudgePrompt } from \"./selectors\";\nimport type { EvalMetric, SelectorOrValue, ValueSelector } from \"./types\";\n\nexport type ExactMatchOptions<Input, Output, Expected = unknown> = {\n name?: string | undefined;\n actual?: ValueSelector<Input, Output, Expected, unknown> | undefined;\n expected?: SelectorOrValue<Input, Output, Expected, unknown> | undefined;\n};\n\nexport function exactMatch<Input, Output, Expected = unknown>(\n options: ExactMatchOptions<Input, Output, Expected> = {},\n): EvalMetric<Input, Output, boolean, Expected> {\n return {\n name: options.name ?? \"exact_match\",\n async evaluate(args) {\n const actual = await resolveActual(options.actual, args);\n const expected = await resolveExpected(options.expected, args);\n if (expected === undefined) {\n return EvalOutcome.invalid(\"No expected value provided for exact match.\");\n }\n const passed = stableComparable(actual) === stableComparable(expected);\n return passed\n ? EvalOutcome.pass(true)\n : EvalOutcome.fail(false, { comment: `Expected ${formatValue(expected)}.` });\n },\n };\n}\n\nexport type ContainsOptions<Input, Output, Expected = unknown> = {\n name?: string | undefined;\n actual?: ValueSelector<Input, Output, Expected, string> | undefined;\n expected?: SelectorOrValue<Input, Output, Expected, string | RegExp> | undefined;\n};\n\nexport function contains<Input, Output, Expected = unknown>(\n options: ContainsOptions<Input, Output, Expected> = {},\n): EvalMetric<Input, Output, boolean, Expected> {\n return {\n name: options.name ?? \"contains\",\n async evaluate(args) {\n const actual = await resolveActualText(options.actual, args);\n const expected = await resolveExpected(options.expected, args);\n if (expected === undefined) {\n return EvalOutcome.invalid(\"No expected value provided for contains.\");\n }\n if (typeof expected !== \"string\" && !(expected instanceof RegExp)) {\n return EvalOutcome.invalid(\"Contains expected value must be a string or RegExp.\");\n }\n const passed = expected instanceof RegExp ? expected.test(actual) : actual.includes(expected);\n return passed\n ? EvalOutcome.pass(true)\n : EvalOutcome.fail(false, { comment: `Output did not contain ${String(expected)}.` });\n },\n };\n}\n\nexport type SemanticSimilarityOptions<Input, Output, Expected = unknown> = {\n name?: string | undefined;\n model: EmbeddingModel;\n threshold: number;\n actual?: ValueSelector<Input, Output, Expected, string> | undefined;\n expected?: SelectorOrValue<Input, Output, Expected, string> | undefined;\n};\n\nexport function semanticSimilarity<Input, Output, Expected = unknown>(\n options: SemanticSimilarityOptions<Input, Output, Expected>,\n): EvalMetric<Input, Output, number, Expected> {\n return {\n name: options.name ?? \"semantic_similarity\",\n async evaluate(args) {\n const actual = await resolveActualText(options.actual, args);\n const expected = await resolveExpected(options.expected, args);\n if (expected === undefined) {\n return EvalOutcome.invalid(\"No expected value provided for semantic similarity.\");\n }\n if (typeof expected !== \"string\") {\n return EvalOutcome.invalid(\"Semantic similarity expected value must be a string.\");\n }\n const [actualEmbedding, expectedEmbedding] = await Promise.all([\n embedText(options.model, actual),\n embedText(options.model, expected),\n ]);\n const score = cosineSimilarity(actualEmbedding.vector, expectedEmbedding.vector);\n return score >= options.threshold\n ? EvalOutcome.pass(score)\n : EvalOutcome.fail(score, { comment: `Similarity below threshold ${options.threshold}.` });\n },\n };\n}\n\nexport type LlmJudgeOptions<Input, Output, SchemaOutput, Expected = unknown> = {\n name?: string | undefined;\n model: CompletionModel;\n schema: ZodSchema<SchemaOutput>;\n passes(value: SchemaOutput): boolean;\n instructions?: string | undefined;\n retries?: number | undefined;\n prompt?: ValueSelector<Input, Output, Expected, string> | undefined;\n};\n\nexport function llmJudge<Input, Output, SchemaOutput, Expected = unknown>(\n options: LlmJudgeOptions<Input, Output, SchemaOutput, Expected>,\n): EvalMetric<Input, Output, SchemaOutput, Expected> {\n const extractor = new ExtractorBuilder(options.model, options.schema)\n .instructions(\n options.instructions ??\n \"Judge the eval case by the requested schema. Submit the judgment using the schema.\",\n )\n .retries(options.retries ?? 0)\n .build();\n\n return {\n name: options.name ?? \"llm_judge\",\n async evaluate(args) {\n try {\n const judgment = await extractor.extract(await resolveJudgePrompt(options.prompt, args));\n return options.passes(judgment) ? EvalOutcome.pass(judgment) : EvalOutcome.fail(judgment);\n } catch (error) {\n return EvalOutcome.invalid(errorMessage(error));\n }\n },\n };\n}\n\nexport type LlmScoreMetricScore = {\n score: number;\n feedback: string;\n};\n\nexport type LlmScoreOptions<Input, Output, Expected = unknown> = {\n name?: string | undefined;\n model: CompletionModel;\n threshold: number;\n criteria: string | string[];\n instructions?: string | undefined;\n retries?: number | undefined;\n prompt?: ValueSelector<Input, Output, Expected, string> | undefined;\n};\n\nexport function llmScore<Input, Output, Expected = unknown>(\n options: LlmScoreOptions<Input, Output, Expected>,\n): EvalMetric<Input, Output, LlmScoreMetricScore, Expected> {\n const criteria = Array.isArray(options.criteria) ? options.criteria.join(\"\\n\") : options.criteria;\n const extractor = new ExtractorBuilder(\n options.model,\n z.object({\n score: z.number(),\n feedback: z.string(),\n }),\n )\n .instructions(\n options.instructions ??\n `Score the eval case against these criteria:\\n${criteria}\\n\\nReturn a score between 0 and 1 and brief feedback.`,\n )\n .retries(options.retries ?? 0)\n .build();\n\n return {\n name: options.name ?? \"llm_score\",\n async evaluate(args) {\n try {\n const score = await extractor.extract(await resolveJudgePrompt(options.prompt, args));\n if (score.score < 0 || score.score > 1) {\n return EvalOutcome.invalid(`Score ${score.score} outside valid range [0, 1].`, {\n score,\n });\n }\n return score.score >= options.threshold\n ? EvalOutcome.pass(score, { comment: score.feedback })\n : EvalOutcome.fail(score, { comment: score.feedback });\n } catch (error) {\n return EvalOutcome.invalid(errorMessage(error));\n }\n },\n };\n}\n","export function defaultOutputValue(output: unknown): unknown {\n if (\n typeof output === \"object\" &&\n output !== null &&\n \"output\" in output &&\n typeof (output as { output?: unknown }).output === \"string\"\n ) {\n return (output as { output: string }).output;\n }\n return output;\n}\n\nexport function stableComparable(value: unknown): string {\n if (typeof value === \"string\") {\n return value;\n }\n return JSON.stringify(value);\n}\n\nexport function formatValue(value: unknown): string {\n if (typeof value === \"string\") {\n return value;\n }\n try {\n return JSON.stringify(value);\n } catch {\n return String(value);\n }\n}\n\nexport function errorMessage(error: unknown): string {\n return error instanceof Error ? error.message : String(error);\n}\n","import type { EvalMetadata } from \"./types\";\n\nexport type EvalOutcome<Score = unknown> =\n | {\n outcome: \"pass\";\n score?: Score | undefined;\n comment?: string | undefined;\n metadata?: EvalMetadata | undefined;\n }\n | {\n outcome: \"fail\";\n score?: Score | undefined;\n comment?: string | undefined;\n metadata?: EvalMetadata | undefined;\n }\n | {\n outcome: \"invalid\";\n reason: string;\n score?: Score | undefined;\n comment?: string | undefined;\n metadata?: EvalMetadata | undefined;\n };\n\nexport const EvalOutcome = {\n pass<Score>(\n score?: Score,\n options: { comment?: string | undefined; metadata?: EvalMetadata | undefined } = {},\n ): EvalOutcome<Score> {\n return {\n outcome: \"pass\",\n ...(score === undefined ? {} : { score }),\n ...(options.comment === undefined ? {} : { comment: options.comment }),\n ...(options.metadata === undefined ? {} : { metadata: options.metadata }),\n };\n },\n\n fail<Score>(\n score?: Score,\n options: { comment?: string | undefined; metadata?: EvalMetadata | undefined } = {},\n ): EvalOutcome<Score> {\n return {\n outcome: \"fail\",\n ...(score === undefined ? {} : { score }),\n ...(options.comment === undefined ? {} : { comment: options.comment }),\n ...(options.metadata === undefined ? {} : { metadata: options.metadata }),\n };\n },\n\n invalid<Score = never>(\n reason: string,\n options: {\n score?: Score | undefined;\n comment?: string | undefined;\n metadata?: EvalMetadata | undefined;\n } = {},\n ): EvalOutcome<Score> {\n return {\n outcome: \"invalid\",\n reason,\n ...(options.score === undefined ? {} : { score: options.score }),\n ...(options.comment === undefined ? {} : { comment: options.comment }),\n ...(options.metadata === undefined ? {} : { metadata: options.metadata }),\n };\n },\n};\n","import { defaultOutputValue, formatValue } from \"./format\";\nimport type { EvalMetricArgs, SelectorOrValue, ValueSelector } from \"./types\";\n\nexport async function resolveActual<Input, Output, Expected>(\n selector: ValueSelector<Input, Output, Expected, unknown> | undefined,\n args: EvalMetricArgs<Input, Output, Expected>,\n): Promise<unknown> {\n return selector === undefined ? defaultOutputValue(args.output) : selector(args);\n}\n\nexport async function resolveActualText<Input, Output, Expected>(\n selector: ValueSelector<Input, Output, Expected, string> | undefined,\n args: EvalMetricArgs<Input, Output, Expected>,\n): Promise<string> {\n const value = selector === undefined ? defaultOutputValue(args.output) : await selector(args);\n return typeof value === \"string\" ? value : JSON.stringify(value);\n}\n\nexport async function resolveExpected<Input, Output, Expected, Value>(\n selectorOrValue: SelectorOrValue<Input, Output, Expected, Value> | undefined,\n args: EvalMetricArgs<Input, Output, Expected>,\n): Promise<Value | Expected | undefined> {\n if (selectorOrValue === undefined) {\n return args.case.expected;\n }\n return typeof selectorOrValue === \"function\"\n ? (selectorOrValue as ValueSelector<Input, Output, Expected, Value>)(args)\n : selectorOrValue;\n}\n\nexport async function resolveJudgePrompt<Input, Output, Expected>(\n selector: ValueSelector<Input, Output, Expected, string> | undefined,\n args: EvalMetricArgs<Input, Output, Expected>,\n): Promise<string> {\n if (selector !== undefined) {\n return selector(args);\n }\n return [\n `Suite: ${args.suiteName}`,\n `Case: ${args.case.id}`,\n `Input: ${formatValue(args.case.input)}`,\n `Expected: ${formatValue(args.case.expected)}`,\n `Output: ${formatValue(defaultOutputValue(args.output))}`,\n ].join(\"\\n\\n\");\n}\n","import { mapWithConcurrency } from \"../internal/concurrency\";\nimport { errorMessage } from \"./format\";\nimport { EvalOutcome, type EvalOutcome as EvalOutcomeType } from \"./outcome\";\nimport type {\n EvalCase,\n EvalCaseResult,\n EvalMetric,\n EvalMetricResult,\n EvalReporter,\n EvalSuiteResult,\n RunEvalSuiteOptions,\n} from \"./types\";\n\nexport async function runEvalSuite<Input, Output, Expected = unknown>(\n options: RunEvalSuiteOptions<Input, Output, Expected>,\n): Promise<EvalSuiteResult<Input, Output, Expected>> {\n const startedAt = Date.now();\n const results = await mapWithConcurrency(\n options.cases,\n Math.max(1, Math.trunc(options.concurrency ?? 1)),\n (testCase) => runEvalCase(options, testCase),\n );\n const counts = countOutcomes(results);\n return {\n name: options.name,\n results,\n ...counts,\n durationMs: Date.now() - startedAt,\n };\n}\n\nasync function runEvalCase<Input, Output, Expected>(\n options: RunEvalSuiteOptions<Input, Output, Expected>,\n testCase: EvalCase<Input, Expected>,\n): Promise<EvalCaseResult<Input, Output, Expected>> {\n let output: Output | undefined;\n let targetError: unknown;\n try {\n output = await options.target(testCase.input, testCase);\n } catch (error) {\n targetError = error;\n }\n\n const metrics: EvalMetricResult[] = [];\n for (const metric of options.metrics) {\n const outcome =\n targetError === undefined\n ? await safeEvaluate(options.name, testCase, output as Output, metric)\n : EvalOutcome.invalid(`Target failed: ${errorMessage(targetError)}`);\n const reporterErrors = await reportOutcome({\n suiteName: options.name,\n testCase,\n output,\n targetError,\n metric,\n outcome,\n reporters: options.reporters ?? [],\n failOnReporterError: options.failOnReporterError === true,\n });\n metrics.push({ metricName: metric.name, outcome, reporterErrors });\n }\n\n return {\n case: testCase,\n ...(output === undefined ? {} : { output }),\n ...(targetError === undefined ? {} : { targetError }),\n metrics,\n };\n}\n\nasync function safeEvaluate<Input, Output, Expected>(\n suiteName: string,\n testCase: EvalCase<Input, Expected>,\n output: Output,\n metric: EvalMetric<Input, Output, unknown, Expected>,\n): Promise<EvalOutcomeType> {\n try {\n return await metric.evaluate({ suiteName, case: testCase, output });\n } catch (error) {\n return EvalOutcome.invalid(errorMessage(error));\n }\n}\n\nasync function reportOutcome<Input, Output, Expected>(args: {\n suiteName: string;\n testCase: EvalCase<Input, Expected>;\n output: Output | undefined;\n targetError: unknown;\n metric: EvalMetric<Input, Output, unknown, Expected>;\n outcome: EvalOutcomeType;\n reporters: Array<EvalReporter<Input, Output, Expected>>;\n failOnReporterError: boolean;\n}): Promise<unknown[]> {\n const errors: unknown[] = [];\n for (const reporter of args.reporters) {\n try {\n await reporter.report({\n suiteName: args.suiteName,\n case: args.testCase,\n output: args.output,\n targetError: args.targetError,\n metric: args.metric,\n outcome: args.outcome,\n });\n } catch (error) {\n if (args.failOnReporterError) {\n throw error;\n }\n errors.push(error);\n }\n }\n return errors;\n}\n\nfunction countOutcomes(results: Array<EvalCaseResult<unknown, unknown, unknown>>): {\n passed: number;\n failed: number;\n invalid: number;\n} {\n let passed = 0;\n let failed = 0;\n let invalid = 0;\n for (const result of results) {\n for (const metric of result.metrics) {\n if (metric.outcome.outcome === \"pass\") passed += 1;\n if (metric.outcome.outcome === \"fail\") failed += 1;\n if (metric.outcome.outcome === \"invalid\") invalid += 1;\n }\n }\n return { passed, failed, invalid };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAkBO,SAAS,gBACd,OACA,UAAkE,CAAC,GACvB;AAC5C,SAAO,OAAO,OAAO,aAAa;AAChC,UAAM,SAAS,QAAQ,SAAS,OAAO,QAAQ,KAAK,OAAO,KAAK;AAChE,UAAM,WAAW,MAAM,MAAM,OAAO,MAAM,EAAE,KAAK;AACjD,WAAO,QAAQ,WAAW,SAAY,WAAW,QAAQ,OAAO,UAAU,QAAQ;AAAA,EACpF;AACF;;;AC3BA,SAAS,SAAS;;;ACAX,SAAS,mBAAmB,QAA0B;AAC3D,MACE,OAAO,WAAW,YAClB,WAAW,QACX,YAAY,UACZ,OAAQ,OAAgC,WAAW,UACnD;AACA,WAAQ,OAA8B;AAAA,EACxC;AACA,SAAO;AACT;AAEO,SAAS,iBAAiB,OAAwB;AACvD,MAAI,OAAO,UAAU,UAAU;AAC7B,WAAO;AAAA,EACT;AACA,SAAO,KAAK,UAAU,KAAK;AAC7B;AAEO,SAAS,YAAY,OAAwB;AAClD,MAAI,OAAO,UAAU,UAAU;AAC7B,WAAO;AAAA,EACT;AACA,MAAI;AACF,WAAO,KAAK,UAAU,KAAK;AAAA,EAC7B,QAAQ;AACN,WAAO,OAAO,KAAK;AAAA,EACrB;AACF;AAEO,SAAS,aAAa,OAAwB;AACnD,SAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC9D;;;ACTO,IAAM,cAAc;AAAA,EACzB,KACE,OACA,UAAiF,CAAC,GAC9D;AACpB,WAAO;AAAA,MACL,SAAS;AAAA,MACT,GAAI,UAAU,SAAY,CAAC,IAAI,EAAE,MAAM;AAAA,MACvC,GAAI,QAAQ,YAAY,SAAY,CAAC,IAAI,EAAE,SAAS,QAAQ,QAAQ;AAAA,MACpE,GAAI,QAAQ,aAAa,SAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;AAAA,IACzE;AAAA,EACF;AAAA,EAEA,KACE,OACA,UAAiF,CAAC,GAC9D;AACpB,WAAO;AAAA,MACL,SAAS;AAAA,MACT,GAAI,UAAU,SAAY,CAAC,IAAI,EAAE,MAAM;AAAA,MACvC,GAAI,QAAQ,YAAY,SAAY,CAAC,IAAI,EAAE,SAAS,QAAQ,QAAQ;AAAA,MACpE,GAAI,QAAQ,aAAa,SAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;AAAA,IACzE;AAAA,EACF;AAAA,EAEA,QACE,QACA,UAII,CAAC,GACe;AACpB,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA,GAAI,QAAQ,UAAU,SAAY,CAAC,IAAI,EAAE,OAAO,QAAQ,MAAM;AAAA,MAC9D,GAAI,QAAQ,YAAY,SAAY,CAAC,IAAI,EAAE,SAAS,QAAQ,QAAQ;AAAA,MACpE,GAAI,QAAQ,aAAa,SAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;AAAA,IACzE;AAAA,EACF;AACF;;;AC7DA,eAAsB,cACpB,UACA,MACkB;AAClB,SAAO,aAAa,SAAY,mBAAmB,KAAK,MAAM,IAAI,SAAS,IAAI;AACjF;AAEA,eAAsB,kBACpB,UACA,MACiB;AACjB,QAAM,QAAQ,aAAa,SAAY,mBAAmB,KAAK,MAAM,IAAI,MAAM,SAAS,IAAI;AAC5F,SAAO,OAAO,UAAU,WAAW,QAAQ,KAAK,UAAU,KAAK;AACjE;AAEA,eAAsB,gBACpB,iBACA,MACuC;AACvC,MAAI,oBAAoB,QAAW;AACjC,WAAO,KAAK,KAAK;AAAA,EACnB;AACA,SAAO,OAAO,oBAAoB,aAC7B,gBAAkE,IAAI,IACvE;AACN;AAEA,eAAsB,mBACpB,UACA,MACiB;AACjB,MAAI,aAAa,QAAW;AAC1B,WAAO,SAAS,IAAI;AAAA,EACtB;AACA,SAAO;AAAA,IACL,UAAU,KAAK,SAAS;AAAA,IACxB,SAAS,KAAK,KAAK,EAAE;AAAA,IACrB,UAAU,YAAY,KAAK,KAAK,KAAK,CAAC;AAAA,IACtC,aAAa,YAAY,KAAK,KAAK,QAAQ,CAAC;AAAA,IAC5C,WAAW,YAAY,mBAAmB,KAAK,MAAM,CAAC,CAAC;AAAA,EACzD,EAAE,KAAK,MAAM;AACf;;;AH5BO,SAAS,WACd,UAAsD,CAAC,GACT;AAC9C,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,YAAM,SAAS,MAAM,cAAc,QAAQ,QAAQ,IAAI;AACvD,YAAM,WAAW,MAAM,gBAAgB,QAAQ,UAAU,IAAI;AAC7D,UAAI,aAAa,QAAW;AAC1B,eAAO,YAAY,QAAQ,6CAA6C;AAAA,MAC1E;AACA,YAAM,SAAS,iBAAiB,MAAM,MAAM,iBAAiB,QAAQ;AACrE,aAAO,SACH,YAAY,KAAK,IAAI,IACrB,YAAY,KAAK,OAAO,EAAE,SAAS,YAAY,YAAY,QAAQ,CAAC,IAAI,CAAC;AAAA,IAC/E;AAAA,EACF;AACF;AAQO,SAAS,SACd,UAAoD,CAAC,GACP;AAC9C,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,YAAM,SAAS,MAAM,kBAAkB,QAAQ,QAAQ,IAAI;AAC3D,YAAM,WAAW,MAAM,gBAAgB,QAAQ,UAAU,IAAI;AAC7D,UAAI,aAAa,QAAW;AAC1B,eAAO,YAAY,QAAQ,0CAA0C;AAAA,MACvE;AACA,UAAI,OAAO,aAAa,YAAY,EAAE,oBAAoB,SAAS;AACjE,eAAO,YAAY,QAAQ,qDAAqD;AAAA,MAClF;AACA,YAAM,SAAS,oBAAoB,SAAS,SAAS,KAAK,MAAM,IAAI,OAAO,SAAS,QAAQ;AAC5F,aAAO,SACH,YAAY,KAAK,IAAI,IACrB,YAAY,KAAK,OAAO,EAAE,SAAS,0BAA0B,OAAO,QAAQ,CAAC,IAAI,CAAC;AAAA,IACxF;AAAA,EACF;AACF;AAUO,SAAS,mBACd,SAC6C;AAC7C,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,YAAM,SAAS,MAAM,kBAAkB,QAAQ,QAAQ,IAAI;AAC3D,YAAM,WAAW,MAAM,gBAAgB,QAAQ,UAAU,IAAI;AAC7D,UAAI,aAAa,QAAW;AAC1B,eAAO,YAAY,QAAQ,qDAAqD;AAAA,MAClF;AACA,UAAI,OAAO,aAAa,UAAU;AAChC,eAAO,YAAY,QAAQ,sDAAsD;AAAA,MACnF;AACA,YAAM,CAAC,iBAAiB,iBAAiB,IAAI,MAAM,QAAQ,IAAI;AAAA,QAC7D,UAAU,QAAQ,OAAO,MAAM;AAAA,QAC/B,UAAU,QAAQ,OAAO,QAAQ;AAAA,MACnC,CAAC;AACD,YAAM,QAAQ,iBAAiB,gBAAgB,QAAQ,kBAAkB,MAAM;AAC/E,aAAO,SAAS,QAAQ,YACpB,YAAY,KAAK,KAAK,IACtB,YAAY,KAAK,OAAO,EAAE,SAAS,8BAA8B,QAAQ,SAAS,IAAI,CAAC;AAAA,IAC7F;AAAA,EACF;AACF;AAYO,SAAS,SACd,SACmD;AACnD,QAAM,YAAY,IAAI,iBAAiB,QAAQ,OAAO,QAAQ,MAAM,EACjE;AAAA,IACC,QAAQ,gBACN;AAAA,EACJ,EACC,QAAQ,QAAQ,WAAW,CAAC,EAC5B,MAAM;AAET,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,UAAI;AACF,cAAM,WAAW,MAAM,UAAU,QAAQ,MAAM,mBAAmB,QAAQ,QAAQ,IAAI,CAAC;AACvF,eAAO,QAAQ,OAAO,QAAQ,IAAI,YAAY,KAAK,QAAQ,IAAI,YAAY,KAAK,QAAQ;AAAA,MAC1F,SAAS,OAAO;AACd,eAAO,YAAY,QAAQ,aAAa,KAAK,CAAC;AAAA,MAChD;AAAA,IACF;AAAA,EACF;AACF;AAiBO,SAAS,SACd,SAC0D;AAC1D,QAAM,WAAW,MAAM,QAAQ,QAAQ,QAAQ,IAAI,QAAQ,SAAS,KAAK,IAAI,IAAI,QAAQ;AACzF,QAAM,YAAY,IAAI;AAAA,IACpB,QAAQ;AAAA,IACR,EAAE,OAAO;AAAA,MACP,OAAO,EAAE,OAAO;AAAA,MAChB,UAAU,EAAE,OAAO;AAAA,IACrB,CAAC;AAAA,EACH,EACG;AAAA,IACC,QAAQ,gBACN;AAAA,EAAgD,QAAQ;AAAA;AAAA;AAAA,EAC5D,EACC,QAAQ,QAAQ,WAAW,CAAC,EAC5B,MAAM;AAET,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,SAAS,MAAM;AACnB,UAAI;AACF,cAAM,QAAQ,MAAM,UAAU,QAAQ,MAAM,mBAAmB,QAAQ,QAAQ,IAAI,CAAC;AACpF,YAAI,MAAM,QAAQ,KAAK,MAAM,QAAQ,GAAG;AACtC,iBAAO,YAAY,QAAQ,SAAS,MAAM,KAAK,gCAAgC;AAAA,YAC7E;AAAA,UACF,CAAC;AAAA,QACH;AACA,eAAO,MAAM,SAAS,QAAQ,YAC1B,YAAY,KAAK,OAAO,EAAE,SAAS,MAAM,SAAS,CAAC,IACnD,YAAY,KAAK,OAAO,EAAE,SAAS,MAAM,SAAS,CAAC;AAAA,MACzD,SAAS,OAAO;AACd,eAAO,YAAY,QAAQ,aAAa,KAAK,CAAC;AAAA,MAChD;AAAA,IACF;AAAA,EACF;AACF;;;AIzKA,eAAsB,aACpB,SACmD;AACnD,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,UAAU,MAAM;AAAA,IACpB,QAAQ;AAAA,IACR,KAAK,IAAI,GAAG,KAAK,MAAM,QAAQ,eAAe,CAAC,CAAC;AAAA,IAChD,CAAC,aAAa,YAAY,SAAS,QAAQ;AAAA,EAC7C;AACA,QAAM,SAAS,cAAc,OAAO;AACpC,SAAO;AAAA,IACL,MAAM,QAAQ;AAAA,IACd;AAAA,IACA,GAAG;AAAA,IACH,YAAY,KAAK,IAAI,IAAI;AAAA,EAC3B;AACF;AAEA,eAAe,YACb,SACA,UACkD;AAClD,MAAI;AACJ,MAAI;AACJ,MAAI;AACF,aAAS,MAAM,QAAQ,OAAO,SAAS,OAAO,QAAQ;AAAA,EACxD,SAAS,OAAO;AACd,kBAAc;AAAA,EAChB;AAEA,QAAM,UAA8B,CAAC;AACrC,aAAW,UAAU,QAAQ,SAAS;AACpC,UAAM,UACJ,gBAAgB,SACZ,MAAM,aAAa,QAAQ,MAAM,UAAU,QAAkB,MAAM,IACnE,YAAY,QAAQ,kBAAkB,aAAa,WAAW,CAAC,EAAE;AACvE,UAAM,iBAAiB,MAAM,cAAc;AAAA,MACzC,WAAW,QAAQ;AAAA,MACnB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,WAAW,QAAQ,aAAa,CAAC;AAAA,MACjC,qBAAqB,QAAQ,wBAAwB;AAAA,IACvD,CAAC;AACD,YAAQ,KAAK,EAAE,YAAY,OAAO,MAAM,SAAS,eAAe,CAAC;AAAA,EACnE;AAEA,SAAO;AAAA,IACL,MAAM;AAAA,IACN,GAAI,WAAW,SAAY,CAAC,IAAI,EAAE,OAAO;AAAA,IACzC,GAAI,gBAAgB,SAAY,CAAC,IAAI,EAAE,YAAY;AAAA,IACnD;AAAA,EACF;AACF;AAEA,eAAe,aACb,WACA,UACA,QACA,QAC0B;AAC1B,MAAI;AACF,WAAO,MAAM,OAAO,SAAS,EAAE,WAAW,MAAM,UAAU,OAAO,CAAC;AAAA,EACpE,SAAS,OAAO;AACd,WAAO,YAAY,QAAQ,aAAa,KAAK,CAAC;AAAA,EAChD;AACF;AAEA,eAAe,cAAuC,MAS/B;AACrB,QAAM,SAAoB,CAAC;AAC3B,aAAW,YAAY,KAAK,WAAW;AACrC,QAAI;AACF,YAAM,SAAS,OAAO;AAAA,QACpB,WAAW,KAAK;AAAA,QAChB,MAAM,KAAK;AAAA,QACX,QAAQ,KAAK;AAAA,QACb,aAAa,KAAK;AAAA,QAClB,QAAQ,KAAK;AAAA,QACb,SAAS,KAAK;AAAA,MAChB,CAAC;AAAA,IACH,SAAS,OAAO;AACd,UAAI,KAAK,qBAAqB;AAC5B,cAAM;AAAA,MACR;AACA,aAAO,KAAK,KAAK;AAAA,IACnB;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,cAAc,SAIrB;AACA,MAAI,SAAS;AACb,MAAI,SAAS;AACb,MAAI,UAAU;AACd,aAAW,UAAU,SAAS;AAC5B,eAAW,UAAU,OAAO,SAAS;AACnC,UAAI,OAAO,QAAQ,YAAY,OAAQ,WAAU;AACjD,UAAI,OAAO,QAAQ,YAAY,OAAQ,WAAU;AACjD,UAAI,OAAO,QAAQ,YAAY,UAAW,YAAW;AAAA,IACvD;AAAA,EACF;AACA,SAAO,EAAE,QAAQ,QAAQ,QAAQ;AACnC;","names":[]}
@@ -1,11 +1,11 @@
1
- import { e as Agent } from '../agent-0UeJ9Rad.js';
1
+ import { e as Agent } from '../agent-B-ls5y_g.js';
2
2
  import { U as Usage, M as Message, C as CompletionModel, e as JsonValue, m as ToolChoice } from '../types-C54aNoCd.js';
3
3
  import { Z as ZodSchema } from '../zod-schema-C7F4clpm.js';
4
4
  import '../memory/index.js';
5
5
  import '../observability/index.js';
6
6
  import '../tool-C3ciF-VG.js';
7
- import '../middleware-BQ7fkEEe.js';
8
- import '../embeddings/index.js';
7
+ import '../middleware-CGiEIaBx.js';
8
+ import '../types-IB2e9u5M.js';
9
9
  import '../vector-store/index.js';
10
10
  import 'zod';
11
11
 
@@ -2,16 +2,17 @@ import {
2
2
  ExtractionError,
3
3
  Extractor,
4
4
  ExtractorBuilder
5
- } from "../chunk-PP4VIN3Y.js";
6
- import "../chunk-6U6PZ5MD.js";
7
- import "../chunk-TP32W7XT.js";
5
+ } from "../chunk-LGETU3RG.js";
6
+ import "../chunk-MOICCK3J.js";
7
+ import "../chunk-N7NMSGZI.js";
8
8
  import "../chunk-XUUY2L2D.js";
9
9
  import "../chunk-XXT2UCAR.js";
10
10
  import "../chunk-YK4WAAS4.js";
11
- import "../chunk-I2HOMD3R.js";
12
- import "../chunk-65QV627O.js";
13
- import "../chunk-A7VDIZQN.js";
11
+ import "../chunk-35GF7P43.js";
12
+ import "../chunk-P425B6GR.js";
14
13
  import "../chunk-6GJDBBDC.js";
14
+ import "../chunk-EFGX3EX5.js";
15
+ import "../chunk-OIMLU4SF.js";
15
16
  export {
16
17
  ExtractionError,
17
18
  Extractor,
package/dist/index.d.ts CHANGED
@@ -1,5 +1,5 @@
1
1
  export { AgentBuilder, MaxTurnsError, PromptCancelledError } from './agent/index.js';
2
- export { A as AgentChildStreamEvent, a as AgentStreamEvent, P as PromptResponse, c as cancelPrompt, b as createHook, r as requestToolApproval, d as runControl, s as skipTool, t as toolCallControl } from './agent-0UeJ9Rad.js';
2
+ export { A as AgentChildStreamEvent, a as AgentStreamEvent, P as PromptResponse, c as cancelPrompt, b as createHook, r as requestToolApproval, d as runControl, s as skipTool, t as toolCallControl } from './agent-B-ls5y_g.js';
3
3
  export { A as AssistantContent, a as AssistantMessage, C as CompletionModel, b as CompletionRequest, c as CompletionResponse, D as Document, I as ImageContent, J as JsonObject, d as JsonPrimitive, e as JsonValue, M as Message, S as SystemMessage, T as Text, f as ToolCall, g as ToolDefinition, h as ToolMessage, i as ToolResult, j as ToolResultContent, U as Usage, k as UserContent, l as UserMessage } from './types-C54aNoCd.js';
4
4
  export { MemoryStore } from './memory/index.js';
5
5
  export { Z as ZodSchema } from './zod-schema-C7F4clpm.js';
@@ -12,7 +12,7 @@ import '@modelcontextprotocol/sdk/client/sse.js';
12
12
  import '@modelcontextprotocol/sdk/client/stdio.js';
13
13
  import '@modelcontextprotocol/sdk/client/streamableHttp.js';
14
14
  import './observability/index.js';
15
- import './middleware-BQ7fkEEe.js';
16
- import './embeddings/index.js';
15
+ import './middleware-CGiEIaBx.js';
16
+ import './types-IB2e9u5M.js';
17
17
  import './vector-store/index.js';
18
18
  import 'zod';
package/dist/index.js CHANGED
@@ -2,10 +2,10 @@ import {
2
2
  SkillValidationError,
3
3
  loadSkills,
4
4
  skill
5
- } from "./chunk-TILE6Z2N.js";
5
+ } from "./chunk-DUP7FMAF.js";
6
6
  import {
7
7
  AgentBuilder
8
- } from "./chunk-6U6PZ5MD.js";
8
+ } from "./chunk-MOICCK3J.js";
9
9
  import {
10
10
  MaxTurnsError,
11
11
  PromptCancelledError,
@@ -15,23 +15,24 @@ import {
15
15
  runControl,
16
16
  skipTool,
17
17
  toolCallControl
18
- } from "./chunk-TP32W7XT.js";
18
+ } from "./chunk-N7NMSGZI.js";
19
19
  import "./chunk-XUUY2L2D.js";
20
20
  import "./chunk-XXT2UCAR.js";
21
21
  import "./chunk-YK4WAAS4.js";
22
22
  import {
23
23
  createThinkTool
24
- } from "./chunk-I2HOMD3R.js";
24
+ } from "./chunk-35GF7P43.js";
25
25
  import {
26
26
  createTool
27
- } from "./chunk-65QV627O.js";
28
- import "./chunk-A7VDIZQN.js";
27
+ } from "./chunk-P425B6GR.js";
29
28
  import {
30
29
  AssistantContent,
31
30
  Message,
32
31
  Usage,
33
32
  UserContent
34
33
  } from "./chunk-6GJDBBDC.js";
34
+ import "./chunk-EFGX3EX5.js";
35
+ import "./chunk-OIMLU4SF.js";
35
36
  export {
36
37
  AgentBuilder,
37
38
  AssistantContent,
@@ -1,8 +1,8 @@
1
- export { e as Agent, j as AgentEventAppendInput, k as AgentEventRecord, h as AgentEventStore, l as AgentEventStoreInclude, i as AgentEventStoreOptions, x as AgentEventStoreRegistration, y as AgentOptions, z as AgentSession, B as AgentToolOptions, E as DEFAULT_MAX_TURNS, D as DynamicContextOptions, F as DynamicContextRegistration, f as DynamicToolOptions, G as DynamicToolRegistration } from '../agent-0UeJ9Rad.js';
1
+ export { e as Agent, j as AgentEventAppendInput, k as AgentEventRecord, h as AgentEventStore, l as AgentEventStoreInclude, i as AgentEventStoreOptions, x as AgentEventStoreRegistration, y as AgentOptions, z as AgentSession, B as AgentToolOptions, E as DEFAULT_MAX_TURNS, D as DynamicContextOptions, F as DynamicContextRegistration, f as DynamicToolOptions, G as DynamicToolRegistration } from '../agent-B-ls5y_g.js';
2
2
  import '../types-C54aNoCd.js';
3
3
  import '../memory/index.js';
4
4
  import '../observability/index.js';
5
5
  import '../tool-C3ciF-VG.js';
6
- import '../middleware-BQ7fkEEe.js';
7
- import '../embeddings/index.js';
6
+ import '../middleware-CGiEIaBx.js';
7
+ import '../types-IB2e9u5M.js';
8
8
  import '../vector-store/index.js';
@@ -2,13 +2,14 @@ import {
2
2
  Agent,
3
3
  AgentSession,
4
4
  DEFAULT_MAX_TURNS
5
- } from "../chunk-TP32W7XT.js";
5
+ } from "../chunk-N7NMSGZI.js";
6
6
  import "../chunk-XUUY2L2D.js";
7
7
  import "../chunk-YK4WAAS4.js";
8
- import "../chunk-I2HOMD3R.js";
9
- import "../chunk-65QV627O.js";
10
- import "../chunk-A7VDIZQN.js";
8
+ import "../chunk-35GF7P43.js";
9
+ import "../chunk-P425B6GR.js";
11
10
  import "../chunk-6GJDBBDC.js";
11
+ import "../chunk-EFGX3EX5.js";
12
+ import "../chunk-OIMLU4SF.js";
12
13
  export {
13
14
  Agent,
14
15
  AgentSession,
@@ -36,6 +36,19 @@ type PdfPageWithPath = {
36
36
  pageNumber: number;
37
37
  text: string;
38
38
  };
39
+
40
+ declare function fileToDocument(file: FileReadWithPath): Document;
41
+ declare function fileLoaderToDocuments(loader: AsyncIterable<FileReadWithPath>): Promise<Document[]>;
42
+ declare function pdfToDocument(pdf: PdfReadWithPath): Document;
43
+ declare function pdfLoaderToDocuments(loader: AsyncIterable<PdfReadWithPath>): Promise<Document[]>;
44
+ declare function pdfPageToDocument(page: PdfPageWithPath): Document;
45
+ declare function pdfPageLoaderToDocuments(loader: AsyncIterable<PdfPageWithPath>): Promise<Document[]>;
46
+
47
+ type LoaderValue$1<T> = T extends {
48
+ ok: true;
49
+ value: infer Value;
50
+ } ? Value : never;
51
+ type UnwrapLoaderResult$1<T> = [LoaderValue$1<T>] extends [never] ? T : LoaderValue$1<T>;
39
52
  declare class FileLoader<T = LoaderResult<FileSource>> implements AsyncIterable<T> {
40
53
  private readonly source;
41
54
  private readonly mode;
@@ -47,9 +60,15 @@ declare class FileLoader<T = LoaderResult<FileSource>> implements AsyncIterable<
47
60
  static fromBytesMany(bytes: Array<Uint8Array | ArrayBuffer>): FileLoader<LoaderResult<FileSource>>;
48
61
  read(): FileLoader<LoaderResult<string>>;
49
62
  readWithPath(): FileLoader<LoaderResult<FileReadWithPath>>;
50
- ignoreErrors(): FileLoader<UnwrapLoaderResult<T>>;
63
+ ignoreErrors(): FileLoader<UnwrapLoaderResult$1<T>>;
51
64
  [Symbol.asyncIterator](): AsyncIterator<T>;
52
65
  }
66
+
67
+ type LoaderValue<T> = T extends {
68
+ ok: true;
69
+ value: infer Value;
70
+ } ? Value : never;
71
+ type UnwrapLoaderResult<T> = [LoaderValue<T>] extends [never] ? T : LoaderValue<T>;
53
72
  declare class PdfFileLoader<T = LoaderResult<PdfSource>> implements AsyncIterable<T> {
54
73
  private readonly source;
55
74
  private readonly mode;
@@ -67,16 +86,5 @@ declare class PdfFileLoader<T = LoaderResult<PdfSource>> implements AsyncIterabl
67
86
  ignoreErrors(): PdfFileLoader<UnwrapLoaderResult<T>>;
68
87
  [Symbol.asyncIterator](): AsyncIterator<T>;
69
88
  }
70
- declare function fileToDocument(file: FileReadWithPath): Document;
71
- declare function fileLoaderToDocuments(loader: AsyncIterable<FileReadWithPath>): Promise<Document[]>;
72
- declare function pdfToDocument(pdf: PdfReadWithPath): Document;
73
- declare function pdfLoaderToDocuments(loader: AsyncIterable<PdfReadWithPath>): Promise<Document[]>;
74
- declare function pdfPageToDocument(page: PdfPageWithPath): Document;
75
- declare function pdfPageLoaderToDocuments(loader: AsyncIterable<PdfPageWithPath>): Promise<Document[]>;
76
- type LoaderValue<T> = T extends {
77
- ok: true;
78
- value: infer Value;
79
- } ? Value : never;
80
- type UnwrapLoaderResult<T> = [LoaderValue<T>] extends [never] ? T : LoaderValue<T>;
81
89
 
82
90
  export { FileLoader, type FileReadWithPath, type FileSource, type LoaderResult, PdfFileLoader, type PdfPage, type PdfPageWithPath, type PdfReadWithPath, type PdfSource, fileLoaderToDocuments, fileToDocument, pdfLoaderToDocuments, pdfPageLoaderToDocuments, pdfPageToDocument, pdfToDocument };