llm-testrunner-components 1.2.4 → 1.3.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +9 -5
- package/dist/cjs/app-chips_5.cjs.entry.js +1 -1
- package/dist/cjs/app-chips_5.cjs.entry.js.map +1 -1
- package/dist/cjs/index.cjs.js +464 -66
- package/dist/cjs/index.cjs.js.map +1 -1
- package/dist/cjs/llm-testrunner.cjs.js +1 -1
- package/dist/cjs/loader.cjs.js +1 -1
- package/dist/collection/components/llm-test-runner/llm-test-runner.js +46 -13
- package/dist/collection/components/llm-test-runner/llm-test-runner.js.map +1 -1
- package/dist/collection/components/llm-test-runner/test-cases/chat-history.css +5 -5
- package/dist/collection/components/llm-test-runner/test-cases/expected-outcome-renderer.js +45 -5
- package/dist/collection/components/llm-test-runner/test-cases/expected-outcome-renderer.js.map +1 -1
- package/dist/collection/components/llm-test-runner/test-cases/llm-test-case-row.css +21 -0
- package/dist/collection/components/llm-test-runner/test-cases/llm-test-case-row.js +2 -2
- package/dist/collection/components/llm-test-runner/test-cases/llm-test-case-row.js.map +1 -1
- package/dist/collection/components/llm-test-runner/test-cases/llm-test-cases.js +2 -2
- package/dist/collection/components/llm-test-runner/test-cases/llm-test-cases.js.map +1 -1
- package/dist/collection/components/llm-test-runner/test-cases/output/response-output.js +1 -1
- package/dist/collection/components/llm-test-runner/test-cases/output/response-output.js.map +1 -1
- package/dist/collection/lib/evaluation/actual-value-resolver.js +52 -0
- package/dist/collection/lib/evaluation/actual-value-resolver.js.map +1 -0
- package/dist/collection/lib/evaluation/evaluation-engine.js +1 -1
- package/dist/collection/lib/evaluation/evaluation-engine.js.map +1 -1
- package/dist/collection/lib/evaluation/evaluation-service.js +55 -17
- package/dist/collection/lib/evaluation/evaluation-service.js.map +1 -1
- package/dist/collection/lib/evaluation/types.js.map +1 -1
- package/dist/collection/lib/import-export/test-suite-importer.js +7 -1
- package/dist/collection/lib/import-export/test-suite-importer.js.map +1 -1
- package/dist/collection/lib/test-cases/test-case-factory.js +5 -0
- package/dist/collection/lib/test-cases/test-case-factory.js.map +1 -1
- package/dist/collection/lib/test-cases/test-case-mutations.js +58 -23
- package/dist/collection/lib/test-cases/test-case-mutations.js.map +1 -1
- package/dist/collection/schemas/expected-outcome.js +39 -0
- package/dist/collection/schemas/expected-outcome.js.map +1 -1
- package/dist/collection/schemas/model-response.js +7 -0
- package/dist/collection/schemas/model-response.js.map +1 -0
- package/dist/collection/schemas/test-case.js +2 -1
- package/dist/collection/schemas/test-case.js.map +1 -1
- package/dist/collection/types/expected-outcome.js.map +1 -1
- package/dist/collection/types/llm-test-runner.js.map +1 -1
- package/dist/components/chat-history.js +1 -1
- package/dist/components/index.js +1 -1
- package/dist/components/llm-test-runner.js +1 -1
- package/dist/components/p-kmtfMXcQ.js +2 -0
- package/dist/components/p-kmtfMXcQ.js.map +1 -0
- package/dist/components/{p-B87Lt3z4.js → p-wzA48RFK.js} +3 -3
- package/dist/components/p-wzA48RFK.js.map +1 -0
- package/dist/esm/app-chips_5.entry.js +1 -1
- package/dist/esm/app-chips_5.entry.js.map +1 -1
- package/dist/esm/index.js +464 -66
- package/dist/esm/index.js.map +1 -1
- package/dist/esm/llm-testrunner.js +1 -1
- package/dist/esm/loader.js +1 -1
- package/dist/llm-testrunner/index.esm.js +2 -2
- package/dist/llm-testrunner/index.esm.js.map +1 -1
- package/dist/llm-testrunner/llm-testrunner.esm.js +1 -1
- package/dist/llm-testrunner/{p-21202f12.entry.js → p-5bf1fc78.entry.js} +2 -2
- package/dist/llm-testrunner/{p-21202f12.entry.js.map → p-5bf1fc78.entry.js.map} +1 -1
- package/dist/types/components/llm-test-runner/llm-test-runner.d.ts +3 -4
- package/dist/types/components/llm-test-runner/test-cases/expected-outcome-renderer.d.ts +1 -0
- package/dist/types/components/llm-test-runner/test-cases/llm-test-case-row.d.ts +1 -0
- package/dist/types/components/llm-test-runner/test-cases/llm-test-cases.d.ts +1 -0
- package/dist/types/components/llm-test-runner/test-cases/output/response-output.d.ts +2 -1
- package/dist/types/components.d.ts +4 -2
- package/dist/types/lib/evaluation/actual-value-resolver.d.ts +9 -0
- package/dist/types/lib/evaluation/evaluation-service.d.ts +2 -2
- package/dist/types/lib/evaluation/types.d.ts +1 -1
- package/dist/types/lib/import-export/test-suite-importer.d.ts +1 -1
- package/dist/types/lib/test-cases/test-case-mutations.d.ts +10 -1
- package/dist/types/schemas/expected-outcome.d.ts +116 -0
- package/dist/types/schemas/model-response.d.ts +7 -0
- package/dist/types/schemas/test-case.d.ts +76 -1
- package/dist/types/types/expected-outcome.d.ts +1 -1
- package/dist/types/types/llm-test-runner.d.ts +4 -2
- package/package.json +1 -1
- package/dist/components/p-B87Lt3z4.js.map +0 -1
- package/dist/components/p-Bx2jqguC.js +0 -2
- package/dist/components/p-Bx2jqguC.js.map +0 -1
|
@@ -10,6 +10,12 @@ export declare const testCaseInputSchema: z.ZodObject<{
|
|
|
10
10
|
expectedOutcome: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
11
11
|
label: z.ZodString;
|
|
12
12
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
13
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
14
|
+
type: z.ZodLiteral<"text">;
|
|
15
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
16
|
+
type: z.ZodLiteral<"custom">;
|
|
17
|
+
extractorId: z.ZodString;
|
|
18
|
+
}, z.core.$strip>], "type">>;
|
|
13
19
|
type: z.ZodLiteral<"text">;
|
|
14
20
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
15
21
|
approach: z.ZodEnum<typeof import("../lib/evaluation/constants").EvaluationApproach>;
|
|
@@ -19,6 +25,12 @@ export declare const testCaseInputSchema: z.ZodObject<{
|
|
|
19
25
|
}, z.core.$strip>, z.ZodObject<{
|
|
20
26
|
label: z.ZodString;
|
|
21
27
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
28
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
29
|
+
type: z.ZodLiteral<"text">;
|
|
30
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
31
|
+
type: z.ZodLiteral<"custom">;
|
|
32
|
+
extractorId: z.ZodString;
|
|
33
|
+
}, z.core.$strip>], "type">>;
|
|
22
34
|
type: z.ZodLiteral<"textarea">;
|
|
23
35
|
rows: z.ZodOptional<z.ZodNumber>;
|
|
24
36
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
@@ -34,6 +46,12 @@ export declare const testCaseInputSchema: z.ZodObject<{
|
|
|
34
46
|
}, z.core.$strip>, z.ZodObject<{
|
|
35
47
|
label: z.ZodString;
|
|
36
48
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
49
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
50
|
+
type: z.ZodLiteral<"text">;
|
|
51
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
52
|
+
type: z.ZodLiteral<"custom">;
|
|
53
|
+
extractorId: z.ZodString;
|
|
54
|
+
}, z.core.$strip>], "type">>;
|
|
37
55
|
type: z.ZodLiteral<"chips-input">;
|
|
38
56
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
39
57
|
approach: z.ZodEnum<typeof import("../lib/evaluation/constants").EvaluationApproach>;
|
|
@@ -43,6 +61,12 @@ export declare const testCaseInputSchema: z.ZodObject<{
|
|
|
43
61
|
}, z.core.$strip>, z.ZodObject<{
|
|
44
62
|
label: z.ZodString;
|
|
45
63
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
64
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
65
|
+
type: z.ZodLiteral<"text">;
|
|
66
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
67
|
+
type: z.ZodLiteral<"custom">;
|
|
68
|
+
extractorId: z.ZodString;
|
|
69
|
+
}, z.core.$strip>], "type">>;
|
|
46
70
|
type: z.ZodLiteral<"select">;
|
|
47
71
|
options: z.ZodArray<z.ZodString>;
|
|
48
72
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
@@ -62,6 +86,12 @@ export declare const testCaseInputArraySchema: z.ZodArray<z.ZodObject<{
|
|
|
62
86
|
expectedOutcome: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
63
87
|
label: z.ZodString;
|
|
64
88
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
89
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
90
|
+
type: z.ZodLiteral<"text">;
|
|
91
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
92
|
+
type: z.ZodLiteral<"custom">;
|
|
93
|
+
extractorId: z.ZodString;
|
|
94
|
+
}, z.core.$strip>], "type">>;
|
|
65
95
|
type: z.ZodLiteral<"text">;
|
|
66
96
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
67
97
|
approach: z.ZodEnum<typeof import("../lib/evaluation/constants").EvaluationApproach>;
|
|
@@ -71,6 +101,12 @@ export declare const testCaseInputArraySchema: z.ZodArray<z.ZodObject<{
|
|
|
71
101
|
}, z.core.$strip>, z.ZodObject<{
|
|
72
102
|
label: z.ZodString;
|
|
73
103
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
104
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
105
|
+
type: z.ZodLiteral<"text">;
|
|
106
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
107
|
+
type: z.ZodLiteral<"custom">;
|
|
108
|
+
extractorId: z.ZodString;
|
|
109
|
+
}, z.core.$strip>], "type">>;
|
|
74
110
|
type: z.ZodLiteral<"textarea">;
|
|
75
111
|
rows: z.ZodOptional<z.ZodNumber>;
|
|
76
112
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
@@ -86,6 +122,12 @@ export declare const testCaseInputArraySchema: z.ZodArray<z.ZodObject<{
|
|
|
86
122
|
}, z.core.$strip>, z.ZodObject<{
|
|
87
123
|
label: z.ZodString;
|
|
88
124
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
125
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
126
|
+
type: z.ZodLiteral<"text">;
|
|
127
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
128
|
+
type: z.ZodLiteral<"custom">;
|
|
129
|
+
extractorId: z.ZodString;
|
|
130
|
+
}, z.core.$strip>], "type">>;
|
|
89
131
|
type: z.ZodLiteral<"chips-input">;
|
|
90
132
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
91
133
|
approach: z.ZodEnum<typeof import("../lib/evaluation/constants").EvaluationApproach>;
|
|
@@ -95,6 +137,12 @@ export declare const testCaseInputArraySchema: z.ZodArray<z.ZodObject<{
|
|
|
95
137
|
}, z.core.$strip>, z.ZodObject<{
|
|
96
138
|
label: z.ZodString;
|
|
97
139
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
140
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
141
|
+
type: z.ZodLiteral<"text">;
|
|
142
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
143
|
+
type: z.ZodLiteral<"custom">;
|
|
144
|
+
extractorId: z.ZodString;
|
|
145
|
+
}, z.core.$strip>], "type">>;
|
|
98
146
|
type: z.ZodLiteral<"select">;
|
|
99
147
|
options: z.ZodArray<z.ZodString>;
|
|
100
148
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
@@ -114,6 +162,12 @@ export declare const testCaseSchema: z.ZodObject<{
|
|
|
114
162
|
expectedOutcome: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
115
163
|
label: z.ZodString;
|
|
116
164
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
165
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
166
|
+
type: z.ZodLiteral<"text">;
|
|
167
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
168
|
+
type: z.ZodLiteral<"custom">;
|
|
169
|
+
extractorId: z.ZodString;
|
|
170
|
+
}, z.core.$strip>], "type">>;
|
|
117
171
|
type: z.ZodLiteral<"text">;
|
|
118
172
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
119
173
|
approach: z.ZodEnum<typeof import("../lib/evaluation/constants").EvaluationApproach>;
|
|
@@ -123,6 +177,12 @@ export declare const testCaseSchema: z.ZodObject<{
|
|
|
123
177
|
}, z.core.$strip>, z.ZodObject<{
|
|
124
178
|
label: z.ZodString;
|
|
125
179
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
180
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
181
|
+
type: z.ZodLiteral<"text">;
|
|
182
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
183
|
+
type: z.ZodLiteral<"custom">;
|
|
184
|
+
extractorId: z.ZodString;
|
|
185
|
+
}, z.core.$strip>], "type">>;
|
|
126
186
|
type: z.ZodLiteral<"textarea">;
|
|
127
187
|
rows: z.ZodOptional<z.ZodNumber>;
|
|
128
188
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
@@ -138,6 +198,12 @@ export declare const testCaseSchema: z.ZodObject<{
|
|
|
138
198
|
}, z.core.$strip>, z.ZodObject<{
|
|
139
199
|
label: z.ZodString;
|
|
140
200
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
201
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
202
|
+
type: z.ZodLiteral<"text">;
|
|
203
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
204
|
+
type: z.ZodLiteral<"custom">;
|
|
205
|
+
extractorId: z.ZodString;
|
|
206
|
+
}, z.core.$strip>], "type">>;
|
|
141
207
|
type: z.ZodLiteral<"chips-input">;
|
|
142
208
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
143
209
|
approach: z.ZodEnum<typeof import("../lib/evaluation/constants").EvaluationApproach>;
|
|
@@ -147,6 +213,12 @@ export declare const testCaseSchema: z.ZodObject<{
|
|
|
147
213
|
}, z.core.$strip>, z.ZodObject<{
|
|
148
214
|
label: z.ZodString;
|
|
149
215
|
placeholder: z.ZodOptional<z.ZodString>;
|
|
216
|
+
evaluationSource: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
217
|
+
type: z.ZodLiteral<"text">;
|
|
218
|
+
}, z.core.$strip>, z.ZodObject<{
|
|
219
|
+
type: z.ZodLiteral<"custom">;
|
|
220
|
+
extractorId: z.ZodString;
|
|
221
|
+
}, z.core.$strip>], "type">>;
|
|
150
222
|
type: z.ZodLiteral<"select">;
|
|
151
223
|
options: z.ZodArray<z.ZodString>;
|
|
152
224
|
evaluationParameters: z.ZodOptional<z.ZodObject<{
|
|
@@ -155,11 +227,14 @@ export declare const testCaseSchema: z.ZodObject<{
|
|
|
155
227
|
}, z.core.$strip>>;
|
|
156
228
|
value: z.ZodString;
|
|
157
229
|
}, z.core.$strip>], "type">>;
|
|
230
|
+
output: z.ZodOptional<z.ZodObject<{
|
|
231
|
+
text: z.ZodOptional<z.ZodString>;
|
|
232
|
+
metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
233
|
+
}, z.core.$strip>>;
|
|
158
234
|
chatHistory: z.ZodObject<{
|
|
159
235
|
enabled: z.ZodBoolean;
|
|
160
236
|
value: z.ZodString;
|
|
161
237
|
}, z.core.$strip>;
|
|
162
|
-
output: z.ZodOptional<z.ZodString>;
|
|
163
238
|
isRunning: z.ZodOptional<z.ZodBoolean>;
|
|
164
239
|
error: z.ZodOptional<z.ZodString>;
|
|
165
240
|
evaluationResult: z.ZodOptional<z.ZodCustom<EvaluationResult, EvaluationResult>>;
|
|
@@ -1 +1 @@
|
|
|
1
|
-
export type { ExpectedOutcomeMode, ExpectedOutcomeSchemaField, ExpectedOutcomeSchema, ExpectedOutcomeField, ExpectedOutcomeFieldType, ExpectedOutcomeBase, TextExpectedOutcomeSchemaField, TextareaExpectedOutcomeSchemaField, ChipsExpectedOutcomeSchemaField, SelectExpectedOutcomeSchemaField, TextExpectedOutcomeField, TextareaExpectedOutcomeField, ChipsExpectedOutcomeField, SelectExpectedOutcomeField, } from '../schemas/expected-outcome';
|
|
1
|
+
export type { EvaluationSource, EvaluationSourceExtractor, EvaluationSourceExtractors, ExpectedOutcomeMode, ExpectedOutcomeSchemaField, ExpectedOutcomeSchema, ExpectedOutcomeField, ExpectedOutcomeFieldType, ExpectedOutcomeBase, TextExpectedOutcomeSchemaField, TextareaExpectedOutcomeSchemaField, ChipsExpectedOutcomeSchemaField, SelectExpectedOutcomeSchemaField, TextExpectedOutcomeField, TextareaExpectedOutcomeField, ChipsExpectedOutcomeField, SelectExpectedOutcomeField, } from '../schemas/expected-outcome';
|
|
@@ -1,10 +1,12 @@
|
|
|
1
1
|
import type { TestCase } from './test-case';
|
|
2
|
-
|
|
2
|
+
import type { ModelResponsePayload } from '../schemas/model-response';
|
|
3
|
+
export type { EvaluationSource, EvaluationSourceExtractor, EvaluationSourceExtractors, ExpectedOutcomeMode, ExpectedOutcomeFieldType, ExpectedOutcomeBase, ExpectedOutcomeSchema, ExpectedOutcomeSchemaField, ExpectedOutcomeField, TextExpectedOutcomeSchemaField, TextareaExpectedOutcomeSchemaField, ChipsExpectedOutcomeSchemaField, SelectExpectedOutcomeSchemaField, TextExpectedOutcomeField, TextareaExpectedOutcomeField, ChipsExpectedOutcomeField, SelectExpectedOutcomeField, } from './expected-outcome';
|
|
4
|
+
export type { ModelResponsePayload, } from '../schemas/model-response';
|
|
3
5
|
export type { TestCaseChatHistory, TestCase, TestCaseInput, } from './test-case';
|
|
4
6
|
export interface LLMRequestPayload {
|
|
5
7
|
prompt: string;
|
|
8
|
+
resolve: (result: ModelResponsePayload) => void;
|
|
6
9
|
chatHistory?: string;
|
|
7
|
-
resolve: (result: string) => void;
|
|
8
10
|
reject: (err: Error | unknown) => void;
|
|
9
11
|
}
|
|
10
12
|
export interface SavePayload {
|