llm-testrunner-components 1.2.3 → 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_4.cjs.entry.js → app-chips_5.cjs.entry.js} +38 -6
- package/dist/cjs/app-chips_5.cjs.entry.js.map +1 -0
- package/dist/cjs/index.cjs.js +499 -68
- 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/collection-manifest.json +1 -0
- package/dist/collection/components/llm-test-runner/llm-test-runner.import-export.test.js +22 -12
- package/dist/collection/components/llm-test-runner/llm-test-runner.import-export.test.js.map +1 -1
- package/dist/collection/components/llm-test-runner/llm-test-runner.js +59 -15
- 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 +101 -0
- package/dist/collection/components/llm-test-runner/test-cases/chat-history.js +105 -0
- package/dist/collection/components/llm-test-runner/test-cases/chat-history.js.map +1 -0
- 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 +12 -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/index.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/form/components/app-chips.js +1 -1
- package/dist/collection/lib/form/components/app-select.js +1 -1
- package/dist/collection/lib/form/components/app-textarea.js +2 -2
- package/dist/collection/lib/import-export/test-suite-exporter.js +4 -0
- package/dist/collection/lib/import-export/test-suite-exporter.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 +7 -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 +8 -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/collection/types/test-case.js.map +1 -1
- package/dist/components/app-chips.js +1 -1
- package/dist/components/app-select.js +1 -1
- package/dist/components/app-textarea.js +1 -1
- package/dist/components/chat-history.d.ts +11 -0
- package/dist/components/chat-history.js +2 -0
- package/dist/components/chat-history.js.map +1 -0
- package/dist/components/index.js +1 -1
- package/dist/components/llm-test-runner.js +1 -1
- package/dist/components/{p-CVtKFBJl.js → p-D2qDAxFN.js} +2 -2
- package/dist/components/{p-Dv7cB5FU.js → p-D4dHUFN9.js} +2 -2
- package/dist/components/{p-CE5-1jfZ.js → p-eN2dLrsr.js} +2 -2
- package/dist/components/p-kmtfMXcQ.js +2 -0
- package/dist/components/p-kmtfMXcQ.js.map +1 -0
- package/dist/components/{p-BcygfrMf.js → p-wzA48RFK.js} +3 -3
- package/dist/components/p-wzA48RFK.js.map +1 -0
- package/dist/esm/{app-chips_4.entry.js → app-chips_5.entry.js} +38 -7
- package/dist/esm/app-chips_5.entry.js.map +1 -0
- package/dist/esm/index.js +499 -68
- 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-5bf1fc78.entry.js +2 -0
- package/dist/llm-testrunner/p-5bf1fc78.entry.js.map +1 -0
- package/dist/react/components.d.ts +6 -1
- package/dist/react/components.d.ts.map +1 -1
- package/dist/react/components.js +9 -0
- package/dist/types/components/llm-test-runner/llm-test-runner.d.ts +4 -4
- package/dist/types/components/llm-test-runner/test-cases/chat-history.d.ts +14 -0
- 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 +6 -0
- package/dist/types/components/llm-test-runner/test-cases/llm-test-cases.d.ts +3 -0
- package/dist/types/components/llm-test-runner/test-cases/output/response-output.d.ts +2 -1
- package/dist/types/components.d.ts +55 -2
- package/dist/types/index.d.ts +1 -1
- 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-exporter.d.ts +4 -0
- 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 +93 -1
- package/dist/types/types/expected-outcome.d.ts +1 -1
- package/dist/types/types/llm-test-runner.d.ts +6 -3
- package/dist/types/types/test-case.d.ts +1 -1
- package/package.json +1 -1
- package/dist/cjs/app-chips_4.cjs.entry.js.map +0 -1
- package/dist/components/p-BcygfrMf.js.map +0 -1
- package/dist/esm/app-chips_4.entry.js.map +0 -1
- package/dist/llm-testrunner/p-5df053b4.entry.js +0 -2
- package/dist/llm-testrunner/p-5df053b4.entry.js.map +0 -1
- /package/dist/components/{p-CVtKFBJl.js.map → p-D2qDAxFN.js.map} +0 -0
- /package/dist/components/{p-Dv7cB5FU.js.map → p-D4dHUFN9.js.map} +0 -0
- /package/dist/components/{p-CE5-1jfZ.js.map → p-eN2dLrsr.js.map} +0 -0
|
@@ -19,7 +19,7 @@ var patchBrowser = () => {
|
|
|
19
19
|
|
|
20
20
|
patchBrowser().then(async (options) => {
|
|
21
21
|
await appGlobals.globalScripts();
|
|
22
|
-
return index.bootstrapLazy([["app-
|
|
22
|
+
return index.bootstrapLazy([["app-chips_5.cjs",[[513,"llm-test-runner",{"delayMs":[2,"delay-ms"],"useSave":[4,"use-save"],"usePromptEditor":[4,"use-prompt-editor"],"resolveExpectedOutcome":[16],"evaluationSourceExtractors":[16],"initialTestCases":[16],"defaultExpectedOutcomeSchema":[16],"testCases":[32],"isRunningAll":[32],"error":[32],"isExportingTestSuite":[32],"isExportingTestResults":[32],"isSaving":[32],"resetSavingState":[64],"getTestCases":[64]}],[513,"app-chips",{"value":[16],"config":[16]}],[513,"app-select",{"value":[1],"config":[16]}],[513,"app-textarea",{"value":[1],"config":[16]}],[513,"chat-history",{"chatHistoryEnabled":[4,"chat-history-enabled"],"chatHistoryValue":[1,"chat-history-value"]}]]]], options);
|
|
23
23
|
});
|
|
24
24
|
|
|
25
25
|
exports.setNonce = index.setNonce;
|
package/dist/cjs/loader.cjs.js
CHANGED
|
@@ -6,7 +6,7 @@ var appGlobals = require('./app-globals-Chb-oJtg.js');
|
|
|
6
6
|
const defineCustomElements = async (win, options) => {
|
|
7
7
|
if (typeof window === 'undefined') return undefined;
|
|
8
8
|
await appGlobals.globalScripts();
|
|
9
|
-
return index.bootstrapLazy([["app-
|
|
9
|
+
return index.bootstrapLazy([["app-chips_5.cjs",[[513,"llm-test-runner",{"delayMs":[2,"delay-ms"],"useSave":[4,"use-save"],"usePromptEditor":[4,"use-prompt-editor"],"resolveExpectedOutcome":[16],"evaluationSourceExtractors":[16],"initialTestCases":[16],"defaultExpectedOutcomeSchema":[16],"testCases":[32],"isRunningAll":[32],"error":[32],"isExportingTestSuite":[32],"isExportingTestResults":[32],"isSaving":[32],"resetSavingState":[64],"getTestCases":[64]}],[513,"app-chips",{"value":[16],"config":[16]}],[513,"app-select",{"value":[1],"config":[16]}],[513,"app-textarea",{"value":[1],"config":[16]}],[513,"chat-history",{"chatHistoryEnabled":[4,"chat-history-enabled"],"chatHistoryValue":[1,"chat-history-value"]}]]]], options);
|
|
10
10
|
};
|
|
11
11
|
|
|
12
12
|
exports.setNonce = index.setNonce;
|
|
@@ -15,6 +15,7 @@ import { downloadFile } from "../../lib/file/file-download";
|
|
|
15
15
|
import { formatTestSuiteAsJson } from "../../lib/import-export/test-suite-exporter";
|
|
16
16
|
import { importTestSuite } from "../../lib/import-export/test-suite-importer";
|
|
17
17
|
import { EvaluationApproach } from "../../lib/evaluation/constants";
|
|
18
|
+
const emptyChatHistory = { enabled: false, value: '' };
|
|
18
19
|
describe('llm-test-runner import/export', () => {
|
|
19
20
|
let page;
|
|
20
21
|
let component;
|
|
@@ -30,7 +31,8 @@ describe('llm-test-runner import/export', () => {
|
|
|
30
31
|
id: '1',
|
|
31
32
|
question: 'What is AI?',
|
|
32
33
|
expectedOutcome: buildExpectedOutcome('artificial intelligence'),
|
|
33
|
-
|
|
34
|
+
chatHistory: { enabled: false, value: '' },
|
|
35
|
+
isRunning: false,
|
|
34
36
|
};
|
|
35
37
|
return {
|
|
36
38
|
...defaults,
|
|
@@ -56,8 +58,9 @@ describe('llm-test-runner import/export', () => {
|
|
|
56
58
|
id: '1',
|
|
57
59
|
question: 'What is AI?',
|
|
58
60
|
expectedOutcome: buildExpectedOutcome('artificial intelligence'),
|
|
59
|
-
|
|
60
|
-
|
|
61
|
+
chatHistory: { ...emptyChatHistory },
|
|
62
|
+
isRunning: false,
|
|
63
|
+
},
|
|
61
64
|
];
|
|
62
65
|
const mockFile = createMockFile(JSON.stringify(mockTestData), 'test-suite.json');
|
|
63
66
|
// Mock the utility functions
|
|
@@ -81,8 +84,9 @@ describe('llm-test-runner import/export', () => {
|
|
|
81
84
|
id: '1',
|
|
82
85
|
question: 'Test question',
|
|
83
86
|
expectedOutcome: buildExpectedOutcome('test'),
|
|
84
|
-
|
|
85
|
-
|
|
87
|
+
chatHistory: { ...emptyChatHistory },
|
|
88
|
+
isRunning: false,
|
|
89
|
+
},
|
|
86
90
|
];
|
|
87
91
|
const normalizedImportData = [
|
|
88
92
|
{
|
|
@@ -98,6 +102,7 @@ describe('llm-test-runner import/export', () => {
|
|
|
98
102
|
},
|
|
99
103
|
},
|
|
100
104
|
],
|
|
105
|
+
chatHistory: { ...emptyChatHistory },
|
|
101
106
|
isRunning: false,
|
|
102
107
|
},
|
|
103
108
|
];
|
|
@@ -121,14 +126,16 @@ describe('llm-test-runner import/export', () => {
|
|
|
121
126
|
id: '1',
|
|
122
127
|
question: 'What is AI?',
|
|
123
128
|
expectedOutcome: buildExpectedOutcome('artificial intelligence'),
|
|
124
|
-
|
|
129
|
+
chatHistory: { ...emptyChatHistory },
|
|
130
|
+
isRunning: false,
|
|
125
131
|
},
|
|
126
132
|
{
|
|
127
133
|
id: '2',
|
|
128
134
|
question: 'What is ML?',
|
|
129
135
|
expectedOutcome: buildExpectedOutcome('machine learning'),
|
|
130
|
-
|
|
131
|
-
|
|
136
|
+
chatHistory: { ...emptyChatHistory },
|
|
137
|
+
isRunning: false,
|
|
138
|
+
},
|
|
132
139
|
];
|
|
133
140
|
const mockFile = createMockFile(JSON.stringify(mockTestData), 'test.json');
|
|
134
141
|
readFileAsync.mockResolvedValue(JSON.stringify(mockTestData));
|
|
@@ -148,20 +155,23 @@ describe('llm-test-runner import/export', () => {
|
|
|
148
155
|
id: '1',
|
|
149
156
|
question: 'Q1',
|
|
150
157
|
expectedOutcome: buildExpectedOutcome('answer1'),
|
|
151
|
-
|
|
158
|
+
chatHistory: { ...emptyChatHistory },
|
|
159
|
+
isRunning: false,
|
|
152
160
|
},
|
|
153
161
|
{
|
|
154
162
|
id: '2',
|
|
155
163
|
question: 'Q2',
|
|
156
164
|
expectedOutcome: buildExpectedOutcome('answer2'),
|
|
157
|
-
|
|
165
|
+
chatHistory: { ...emptyChatHistory },
|
|
166
|
+
isRunning: false,
|
|
158
167
|
},
|
|
159
168
|
{
|
|
160
169
|
id: '3',
|
|
161
170
|
question: 'Q3',
|
|
162
171
|
expectedOutcome: buildExpectedOutcome('answer3'),
|
|
163
|
-
|
|
164
|
-
|
|
172
|
+
chatHistory: { ...emptyChatHistory },
|
|
173
|
+
isRunning: false,
|
|
174
|
+
},
|
|
165
175
|
];
|
|
166
176
|
const mockFile = createMockFile(JSON.stringify(mockTestData), 'test.json');
|
|
167
177
|
readFileAsync.mockResolvedValue(JSON.stringify(mockTestData));
|
package/dist/collection/components/llm-test-runner/llm-test-runner.import-export.test.js.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"llm-test-runner.import-export.test.js","sourceRoot":"","sources":["../../../src/components/llm-test-runner/llm-test-runner.import-export.test.ts"],"names":[],"mappings":"AAAA,qBAAqB;AAErB,IAAI,CAAC,IAAI,CAAC,wCAAwC,EAAE,GAAG,EAAE,CAAC,CAAC;IACzD,mBAAmB,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,kBAAkB,CAAC,GAAG,EAAE,CAAC,CAAC;QACvD,gBAAgB,EAAE,IAAI,CAAC,EAAE,EAAE;KAC5B,CAAC,CAAC;CACJ,CAAC,CAAC,CAAC;AAEJ,IAAI,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAC;AACxC,IAAI,CAAC,IAAI,CAAC,8BAA8B,CAAC,CAAC;AAC1C,IAAI,CAAC,IAAI,CAAC,6CAA6C,CAAC,CAAC;AACzD,IAAI,CAAC,IAAI,CAAC,6CAA6C,CAAC,CAAC;AAEzD,OAAO,EAAE,WAAW,EAAY,MAAM,uBAAuB,CAAC;AAC9D,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAElD,OAAO,EAAE,aAAa,EAAE,MAAM,4BAA4B,CAAC;AAC3D,OAAO,EAAE,YAAY,EAAE,MAAM,8BAA8B,CAAC;AAC5D,OAAO,EAAE,qBAAqB,EAAE,MAAM,6CAA6C,CAAC;AACpF,OAAO,EAAE,eAAe,EAAE,MAAM,6CAA6C,CAAC;AAC9E,OAAO,EAAE,kBAAkB,EAAE,MAAM,gCAAgC,CAAC;AAEpE,QAAQ,CAAC,+BAA+B,EAAE,GAAG,EAAE;IAC7C,IAAI,IAAc,CAAC;IACnB,IAAI,SAAc,CAAC;IACnB,MAAM,oBAAoB,GAAG,CAAC,KAAa,EAAE,EAAE,CAAC;QAC9C;YACE,IAAI,EAAE,UAAmB;YACzB,KAAK,EAAE,kBAAkB;YACzB,KAAK;SACN;KACF,CAAC;IAEF,MAAM,kBAAkB,GAAG,CAAC,YAA+B,EAAE,EAAY,EAAE;QACzE,MAAM,QAAQ,GAAa;YACzB,EAAE,EAAE,GAAG;YACP,QAAQ,EAAE,aAAa;YACvB,eAAe,EAAE,oBAAoB,CAAC,yBAAyB,CAAC;YAChE,SAAS,EAAE,KAAK;SACjB,CAAC;QAEF,OAAO;YACL,GAAG,QAAQ;YACX,GAAG,SAAS;SACb,CAAC;IACJ,CAAC,CAAC;IAEF,MAAM,cAAc,GAAG,CAAC,OAAe,EAAE,QAAgB,EAAE,OAAe,kBAAkB,EAAQ,EAAE;QACpG,OAAO,IAAI,IAAI,CAAC,CAAC,OAAO,CAAC,EAAE,QAAQ,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC;IACjD,CAAC,CAAC;IAEF,UAAU,CAAC,KAAK,IAAI,EAAE;QACpB,IAAI,CAAC,aAAa,EAAE,CAAC;QAErB,IAAI,GAAG,MAAM,WAAW,CAAC;YACvB,UAAU,EAAE,CAAC,aAAa,CAAC;YAC3B,IAAI,EAAE,qCAAqC;SAC5C,CAAC,CAAC;QACH,SAAS,GAAG,IAAI,CAAC,YAAY,CAAC;QAE9B,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;IAC9B,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,mCAAmC,EAAE,GAAG,EAAE;QACjD,EAAE,CAAC,0CAA0C,EAAE,KAAK,IAAI,EAAE;YACxD,MAAM,YAAY,GAAG;gBACnB;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,aAAa;oBACvB,eAAe,EAAE,oBAAoB,CAAC,yBAAyB,CAAC;oBAChE,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YAEF,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,EAAE,iBAAiB,CAAC,CAAC;YAEjF,6BAA6B;YAC5B,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC,CAAC;YAC5E,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,YAAY;aACxB,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC;YAC5C,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC;gBAC3C,EAAE,EAAE,GAAG;gBACP,QAAQ,EAAE,aAAa;gBACvB,eAAe,EAAE,oBAAoB,CAAC,yBAAyB,CAAC;aACjE,CAAC,CAAC;QACL,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,4DAA4D,EAAE,KAAK,IAAI,EAAE;YAC1E,MAAM,aAAa,GAAG;gBACpB;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,eAAe;oBACzB,eAAe,EAAE,oBAAoB,CAAC,MAAM,CAAC;oBAC7C,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YACF,MAAM,oBAAoB,GAAG;gBAC3B;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,eAAe;oBACzB,eAAe,EAAE;wBACf;4BACE,IAAI,EAAE,UAAmB;4BACzB,KAAK,EAAE,kBAAkB;4BACzB,KAAK,EAAE,MAAM;4BACb,oBAAoB,EAAE;gCACpB,QAAQ,EAAE,kBAAkB,CAAC,KAAK;6BACnC;yBACF;qBACF;oBACD,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YAEF,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,aAAa,CAAC,EAAE,WAAW,CAAC,CAAC;YAE3E,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,SAAS,CAAC,aAAa,CAAC,CAAC,CAAC;YAC7E,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,oBAAoB;aAChC,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CACJ,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,oBAAoB,EAAE,QAAQ,CACzE,CAAC,IAAI,CAAC,kBAAkB,CAAC,KAAK,CAAC,CAAC;YACjC,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;YACrE,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,CAC1D,kBAAkB,CACnB,CAAC;QACJ,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,mCAAmC,EAAE,GAAG,EAAE;QACjD,EAAE,CAAC,gDAAgD,EAAE,KAAK,IAAI,EAAE;YAC9D,MAAM,YAAY,GAAG;gBACnB;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,aAAa;oBACvB,eAAe,EAAE,oBAAoB,CAAC,yBAAyB,CAAC;oBAChE,SAAS,EAAE,KAAK;iBACjB;gBACD;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,aAAa;oBACvB,eAAe,EAAE,oBAAoB,CAAC,kBAAkB,CAAC;oBACzD,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YAEF,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,EAAE,WAAW,CAAC,CAAC;YAE1E,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC,CAAC;YAC5E,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,YAAY;aACxB,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC;YAC5C,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC;YAC5D,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC;QAC9D,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,kDAAkD,EAAE,KAAK,IAAI,EAAE;YAChE,MAAM,YAAY,GAAG;gBACnB;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,IAAI;oBACd,eAAe,EAAE,oBAAoB,CAAC,SAAS,CAAC;oBAChD,SAAS,EAAE,KAAK;iBACjB;gBACD;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,IAAI;oBACd,eAAe,EAAE,oBAAoB,CAAC,SAAS,CAAC;oBAChD,SAAS,EAAE,KAAK;iBACjB;gBACD;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,IAAI;oBACd,eAAe,EAAE,oBAAoB,CAAC,SAAS,CAAC;oBAChD,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YAEF,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,EAAE,WAAW,CAAC,CAAC;YAE1E,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC,CAAC;YAC5E,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,YAAY;aACxB,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,GAAG,GAAG,SAAS,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,EAAY,EAAE,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC;YAC7D,MAAM,CAAC,IAAI,GAAG,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;QAC7C,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,8BAA8B,EAAE,GAAG,EAAE;QAC5C,EAAE,CAAC,kCAAkC,EAAE,KAAK,IAAI,EAAE;YAChD,MAAM,QAAQ,GAAG,cAAc,CAAC,cAAc,EAAE,UAAU,EAAE,YAAY,CAAC,CAAC;YAE1E,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,+CAA+C,CAAC,CAAC;QAChF,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,2CAA2C,EAAE,KAAK,IAAI,EAAE;YACzD,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,EAAE,WAAW,CAAC,CAAC;YAElD,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,CAAC;YACpD,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,KAAK;gBACd,KAAK,EAAE,gDAAgD;aACxD,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,CAAC,wBAAwB,CAAC,CAAC;QAC9D,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,gCAAgC,EAAE,KAAK,IAAI,EAAE;YAC9C,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,EAAE,WAAW,CAAC,CAAC;YAElD,aAA2B,CAAC,iBAAiB,CAAC,IAAI,KAAK,CAAC,qBAAqB,CAAC,CAAC,CAAC;YAEjF,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,UAAU,EAAE,CAAC;QACvC,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,mBAAmB,EAAE,GAAG,EAAE;QACjC,EAAE,CAAC,iDAAiD,EAAE,KAAK,IAAI,EAAE;YAC/D,SAAS,CAAC,SAAS,GAAG,CAAC,kBAAkB,EAAE,CAAC,CAAC;YAE7C,MAAM,eAAe,GAAG,IAAI,CAAC,SAAS,CAAC,CAAC,kBAAkB,EAAE,CAAC,CAAC,CAAC;YAC9D,qBAAmC,CAAC,eAAe,CAAC,eAAe,CAAC,CAAC;YACrE,YAA0B,CAAC,kBAAkB,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YAEzD,MAAM,SAAS,CAAC,qBAAqB,EAAE,CAAC;YACxC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,qBAAqB,CAAC,CAAC,oBAAoB,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC;YACxE,MAAM,CAAC,YAAY,CAAC,CAAC,oBAAoB,CACvC,eAAe,EACf,iBAAiB,EACjB,kBAAkB,CACnB,CAAC;QACJ,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,+BAA+B,EAAE,KAAK,IAAI,EAAE;YAC7C,SAAS,CAAC,SAAS,GAAG,CAAC,kBAAkB,EAAE,CAAC,CAAC;YAE5C,qBAAmC,CAAC,eAAe,CAAC,IAAI,CAAC,CAAC;YAC1D,YAA0B,CAAC,kBAAkB,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YAEzD,MAAM,CAAC,SAAS,CAAC,oBAAoB,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YAEnD,MAAM,aAAa,GAAG,SAAS,CAAC,qBAAqB,EAAE,CAAC;YAExD,gBAAgB;YAChB,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,yBAAyB;YACzB,MAAM,aAAa,CAAC;YACpB,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,oBAAoB,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACrD,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,mCAAmC,EAAE,KAAK,IAAI,EAAE;YACjD,MAAM,SAAS,GAAG;gBAChB,kBAAkB,CAAC,EAAE,EAAE,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC;gBAC/C,kBAAkB,CAAC,EAAE,EAAE,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC;gBAC/C,kBAAkB,CAAC,EAAE,EAAE,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC;aAChD,CAAC;YAEF,SAAS,CAAC,SAAS,GAAG,SAAS,CAAC;YAE/B,qBAAmC,CAAC,eAAe,CAAC,IAAI,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC;YAC/E,YAA0B,CAAC,kBAAkB,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YAEzD,MAAM,SAAS,CAAC,qBAAqB,EAAE,CAAC;YACxC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,qBAAqB,CAAC,CAAC,oBAAoB,CAAC,SAAS,CAAC,CAAC;QAChE,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,4BAA4B,EAAE,GAAG,EAAE;QAC1C,EAAE,CAAC,kCAAkC,EAAE,GAAG,EAAE;YAC1C,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,WAAW,EAAE,CAAC;YAC1C,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACxD,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,6BAA6B,EAAE,GAAG,EAAE;YACrC,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,WAAW,EAAE,CAAC;YACtC,MAAM,CAAC,OAAO,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QAChD,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,qBAAqB,EAAE,KAAK,IAAI,EAAE;YACnC,SAAS,CAAC,KAAK,GAAG,YAAY,CAAC;YAC/B,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;YAE3C,SAAS,CAAC,KAAK,GAAG,EAAE,CAAC;YACrB,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACnC,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;AACL,CAAC,CAAC,CAAC","sourcesContent":["/* eslint-env jest */\n\njest.mock('../../lib/evaluation/evaluation-engine', () => ({\n LLMEvaluationEngine: jest.fn().mockImplementation(() => ({\n evaluateResponse: jest.fn()\n }))\n}));\n\njest.mock('../../lib/file/file-reader');\njest.mock('../../lib/file/file-download');\njest.mock('../../lib/import-export/test-suite-exporter');\njest.mock('../../lib/import-export/test-suite-importer');\n\nimport { newSpecPage, SpecPage } from '@stencil/core/testing';\nimport { LLMTestRunner } from './llm-test-runner';\nimport { TestCase } from '../../types/llm-test-runner';\nimport { readFileAsync } from '../../lib/file/file-reader';\nimport { downloadFile } from '../../lib/file/file-download';\nimport { formatTestSuiteAsJson } from '../../lib/import-export/test-suite-exporter';\nimport { importTestSuite } from '../../lib/import-export/test-suite-importer';\nimport { EvaluationApproach } from '../../lib/evaluation/constants';\n\ndescribe('llm-test-runner import/export', () => {\n let page: SpecPage;\n let component: any;\n const buildExpectedOutcome = (value: string) => [\n {\n type: 'textarea' as const,\n label: 'Expected Outcome',\n value,\n },\n ];\n\n const createMockTestCase = (overrides: Partial<TestCase> = {}): TestCase => {\n const defaults: TestCase = {\n id: '1',\n question: 'What is AI?',\n expectedOutcome: buildExpectedOutcome('artificial intelligence'),\n isRunning: false\n };\n\n return {\n ...defaults,\n ...overrides,\n };\n };\n\n const createMockFile = (content: string, filename: string, type: string = 'application/json'): File => {\n return new File([content], filename, { type });\n };\n\n beforeEach(async () => {\n jest.clearAllMocks();\n \n page = await newSpecPage({\n components: [LLMTestRunner],\n html: '<llm-test-runner></llm-test-runner>',\n });\n component = page.rootInstance;\n \n await page.waitForChanges();\n });\n\n describe('File Import - Basic Functionality', () => {\n it('should import valid test cases structure', async () => {\n const mockTestData = [\n {\n id: '1',\n question: 'What is AI?',\n expectedOutcome: buildExpectedOutcome('artificial intelligence'),\n isRunning: false\n }\n ];\n\n const mockFile = createMockFile(JSON.stringify(mockTestData), 'test-suite.json');\n\n // Mock the utility functions\n (readFileAsync as jest.Mock).mockResolvedValue(JSON.stringify(mockTestData));\n (importTestSuite as jest.Mock).mockReturnValue({\n success: true,\n testCases: mockTestData\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.testCases).toHaveLength(1);\n expect(component.testCases[0]).toMatchObject({\n id: '1',\n question: 'What is AI?',\n expectedOutcome: buildExpectedOutcome('artificial intelligence'),\n });\n });\n\n it('should use default evaluation parameters when not provided', async () => {\n const rawImportData = [\n {\n id: '1',\n question: 'Test question',\n expectedOutcome: buildExpectedOutcome('test'),\n isRunning: false\n }\n ];\n const normalizedImportData = [\n {\n id: '1',\n question: 'Test question',\n expectedOutcome: [\n {\n type: 'textarea' as const,\n label: 'Expected Outcome',\n value: 'test',\n evaluationParameters: {\n approach: EvaluationApproach.EXACT,\n },\n },\n ],\n isRunning: false,\n },\n ];\n\n const mockFile = createMockFile(JSON.stringify(rawImportData), 'test.json');\n\n (readFileAsync as jest.Mock).mockResolvedValue(JSON.stringify(rawImportData));\n (importTestSuite as jest.Mock).mockReturnValue({\n success: true,\n testCases: normalizedImportData,\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(\n component.testCases[0].expectedOutcome[0].evaluationParameters?.approach,\n ).toBe(EvaluationApproach.EXACT);\n expect(component.testCases[0].expectedOutcome[0].value).toBe('test');\n expect(component.testCases[0].expectedOutcome[0].label).toBe(\n 'Expected Outcome',\n );\n });\n });\n\n describe('File Import - Multiple Test Cases', () => {\n it('should import multiple test cases successfully', async () => {\n const mockTestData = [\n {\n id: '1',\n question: 'What is AI?',\n expectedOutcome: buildExpectedOutcome('artificial intelligence'),\n isRunning: false\n },\n {\n id: '2',\n question: 'What is ML?',\n expectedOutcome: buildExpectedOutcome('machine learning'),\n isRunning: false\n }\n ];\n\n const mockFile = createMockFile(JSON.stringify(mockTestData), 'test.json');\n\n (readFileAsync as jest.Mock).mockResolvedValue(JSON.stringify(mockTestData));\n (importTestSuite as jest.Mock).mockReturnValue({\n success: true,\n testCases: mockTestData\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.testCases).toHaveLength(2);\n expect(component.testCases[0].question).toBe('What is AI?');\n expect(component.testCases[1].question).toBe('What is ML?');\n });\n\n it('should handle unique IDs for imported test cases', async () => {\n const mockTestData = [\n {\n id: '1',\n question: 'Q1',\n expectedOutcome: buildExpectedOutcome('answer1'),\n isRunning: false\n },\n {\n id: '2',\n question: 'Q2',\n expectedOutcome: buildExpectedOutcome('answer2'),\n isRunning: false\n },\n {\n id: '3',\n question: 'Q3',\n expectedOutcome: buildExpectedOutcome('answer3'),\n isRunning: false\n }\n ];\n\n const mockFile = createMockFile(JSON.stringify(mockTestData), 'test.json');\n\n (readFileAsync as jest.Mock).mockResolvedValue(JSON.stringify(mockTestData));\n (importTestSuite as jest.Mock).mockReturnValue({\n success: true,\n testCases: mockTestData\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n const ids = component.testCases.map((tc: TestCase) => tc.id);\n expect(new Set(ids).size).toBe(ids.length);\n });\n });\n\n describe('File Import - Error Handling', () => {\n it('should reject non-JSON file type', async () => {\n const mockFile = createMockFile('some content', 'test.txt', 'text/plain');\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.error).toBe('Invalid file type. Please select a JSON file.');\n });\n\n it('should handle import failure from utility', async () => {\n const mockFile = createMockFile('{}', 'test.json');\n\n (readFileAsync as jest.Mock).mockResolvedValue('{}');\n (importTestSuite as jest.Mock).mockReturnValue({\n success: false,\n error: 'Invalid JSON structure. Expected a JSON array.'\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.error).toContain('Invalid JSON structure');\n });\n\n it('should handle file read errors', async () => {\n const mockFile = createMockFile('{}', 'test.json');\n\n (readFileAsync as jest.Mock).mockRejectedValue(new Error('Failed to read file'));\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.error).toBeTruthy();\n });\n });\n\n describe('Export Test Suite', () => {\n it('should export test suite with correct structure', async () => {\n component.testCases = [createMockTestCase()];\n\n const mockJsonContent = JSON.stringify([createMockTestCase()]);\n (formatTestSuiteAsJson as jest.Mock).mockReturnValue(mockJsonContent);\n (downloadFile as jest.Mock).mockImplementation(() => {});\n\n await component.handleExportTestSuite();\n await page.waitForChanges();\n\n expect(formatTestSuiteAsJson).toHaveBeenCalledWith(component.testCases);\n expect(downloadFile).toHaveBeenCalledWith(\n mockJsonContent,\n 'test-suite.json',\n 'application/json'\n );\n });\n\n it('should handle exporting state', async () => {\n component.testCases = [createMockTestCase()];\n\n (formatTestSuiteAsJson as jest.Mock).mockReturnValue('[]');\n (downloadFile as jest.Mock).mockImplementation(() => {});\n\n expect(component.isExportingTestSuite).toBe(false);\n\n const exportPromise = component.handleExportTestSuite();\n \n // During export\n await page.waitForChanges();\n\n // After export completes\n await exportPromise;\n await page.waitForChanges();\n\n expect(component.isExportingTestSuite).toBe(false);\n });\n\n it('should export multiple test cases', async () => {\n const testCases = [\n createMockTestCase({ id: '1', question: 'Q1' }),\n createMockTestCase({ id: '2', question: 'Q2' }),\n createMockTestCase({ id: '3', question: 'Q3' })\n ];\n\n component.testCases = testCases;\n\n (formatTestSuiteAsJson as jest.Mock).mockReturnValue(JSON.stringify(testCases));\n (downloadFile as jest.Mock).mockImplementation(() => {});\n\n await component.handleExportTestSuite();\n await page.waitForChanges();\n\n expect(formatTestSuiteAsJson).toHaveBeenCalledWith(testCases);\n });\n });\n\n describe('Component State Management', () => {\n it('should maintain test cases state', () => {\n expect(component.testCases).toBeDefined();\n expect(Array.isArray(component.testCases)).toBe(true);\n });\n\n it('should maintain error state', () => {\n expect(component.error).toBeDefined();\n expect(typeof component.error).toBe('string');\n });\n\n it('should clear errors', async () => {\n component.error = 'Test error';\n await page.waitForChanges();\n\n expect(component.error).toBe('Test error');\n\n component.error = '';\n await page.waitForChanges();\n\n expect(component.error).toBe('');\n });\n });\n});"]}
|
|
1
|
+
{"version":3,"file":"llm-test-runner.import-export.test.js","sourceRoot":"","sources":["../../../src/components/llm-test-runner/llm-test-runner.import-export.test.ts"],"names":[],"mappings":"AAAA,qBAAqB;AAErB,IAAI,CAAC,IAAI,CAAC,wCAAwC,EAAE,GAAG,EAAE,CAAC,CAAC;IACzD,mBAAmB,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,kBAAkB,CAAC,GAAG,EAAE,CAAC,CAAC;QACvD,gBAAgB,EAAE,IAAI,CAAC,EAAE,EAAE;KAC5B,CAAC,CAAC;CACJ,CAAC,CAAC,CAAC;AAEJ,IAAI,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAC;AACxC,IAAI,CAAC,IAAI,CAAC,8BAA8B,CAAC,CAAC;AAC1C,IAAI,CAAC,IAAI,CAAC,6CAA6C,CAAC,CAAC;AACzD,IAAI,CAAC,IAAI,CAAC,6CAA6C,CAAC,CAAC;AAEzD,OAAO,EAAE,WAAW,EAAY,MAAM,uBAAuB,CAAC;AAC9D,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAElD,OAAO,EAAE,aAAa,EAAE,MAAM,4BAA4B,CAAC;AAC3D,OAAO,EAAE,YAAY,EAAE,MAAM,8BAA8B,CAAC;AAC5D,OAAO,EAAE,qBAAqB,EAAE,MAAM,6CAA6C,CAAC;AACpF,OAAO,EAAE,eAAe,EAAE,MAAM,6CAA6C,CAAC;AAC9E,OAAO,EAAE,kBAAkB,EAAE,MAAM,gCAAgC,CAAC;AAEpE,MAAM,gBAAgB,GAAG,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,EAAE,EAAW,CAAC;AAEhE,QAAQ,CAAC,+BAA+B,EAAE,GAAG,EAAE;IAC7C,IAAI,IAAc,CAAC;IACnB,IAAI,SAAc,CAAC;IACnB,MAAM,oBAAoB,GAAG,CAAC,KAAa,EAAE,EAAE,CAAC;QAC9C;YACE,IAAI,EAAE,UAAmB;YACzB,KAAK,EAAE,kBAAkB;YACzB,KAAK;SACN;KACF,CAAC;IAEF,MAAM,kBAAkB,GAAG,CAAC,YAA+B,EAAE,EAAY,EAAE;QACzE,MAAM,QAAQ,GAAa;YACzB,EAAE,EAAE,GAAG;YACP,QAAQ,EAAE,aAAa;YACvB,eAAe,EAAE,oBAAoB,CAAC,yBAAyB,CAAC;YAChE,WAAW,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,EAAE,EAAE;YAC1C,SAAS,EAAE,KAAK;SACjB,CAAC;QAEF,OAAO;YACL,GAAG,QAAQ;YACX,GAAG,SAAS;SACb,CAAC;IACJ,CAAC,CAAC;IAEF,MAAM,cAAc,GAAG,CAAC,OAAe,EAAE,QAAgB,EAAE,OAAe,kBAAkB,EAAQ,EAAE;QACpG,OAAO,IAAI,IAAI,CAAC,CAAC,OAAO,CAAC,EAAE,QAAQ,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC;IACjD,CAAC,CAAC;IAEF,UAAU,CAAC,KAAK,IAAI,EAAE;QACpB,IAAI,CAAC,aAAa,EAAE,CAAC;QAErB,IAAI,GAAG,MAAM,WAAW,CAAC;YACvB,UAAU,EAAE,CAAC,aAAa,CAAC;YAC3B,IAAI,EAAE,qCAAqC;SAC5C,CAAC,CAAC;QACH,SAAS,GAAG,IAAI,CAAC,YAAY,CAAC;QAE9B,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;IAC9B,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,mCAAmC,EAAE,GAAG,EAAE;QACjD,EAAE,CAAC,0CAA0C,EAAE,KAAK,IAAI,EAAE;YACxD,MAAM,YAAY,GAAG;gBACnB;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,aAAa;oBACvB,eAAe,EAAE,oBAAoB,CAAC,yBAAyB,CAAC;oBAChE,WAAW,EAAE,EAAE,GAAG,gBAAgB,EAAE;oBACpC,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YAEF,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,EAAE,iBAAiB,CAAC,CAAC;YAEjF,6BAA6B;YAC5B,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC,CAAC;YAC5E,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,YAAY;aACxB,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC;YAC5C,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC;gBAC3C,EAAE,EAAE,GAAG;gBACP,QAAQ,EAAE,aAAa;gBACvB,eAAe,EAAE,oBAAoB,CAAC,yBAAyB,CAAC;aACjE,CAAC,CAAC;QACL,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,4DAA4D,EAAE,KAAK,IAAI,EAAE;YAC1E,MAAM,aAAa,GAAG;gBACpB;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,eAAe;oBACzB,eAAe,EAAE,oBAAoB,CAAC,MAAM,CAAC;oBAC7C,WAAW,EAAE,EAAE,GAAG,gBAAgB,EAAE;oBACpC,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YACF,MAAM,oBAAoB,GAAG;gBAC3B;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,eAAe;oBACzB,eAAe,EAAE;wBACf;4BACE,IAAI,EAAE,UAAmB;4BACzB,KAAK,EAAE,kBAAkB;4BACzB,KAAK,EAAE,MAAM;4BACb,oBAAoB,EAAE;gCACpB,QAAQ,EAAE,kBAAkB,CAAC,KAAK;6BACnC;yBACF;qBACF;oBACD,WAAW,EAAE,EAAE,GAAG,gBAAgB,EAAE;oBACpC,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YAEF,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,aAAa,CAAC,EAAE,WAAW,CAAC,CAAC;YAE3E,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,SAAS,CAAC,aAAa,CAAC,CAAC,CAAC;YAC7E,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,oBAAoB;aAChC,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CACJ,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,oBAAoB,EAAE,QAAQ,CACzE,CAAC,IAAI,CAAC,kBAAkB,CAAC,KAAK,CAAC,CAAC;YACjC,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;YACrE,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,CAC1D,kBAAkB,CACnB,CAAC;QACJ,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,mCAAmC,EAAE,GAAG,EAAE;QACjD,EAAE,CAAC,gDAAgD,EAAE,KAAK,IAAI,EAAE;YAC9D,MAAM,YAAY,GAAG;gBACnB;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,aAAa;oBACvB,eAAe,EAAE,oBAAoB,CAAC,yBAAyB,CAAC;oBAChE,WAAW,EAAE,EAAE,GAAG,gBAAgB,EAAE;oBACpC,SAAS,EAAE,KAAK;iBACjB;gBACD;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,aAAa;oBACvB,eAAe,EAAE,oBAAoB,CAAC,kBAAkB,CAAC;oBACzD,WAAW,EAAE,EAAE,GAAG,gBAAgB,EAAE;oBACpC,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YAEF,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,EAAE,WAAW,CAAC,CAAC;YAE1E,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC,CAAC;YAC5E,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,YAAY;aACxB,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC;YAC5C,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC;YAC5D,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC;QAC9D,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,kDAAkD,EAAE,KAAK,IAAI,EAAE;YAChE,MAAM,YAAY,GAAG;gBACnB;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,IAAI;oBACd,eAAe,EAAE,oBAAoB,CAAC,SAAS,CAAC;oBAChD,WAAW,EAAE,EAAE,GAAG,gBAAgB,EAAE;oBACpC,SAAS,EAAE,KAAK;iBACjB;gBACD;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,IAAI;oBACd,eAAe,EAAE,oBAAoB,CAAC,SAAS,CAAC;oBAChD,WAAW,EAAE,EAAE,GAAG,gBAAgB,EAAE;oBACpC,SAAS,EAAE,KAAK;iBACjB;gBACD;oBACE,EAAE,EAAE,GAAG;oBACP,QAAQ,EAAE,IAAI;oBACd,eAAe,EAAE,oBAAoB,CAAC,SAAS,CAAC;oBAChD,WAAW,EAAE,EAAE,GAAG,gBAAgB,EAAE;oBACpC,SAAS,EAAE,KAAK;iBACjB;aACF,CAAC;YAEF,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,EAAE,WAAW,CAAC,CAAC;YAE1E,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC,CAAC;YAC5E,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,YAAY;aACxB,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,GAAG,GAAG,SAAS,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,EAAY,EAAE,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC;YAC7D,MAAM,CAAC,IAAI,GAAG,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;QAC7C,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,8BAA8B,EAAE,GAAG,EAAE;QAC5C,EAAE,CAAC,kCAAkC,EAAE,KAAK,IAAI,EAAE;YAChD,MAAM,QAAQ,GAAG,cAAc,CAAC,cAAc,EAAE,UAAU,EAAE,YAAY,CAAC,CAAC;YAE1E,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,+CAA+C,CAAC,CAAC;QAChF,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,2CAA2C,EAAE,KAAK,IAAI,EAAE;YACzD,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,EAAE,WAAW,CAAC,CAAC;YAElD,aAA2B,CAAC,iBAAiB,CAAC,IAAI,CAAC,CAAC;YACpD,eAA6B,CAAC,eAAe,CAAC;gBAC7C,OAAO,EAAE,KAAK;gBACd,KAAK,EAAE,gDAAgD;aACxD,CAAC,CAAC;YAEH,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,CAAC,wBAAwB,CAAC,CAAC;QAC9D,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,gCAAgC,EAAE,KAAK,IAAI,EAAE;YAC9C,MAAM,QAAQ,GAAG,cAAc,CAAC,IAAI,EAAE,WAAW,CAAC,CAAC;YAElD,aAA2B,CAAC,iBAAiB,CAAC,IAAI,KAAK,CAAC,qBAAqB,CAAC,CAAC,CAAC;YAEjF,MAAM,SAAS,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACvC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,UAAU,EAAE,CAAC;QACvC,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,mBAAmB,EAAE,GAAG,EAAE;QACjC,EAAE,CAAC,iDAAiD,EAAE,KAAK,IAAI,EAAE;YAC/D,SAAS,CAAC,SAAS,GAAG,CAAC,kBAAkB,EAAE,CAAC,CAAC;YAE7C,MAAM,eAAe,GAAG,IAAI,CAAC,SAAS,CAAC,CAAC,kBAAkB,EAAE,CAAC,CAAC,CAAC;YAC9D,qBAAmC,CAAC,eAAe,CAAC,eAAe,CAAC,CAAC;YACrE,YAA0B,CAAC,kBAAkB,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YAEzD,MAAM,SAAS,CAAC,qBAAqB,EAAE,CAAC;YACxC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,qBAAqB,CAAC,CAAC,oBAAoB,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC;YACxE,MAAM,CAAC,YAAY,CAAC,CAAC,oBAAoB,CACvC,eAAe,EACf,iBAAiB,EACjB,kBAAkB,CACnB,CAAC;QACJ,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,+BAA+B,EAAE,KAAK,IAAI,EAAE;YAC7C,SAAS,CAAC,SAAS,GAAG,CAAC,kBAAkB,EAAE,CAAC,CAAC;YAE5C,qBAAmC,CAAC,eAAe,CAAC,IAAI,CAAC,CAAC;YAC1D,YAA0B,CAAC,kBAAkB,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YAEzD,MAAM,CAAC,SAAS,CAAC,oBAAoB,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YAEnD,MAAM,aAAa,GAAG,SAAS,CAAC,qBAAqB,EAAE,CAAC;YAExD,gBAAgB;YAChB,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,yBAAyB;YACzB,MAAM,aAAa,CAAC;YACpB,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,oBAAoB,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACrD,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,mCAAmC,EAAE,KAAK,IAAI,EAAE;YACjD,MAAM,SAAS,GAAG;gBAChB,kBAAkB,CAAC,EAAE,EAAE,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC;gBAC/C,kBAAkB,CAAC,EAAE,EAAE,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC;gBAC/C,kBAAkB,CAAC,EAAE,EAAE,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC;aAChD,CAAC;YAEF,SAAS,CAAC,SAAS,GAAG,SAAS,CAAC;YAE/B,qBAAmC,CAAC,eAAe,CAAC,IAAI,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC;YAC/E,YAA0B,CAAC,kBAAkB,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YAEzD,MAAM,SAAS,CAAC,qBAAqB,EAAE,CAAC;YACxC,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,qBAAqB,CAAC,CAAC,oBAAoB,CAAC,SAAS,CAAC,CAAC;QAChE,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;IAEH,QAAQ,CAAC,4BAA4B,EAAE,GAAG,EAAE;QAC1C,EAAE,CAAC,kCAAkC,EAAE,GAAG,EAAE;YAC1C,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,WAAW,EAAE,CAAC;YAC1C,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACxD,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,6BAA6B,EAAE,GAAG,EAAE;YACrC,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,WAAW,EAAE,CAAC;YACtC,MAAM,CAAC,OAAO,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QAChD,CAAC,CAAC,CAAC;QAEH,EAAE,CAAC,qBAAqB,EAAE,KAAK,IAAI,EAAE;YACnC,SAAS,CAAC,KAAK,GAAG,YAAY,CAAC;YAC/B,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;YAE3C,SAAS,CAAC,KAAK,GAAG,EAAE,CAAC;YACrB,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;YAE5B,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACnC,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;AACL,CAAC,CAAC,CAAC","sourcesContent":["/* eslint-env jest */\n\njest.mock('../../lib/evaluation/evaluation-engine', () => ({\n LLMEvaluationEngine: jest.fn().mockImplementation(() => ({\n evaluateResponse: jest.fn()\n }))\n}));\n\njest.mock('../../lib/file/file-reader');\njest.mock('../../lib/file/file-download');\njest.mock('../../lib/import-export/test-suite-exporter');\njest.mock('../../lib/import-export/test-suite-importer');\n\nimport { newSpecPage, SpecPage } from '@stencil/core/testing';\nimport { LLMTestRunner } from './llm-test-runner';\nimport { TestCase } from '../../types/llm-test-runner';\nimport { readFileAsync } from '../../lib/file/file-reader';\nimport { downloadFile } from '../../lib/file/file-download';\nimport { formatTestSuiteAsJson } from '../../lib/import-export/test-suite-exporter';\nimport { importTestSuite } from '../../lib/import-export/test-suite-importer';\nimport { EvaluationApproach } from '../../lib/evaluation/constants';\n\nconst emptyChatHistory = { enabled: false, value: '' } as const;\n\ndescribe('llm-test-runner import/export', () => {\n let page: SpecPage;\n let component: any;\n const buildExpectedOutcome = (value: string) => [\n {\n type: 'textarea' as const,\n label: 'Expected Outcome',\n value,\n },\n ];\n\n const createMockTestCase = (overrides: Partial<TestCase> = {}): TestCase => {\n const defaults: TestCase = {\n id: '1',\n question: 'What is AI?',\n expectedOutcome: buildExpectedOutcome('artificial intelligence'),\n chatHistory: { enabled: false, value: '' },\n isRunning: false,\n };\n\n return {\n ...defaults,\n ...overrides,\n };\n };\n\n const createMockFile = (content: string, filename: string, type: string = 'application/json'): File => {\n return new File([content], filename, { type });\n };\n\n beforeEach(async () => {\n jest.clearAllMocks();\n \n page = await newSpecPage({\n components: [LLMTestRunner],\n html: '<llm-test-runner></llm-test-runner>',\n });\n component = page.rootInstance;\n \n await page.waitForChanges();\n });\n\n describe('File Import - Basic Functionality', () => {\n it('should import valid test cases structure', async () => {\n const mockTestData = [\n {\n id: '1',\n question: 'What is AI?',\n expectedOutcome: buildExpectedOutcome('artificial intelligence'),\n chatHistory: { ...emptyChatHistory },\n isRunning: false,\n },\n ];\n\n const mockFile = createMockFile(JSON.stringify(mockTestData), 'test-suite.json');\n\n // Mock the utility functions\n (readFileAsync as jest.Mock).mockResolvedValue(JSON.stringify(mockTestData));\n (importTestSuite as jest.Mock).mockReturnValue({\n success: true,\n testCases: mockTestData\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.testCases).toHaveLength(1);\n expect(component.testCases[0]).toMatchObject({\n id: '1',\n question: 'What is AI?',\n expectedOutcome: buildExpectedOutcome('artificial intelligence'),\n });\n });\n\n it('should use default evaluation parameters when not provided', async () => {\n const rawImportData = [\n {\n id: '1',\n question: 'Test question',\n expectedOutcome: buildExpectedOutcome('test'),\n chatHistory: { ...emptyChatHistory },\n isRunning: false,\n },\n ];\n const normalizedImportData = [\n {\n id: '1',\n question: 'Test question',\n expectedOutcome: [\n {\n type: 'textarea' as const,\n label: 'Expected Outcome',\n value: 'test',\n evaluationParameters: {\n approach: EvaluationApproach.EXACT,\n },\n },\n ],\n chatHistory: { ...emptyChatHistory },\n isRunning: false,\n },\n ];\n\n const mockFile = createMockFile(JSON.stringify(rawImportData), 'test.json');\n\n (readFileAsync as jest.Mock).mockResolvedValue(JSON.stringify(rawImportData));\n (importTestSuite as jest.Mock).mockReturnValue({\n success: true,\n testCases: normalizedImportData,\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(\n component.testCases[0].expectedOutcome[0].evaluationParameters?.approach,\n ).toBe(EvaluationApproach.EXACT);\n expect(component.testCases[0].expectedOutcome[0].value).toBe('test');\n expect(component.testCases[0].expectedOutcome[0].label).toBe(\n 'Expected Outcome',\n );\n });\n });\n\n describe('File Import - Multiple Test Cases', () => {\n it('should import multiple test cases successfully', async () => {\n const mockTestData = [\n {\n id: '1',\n question: 'What is AI?',\n expectedOutcome: buildExpectedOutcome('artificial intelligence'),\n chatHistory: { ...emptyChatHistory },\n isRunning: false,\n },\n {\n id: '2',\n question: 'What is ML?',\n expectedOutcome: buildExpectedOutcome('machine learning'),\n chatHistory: { ...emptyChatHistory },\n isRunning: false,\n },\n ];\n\n const mockFile = createMockFile(JSON.stringify(mockTestData), 'test.json');\n\n (readFileAsync as jest.Mock).mockResolvedValue(JSON.stringify(mockTestData));\n (importTestSuite as jest.Mock).mockReturnValue({\n success: true,\n testCases: mockTestData\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.testCases).toHaveLength(2);\n expect(component.testCases[0].question).toBe('What is AI?');\n expect(component.testCases[1].question).toBe('What is ML?');\n });\n\n it('should handle unique IDs for imported test cases', async () => {\n const mockTestData = [\n {\n id: '1',\n question: 'Q1',\n expectedOutcome: buildExpectedOutcome('answer1'),\n chatHistory: { ...emptyChatHistory },\n isRunning: false,\n },\n {\n id: '2',\n question: 'Q2',\n expectedOutcome: buildExpectedOutcome('answer2'),\n chatHistory: { ...emptyChatHistory },\n isRunning: false,\n },\n {\n id: '3',\n question: 'Q3',\n expectedOutcome: buildExpectedOutcome('answer3'),\n chatHistory: { ...emptyChatHistory },\n isRunning: false,\n },\n ];\n\n const mockFile = createMockFile(JSON.stringify(mockTestData), 'test.json');\n\n (readFileAsync as jest.Mock).mockResolvedValue(JSON.stringify(mockTestData));\n (importTestSuite as jest.Mock).mockReturnValue({\n success: true,\n testCases: mockTestData\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n const ids = component.testCases.map((tc: TestCase) => tc.id);\n expect(new Set(ids).size).toBe(ids.length);\n });\n });\n\n describe('File Import - Error Handling', () => {\n it('should reject non-JSON file type', async () => {\n const mockFile = createMockFile('some content', 'test.txt', 'text/plain');\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.error).toBe('Invalid file type. Please select a JSON file.');\n });\n\n it('should handle import failure from utility', async () => {\n const mockFile = createMockFile('{}', 'test.json');\n\n (readFileAsync as jest.Mock).mockResolvedValue('{}');\n (importTestSuite as jest.Mock).mockReturnValue({\n success: false,\n error: 'Invalid JSON structure. Expected a JSON array.'\n });\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.error).toContain('Invalid JSON structure');\n });\n\n it('should handle file read errors', async () => {\n const mockFile = createMockFile('{}', 'test.json');\n\n (readFileAsync as jest.Mock).mockRejectedValue(new Error('Failed to read file'));\n\n await component.handleImport(mockFile);\n await page.waitForChanges();\n\n expect(component.error).toBeTruthy();\n });\n });\n\n describe('Export Test Suite', () => {\n it('should export test suite with correct structure', async () => {\n component.testCases = [createMockTestCase()];\n\n const mockJsonContent = JSON.stringify([createMockTestCase()]);\n (formatTestSuiteAsJson as jest.Mock).mockReturnValue(mockJsonContent);\n (downloadFile as jest.Mock).mockImplementation(() => {});\n\n await component.handleExportTestSuite();\n await page.waitForChanges();\n\n expect(formatTestSuiteAsJson).toHaveBeenCalledWith(component.testCases);\n expect(downloadFile).toHaveBeenCalledWith(\n mockJsonContent,\n 'test-suite.json',\n 'application/json'\n );\n });\n\n it('should handle exporting state', async () => {\n component.testCases = [createMockTestCase()];\n\n (formatTestSuiteAsJson as jest.Mock).mockReturnValue('[]');\n (downloadFile as jest.Mock).mockImplementation(() => {});\n\n expect(component.isExportingTestSuite).toBe(false);\n\n const exportPromise = component.handleExportTestSuite();\n \n // During export\n await page.waitForChanges();\n\n // After export completes\n await exportPromise;\n await page.waitForChanges();\n\n expect(component.isExportingTestSuite).toBe(false);\n });\n\n it('should export multiple test cases', async () => {\n const testCases = [\n createMockTestCase({ id: '1', question: 'Q1' }),\n createMockTestCase({ id: '2', question: 'Q2' }),\n createMockTestCase({ id: '3', question: 'Q3' })\n ];\n\n component.testCases = testCases;\n\n (formatTestSuiteAsJson as jest.Mock).mockReturnValue(JSON.stringify(testCases));\n (downloadFile as jest.Mock).mockImplementation(() => {});\n\n await component.handleExportTestSuite();\n await page.waitForChanges();\n\n expect(formatTestSuiteAsJson).toHaveBeenCalledWith(testCases);\n });\n });\n\n describe('Component State Management', () => {\n it('should maintain test cases state', () => {\n expect(component.testCases).toBeDefined();\n expect(Array.isArray(component.testCases)).toBe(true);\n });\n\n it('should maintain error state', () => {\n expect(component.error).toBeDefined();\n expect(typeof component.error).toBe('string');\n });\n\n it('should clear errors', async () => {\n component.error = 'Test error';\n await page.waitForChanges();\n\n expect(component.error).toBe('Test error');\n\n component.error = '';\n await page.waitForChanges();\n\n expect(component.error).toBe('');\n });\n });\n});"]}
|
|
@@ -11,7 +11,7 @@ import { resolveDynamicExpectedOutcomes, } from "../../lib/test-cases/dynamic-ex
|
|
|
11
11
|
import * as TestCaseMutations from "../../lib/test-cases/test-case-mutations";
|
|
12
12
|
import { EvaluationService } from "../../lib/evaluation/evaluation-service";
|
|
13
13
|
import { validateTestCaseInputArray } from "../../schemas/test-case";
|
|
14
|
-
import { validateExpectedOutcomeSchema } from "../../schemas/expected-outcome";
|
|
14
|
+
import { getExtractorIds, validateExpectedOutcomeArrayWithExtractors, validateExpectedOutcomeSchema, } from "../../schemas/expected-outcome";
|
|
15
15
|
import { LLMTestRunnerHeader } from "./header/llm-test-runner-header";
|
|
16
16
|
import { LLMTestCases } from "./test-cases/llm-test-cases";
|
|
17
17
|
export class LLMTestRunner {
|
|
@@ -21,6 +21,7 @@ export class LLMTestRunner {
|
|
|
21
21
|
useSave = false;
|
|
22
22
|
usePromptEditor = false;
|
|
23
23
|
resolveExpectedOutcome;
|
|
24
|
+
evaluationSourceExtractors;
|
|
24
25
|
initialTestCases;
|
|
25
26
|
defaultExpectedOutcomeSchema;
|
|
26
27
|
testCases = [
|
|
@@ -34,6 +35,7 @@ export class LLMTestRunner {
|
|
|
34
35
|
value: '',
|
|
35
36
|
},
|
|
36
37
|
],
|
|
38
|
+
chatHistory: { enabled: false, value: '' },
|
|
37
39
|
isRunning: false,
|
|
38
40
|
},
|
|
39
41
|
];
|
|
@@ -56,6 +58,12 @@ export class LLMTestRunner {
|
|
|
56
58
|
// Initialize testCases from prop if provided
|
|
57
59
|
if (this.initialTestCases !== undefined) {
|
|
58
60
|
validateTestCaseInputArray(this.initialTestCases);
|
|
61
|
+
const extractorIds = getExtractorIds(this.evaluationSourceExtractors);
|
|
62
|
+
if (extractorIds.length > 0) {
|
|
63
|
+
this.initialTestCases.forEach(testCase => {
|
|
64
|
+
validateExpectedOutcomeArrayWithExtractors(testCase.expectedOutcome, extractorIds);
|
|
65
|
+
});
|
|
66
|
+
}
|
|
59
67
|
this.testCases = this.initialTestCases.map((rawTestCase, index) => {
|
|
60
68
|
try {
|
|
61
69
|
return createTestCaseFromInput(rawTestCase);
|
|
@@ -79,8 +87,6 @@ export class LLMTestRunner {
|
|
|
79
87
|
this.testCases = [];
|
|
80
88
|
}
|
|
81
89
|
}
|
|
82
|
-
componentDidLoad() { }
|
|
83
|
-
disconnectedCallback() { }
|
|
84
90
|
async resetSavingState() {
|
|
85
91
|
this.isSaving = false;
|
|
86
92
|
}
|
|
@@ -91,6 +97,12 @@ export class LLMTestRunner {
|
|
|
91
97
|
const { testCaseId, key, value } = event.detail;
|
|
92
98
|
this.testCases = this.testCases.map(tc => tc.id === testCaseId ? { ...tc, [key]: value } : tc);
|
|
93
99
|
};
|
|
100
|
+
handleChatHistoryChange = (event) => {
|
|
101
|
+
const { testCaseId, enabled, value } = event.detail;
|
|
102
|
+
this.updateTestCase(testCaseId, {
|
|
103
|
+
chatHistory: { enabled, value },
|
|
104
|
+
});
|
|
105
|
+
};
|
|
94
106
|
addNewTestCase() {
|
|
95
107
|
try {
|
|
96
108
|
const schema = this.getResolvedExpectedOutcomeSchema();
|
|
@@ -107,13 +119,17 @@ export class LLMTestRunner {
|
|
|
107
119
|
updateTestCase(id, updates) {
|
|
108
120
|
this.testCases = this.testCases.map(tc => tc.id === id ? { ...tc, ...updates } : tc);
|
|
109
121
|
}
|
|
110
|
-
|
|
122
|
+
requestLlmResponse(testCase) {
|
|
111
123
|
return new Promise((resolve, reject) => {
|
|
112
|
-
|
|
124
|
+
const payload = {
|
|
113
125
|
prompt: testCase.question,
|
|
114
126
|
resolve,
|
|
115
127
|
reject,
|
|
116
|
-
}
|
|
128
|
+
};
|
|
129
|
+
if (testCase.chatHistory?.enabled) {
|
|
130
|
+
payload.chatHistory = testCase.chatHistory.value;
|
|
131
|
+
}
|
|
132
|
+
this.llmRequest.emit(payload);
|
|
117
133
|
});
|
|
118
134
|
}
|
|
119
135
|
throwError(reason) {
|
|
@@ -126,14 +142,14 @@ export class LLMTestRunner {
|
|
|
126
142
|
const startTime = Date.now();
|
|
127
143
|
this.updateTestCase(testCase.id, { isRunning: true });
|
|
128
144
|
const [llmSettled, resolutionSettled] = await Promise.allSettled([
|
|
129
|
-
this.
|
|
145
|
+
this.requestLlmResponse(testCase),
|
|
130
146
|
resolveDynamicExpectedOutcomes(testCase, this.resolveExpectedOutcome),
|
|
131
147
|
]);
|
|
132
148
|
const responseTime = Date.now() - startTime;
|
|
133
149
|
if (llmSettled.status === 'rejected') {
|
|
134
150
|
this.updateTestCase(testCase.id, {
|
|
135
151
|
isRunning: false,
|
|
136
|
-
output:
|
|
152
|
+
output: undefined,
|
|
137
153
|
error: this.addErrorMessage(llmSettled.reason, 'Unknown error'),
|
|
138
154
|
responseTime,
|
|
139
155
|
});
|
|
@@ -181,7 +197,7 @@ export class LLMTestRunner {
|
|
|
181
197
|
this.updateTestCase(testCase.id, {
|
|
182
198
|
evaluationResult: result,
|
|
183
199
|
});
|
|
184
|
-
});
|
|
200
|
+
}, this.evaluationSourceExtractors);
|
|
185
201
|
}
|
|
186
202
|
async runAllTests() {
|
|
187
203
|
this.isRunningAll = true;
|
|
@@ -212,7 +228,7 @@ export class LLMTestRunner {
|
|
|
212
228
|
this.error = '';
|
|
213
229
|
try {
|
|
214
230
|
const content = await readFileAsync(file);
|
|
215
|
-
const result = importTestSuite(content);
|
|
231
|
+
const result = importTestSuite(content, getExtractorIds(this.evaluationSourceExtractors));
|
|
216
232
|
if (!result.success) {
|
|
217
233
|
this.error = result.error || 'Unknown error occurred during import.';
|
|
218
234
|
return;
|
|
@@ -273,7 +289,7 @@ export class LLMTestRunner {
|
|
|
273
289
|
}
|
|
274
290
|
}
|
|
275
291
|
render() {
|
|
276
|
-
return (h("div", { key: '
|
|
292
|
+
return (h("div", { key: '7433beaa1d60d48f65600c43e11b302b892a7bca', class: "test-runner-container" }, h(LLMTestRunnerHeader, { key: '8083cc39376e7a710bd3f52efb184b959e885a87', isExportingTestSuite: this.isExportingTestSuite, isExportingTestResults: this.isExportingTestResults, isRunningAll: this.isRunningAll, useSave: this.useSave, isSaving: this.isSaving, usePromptEditor: this.usePromptEditor, onImport: file => this.handleImport(file), onExportSuite: () => this.handleExportTestSuite(), onExportResults: () => this.handleExportTestResults(), onRunAll: () => this.runAllTests(), onSave: () => this.handleSave() }), h(ErrorMessage, { key: 'ddced98c13cd595c4cfb6eef11b27cb173769518', message: this.error, onClear: () => (this.error = '') }), h("div", { key: '8d6f65c4d68d34869b644709eacb97fec93683c6', class: "test-runner-container__content" }, h(LLMTestCases, { key: '5ccb186132b23af6209209b0a14086e03cf790af', testCases: this.testCases, dynamicResolutionSupported: !!this.resolveExpectedOutcome, extractorIds: getExtractorIds(this.evaluationSourceExtractors), onRun: testCase => this.runSingleTest(testCase).catch(() => { }), onDelete: id => this.deleteTestCase(id), onAddTestCase: () => this.addNewTestCase(), handleTestCaseChange: this.handleTestCaseChange, onExpectedOutcomeChange: this.handleExpectedOutcomeChange, onChatHistoryChange: this.handleChatHistoryChange }))));
|
|
277
293
|
}
|
|
278
294
|
static get is() { return "llm-test-runner"; }
|
|
279
295
|
static get encapsulation() { return "shadow"; }
|
|
@@ -354,7 +370,7 @@ export class LLMTestRunner {
|
|
|
354
370
|
"mutable": false,
|
|
355
371
|
"complexType": {
|
|
356
372
|
"original": "ExpectedOutcomeResolver",
|
|
357
|
-
"resolved": "(resolutionQuery: string, context: { testCase: { id: string; question: string; expectedOutcome: ({ label: string; type: \"text\"; value: string; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"textarea\"; value: string; placeholder?: string; rows?: number; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; outcomeMode?: \"static\" | \"dynamic\"; resolutionQuery?: string; } | { label: string; type: \"chips-input\"; value: string[]; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"select\"; options: string[]; value: string; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; })[]; output?: string; isRunning?: boolean; error?: string; evaluationResult?: EvaluationResult; responseTime?: number; }; fieldIndex: number; }) => Promise<string>",
|
|
373
|
+
"resolved": "(resolutionQuery: string, context: { testCase: { id: string; question: string; expectedOutcome: ({ label: string; type: \"text\"; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"textarea\"; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; rows?: number; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; outcomeMode?: \"static\" | \"dynamic\"; resolutionQuery?: string; } | { label: string; type: \"chips-input\"; value: string[]; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"select\"; options: string[]; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; })[]; chatHistory: { enabled: boolean; value: string; }; output?: { text?: string; metadata?: Record<string, unknown>; }; isRunning?: boolean; error?: string; evaluationResult?: EvaluationResult; responseTime?: number; }; fieldIndex: number; }) => Promise<string>",
|
|
358
374
|
"references": {
|
|
359
375
|
"ExpectedOutcomeResolver": {
|
|
360
376
|
"location": "import",
|
|
@@ -373,12 +389,36 @@ export class LLMTestRunner {
|
|
|
373
389
|
"getter": false,
|
|
374
390
|
"setter": false
|
|
375
391
|
},
|
|
392
|
+
"evaluationSourceExtractors": {
|
|
393
|
+
"type": "unknown",
|
|
394
|
+
"mutable": false,
|
|
395
|
+
"complexType": {
|
|
396
|
+
"original": "EvaluationSourceExtractors",
|
|
397
|
+
"resolved": "((payload: { text?: string; metadata?: Record<string, unknown>; }) => string | Promise<string>) | string",
|
|
398
|
+
"references": {
|
|
399
|
+
"EvaluationSourceExtractors": {
|
|
400
|
+
"location": "import",
|
|
401
|
+
"path": "../../types/llm-test-runner",
|
|
402
|
+
"id": "src/types/llm-test-runner.ts::EvaluationSourceExtractors",
|
|
403
|
+
"referenceLocation": "EvaluationSourceExtractors"
|
|
404
|
+
}
|
|
405
|
+
}
|
|
406
|
+
},
|
|
407
|
+
"required": false,
|
|
408
|
+
"optional": true,
|
|
409
|
+
"docs": {
|
|
410
|
+
"tags": [],
|
|
411
|
+
"text": ""
|
|
412
|
+
},
|
|
413
|
+
"getter": false,
|
|
414
|
+
"setter": false
|
|
415
|
+
},
|
|
376
416
|
"initialTestCases": {
|
|
377
417
|
"type": "unknown",
|
|
378
418
|
"mutable": false,
|
|
379
419
|
"complexType": {
|
|
380
420
|
"original": "TestCase[]",
|
|
381
|
-
"resolved": "{ id: string; question: string; expectedOutcome: ({ label: string; type: \"text\"; value: string; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"textarea\"; value: string; placeholder?: string; rows?: number; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; outcomeMode?: \"static\" | \"dynamic\"; resolutionQuery?: string; } | { label: string; type: \"chips-input\"; value: string[]; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"select\"; options: string[]; value: string; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; })[]; output?: string; isRunning?: boolean; error?: string; evaluationResult?: EvaluationResult; responseTime?: number; }[]",
|
|
421
|
+
"resolved": "{ id: string; question: string; expectedOutcome: ({ label: string; type: \"text\"; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"textarea\"; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; rows?: number; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; outcomeMode?: \"static\" | \"dynamic\"; resolutionQuery?: string; } | { label: string; type: \"chips-input\"; value: string[]; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"select\"; options: string[]; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; })[]; chatHistory: { enabled: boolean; value: string; }; output?: { text?: string; metadata?: Record<string, unknown>; }; isRunning?: boolean; error?: string; evaluationResult?: EvaluationResult; responseTime?: number; }[]",
|
|
382
422
|
"references": {
|
|
383
423
|
"TestCase": {
|
|
384
424
|
"location": "import",
|
|
@@ -402,7 +442,7 @@ export class LLMTestRunner {
|
|
|
402
442
|
"mutable": false,
|
|
403
443
|
"complexType": {
|
|
404
444
|
"original": "ExpectedOutcomeSchema",
|
|
405
|
-
"resolved": "({ label: string; type: \"text\"; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"textarea\"; placeholder?: string; rows?: number; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"chips-input\"; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"select\"; options: string[]; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; })[]",
|
|
445
|
+
"resolved": "({ label: string; type: \"text\"; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"textarea\"; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; rows?: number; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"chips-input\"; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"select\"; options: string[]; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; })[]",
|
|
406
446
|
"references": {
|
|
407
447
|
"ExpectedOutcomeSchema": {
|
|
408
448
|
"location": "import",
|
|
@@ -508,6 +548,10 @@ export class LLMTestRunner {
|
|
|
508
548
|
"location": "global",
|
|
509
549
|
"id": "global::Promise"
|
|
510
550
|
},
|
|
551
|
+
"Record": {
|
|
552
|
+
"location": "global",
|
|
553
|
+
"id": "global::Record"
|
|
554
|
+
},
|
|
511
555
|
"EvaluationResult": {
|
|
512
556
|
"location": "import",
|
|
513
557
|
"path": "../../lib/evaluation/types",
|
|
@@ -521,7 +565,7 @@ export class LLMTestRunner {
|
|
|
521
565
|
"referenceLocation": "TestCase"
|
|
522
566
|
}
|
|
523
567
|
},
|
|
524
|
-
"return": "Promise<{ id: string; question: string; expectedOutcome: ({ label: string; type: \"text\"; value: string; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"textarea\"; value: string; placeholder?: string; rows?: number; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; outcomeMode?: \"static\" | \"dynamic\"; resolutionQuery?: string; } | { label: string; type: \"chips-input\"; value: string[]; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"select\"; options: string[]; value: string; placeholder?: string; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; })[]; output?: string; isRunning?: boolean; error?: string; evaluationResult?: EvaluationResult; responseTime?: number; }[]>"
|
|
568
|
+
"return": "Promise<{ id: string; question: string; expectedOutcome: ({ label: string; type: \"text\"; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"textarea\"; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; rows?: number; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; outcomeMode?: \"static\" | \"dynamic\"; resolutionQuery?: string; } | { label: string; type: \"chips-input\"; value: string[]; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; } | { label: string; type: \"select\"; options: string[]; value: string; placeholder?: string; evaluationSource?: { type: \"text\"; } | { type: \"custom\"; extractorId: string; }; evaluationParameters?: { approach: EvaluationApproach; threshold?: number; }; })[]; chatHistory: { enabled: boolean; value: string; }; output?: { text?: string; metadata?: Record<string, unknown>; }; isRunning?: boolean; error?: string; evaluationResult?: EvaluationResult; responseTime?: number; }[]>"
|
|
525
569
|
},
|
|
526
570
|
"docs": {
|
|
527
571
|
"text": "",
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"llm-test-runner.js","sourceRoot":"","sources":["../../../src/components/llm-test-runner/llm-test-runner.tsx"],"names":[],"mappings":"AAAA,OAAO,EACL,SAAS,EACT,KAAK,EACL,IAAI,EACJ,CAAC,EAED,KAAK,EACL,MAAM,GACP,MAAM,eAAe,CAAC;AAEvB,OAAO,EAAE,YAAY,EAAE,MAAM,gCAAgC,CAAC;AAC9D,OAAO,EAAE,kBAAkB,EAAE,MAAM,qDAAqD,CAAC;AAOzF,OAAO,EAAE,aAAa,EAAE,MAAM,4BAA4B,CAAC;AAC3D,OAAO,EAAE,YAAY,EAAE,MAAM,8BAA8B,CAAC;AAC5D,OAAO,EAAE,qBAAqB,EAAE,MAAM,6CAA6C,CAAC;AACpF,OAAO,EAAE,sBAAsB,EAAE,MAAM,0CAA0C,CAAC;AAClF,OAAO,EAAE,eAAe,EAAE,MAAM,6CAA6C,CAAC;AAC9E,OAAO,EACL,cAAc,EACd,uBAAuB,EACvB,+BAA+B,GAChC,MAAM,wCAAwC,CAAC;AAChD,OAAO,EAEL,8BAA8B,GAC/B,MAAM,wDAAwD,CAAC;AAChE,OAAO,KAAK,iBAAiB,MAAM,0CAA0C,CAAC;AAC9E,OAAO,EAAE,iBAAiB,EAAE,MAAM,yCAAyC,CAAC;AAC5E,OAAO,EAAE,0BAA0B,EAAE,MAAM,yBAAyB,CAAC;AACrE,OAAO,EAAE,6BAA6B,EAAE,MAAM,gCAAgC,CAAC;AAC/E,OAAO,EAAE,mBAAmB,EAAE,MAAM,iCAAiC,CAAC;AACtE,OAAO,EAAE,YAAY,EAAE,MAAM,6BAA6B,CAAC;AAoB3D,MAAM,OAAO,aAAa;IACf,UAAU,CAAkC;IAC5C,IAAI,CAA4B;IACjC,OAAO,GAAY,GAAG,CAAC;IACvB,OAAO,GAAa,KAAK,CAAC;IAC1B,eAAe,GAAa,KAAK,CAAC;IAClC,sBAAsB,CAA2B;IACjD,gBAAgB,CAAc;IAC9B,4BAA4B,CAAyB;IACpD,SAAS,GAAe;QAC/B;YACE,EAAE,EAAE,GAAG;YACP,QAAQ,EAAE,EAAE;YACZ,eAAe,EAAE;gBACf;oBACE,IAAI,EAAE,UAAU;oBAChB,KAAK,EAAE,kBAAkB;oBACzB,KAAK,EAAE,EAAE;iBACV;aACF;YACD,SAAS,EAAE,KAAK;SACjB;KACF,CAAC;IACO,YAAY,GAAY,KAAK,CAAC;IAC9B,KAAK,GAAW,EAAE,CAAC;IACnB,oBAAoB,GAAY,KAAK,CAAC;IACtC,sBAAsB,GAAY,KAAK,CAAC;IACxC,QAAQ,GAAY,KAAK,CAAC;IAE3B,iBAAiB,CAAoB;IAErC,gCAAgC;QACtC,IAAI,IAAI,CAAC,4BAA4B,KAAK,SAAS,EAAE,CAAC;YACpD,OAAO,+BAA+B,CAAC;QACzC,CAAC;QAED,6BAA6B,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAC;QACjE,OAAO,IAAI,CAAC,4BAA4B,CAAC;IAC3C,CAAC;IAED,iBAAiB;QACf,IAAI,CAAC,iBAAiB,GAAG,IAAI,iBAAiB,EAAE,CAAC;QACjD,IAAI,CAAC;YACH,6CAA6C;YAC7C,IAAI,IAAI,CAAC,gBAAgB,KAAK,SAAS,EAAE,CAAC;gBACxC,0BAA0B,CAAC,IAAI,CAAC,gBAAgB,CAAC,CAAC;gBAClD,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,KAAK,EAAE,EAAE;oBAChE,IAAI,CAAC;wBACH,OAAO,uBAAuB,CAAC,WAAW,CAAC,CAAC;oBAC9C,CAAC;oBAAC,OAAO,GAAG,EAAE,CAAC;wBACb,MAAM,OAAO,GAAG,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,eAAe,CAAC;wBACrE,MAAM,IAAI,KAAK,CAAC,sCAAsC,KAAK,KAAK,OAAO,EAAE,CAAC,CAAC;oBAC7E,CAAC;gBACH,CAAC,CAAC,CAAC;YACL,CAAC;iBAAM,CAAC;gBACN,MAAM,MAAM,GAAG,IAAI,CAAC,gCAAgC,EAAE,CAAC;gBACvD,IAAI,CAAC,SAAS,GAAG,CAAC,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC;YAC5C,CAAC;QACH,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,IAAI,CAAC,KAAK;gBACR,GAAG,YAAY,KAAK;oBAClB,CAAC,CAAC,GAAG,CAAC,OAAO;oBACb,CAAC,CAAC,gDAAgD,CAAC;YACvD,IAAI,CAAC,SAAS,GAAG,EAAE,CAAC;QACtB,CAAC;IACH,CAAC;IAED,gBAAgB,KAAI,CAAC;IAErB,oBAAoB,KAAI,CAAC;IAGzB,KAAK,CAAC,gBAAgB;QACpB,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC;IACxB,CAAC;IAGD,KAAK,CAAC,YAAY;QAChB,OAAO,IAAI,CAAC,SAAS,CAAC;IACxB,CAAC;IAEO,oBAAoB,GAAG,CAC7B,KAAsE,EACtE,EAAE;QACF,MAAM,EAAE,UAAU,EAAE,GAAG,EAAE,KAAK,EAAE,GAAG,KAAK,CAAC,MAAM,CAAC;QAChD,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,CACvC,EAAE,CAAC,EAAE,KAAK,UAAU,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,CACpD,CAAC;IACJ,CAAC,CAAC;IAEM,cAAc;QACpB,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,IAAI,CAAC,gCAAgC,EAAE,CAAC;YACvD,MAAM,WAAW,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC;YAC3C,IAAI,CAAC,SAAS,GAAG,CAAC,GAAG,IAAI,CAAC,SAAS,EAAE,WAAW,CAAC,CAAC;QACpD,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,IAAI,CAAC,KAAK;gBACR,GAAG,YAAY,KAAK;oBAClB,CAAC,CAAC,GAAG,CAAC,OAAO;oBACb,CAAC,CAAC,gDAAgD,CAAC;QACzD,CAAC;IACH,CAAC;IAEO,cAAc,CAAC,EAAU,EAAE,OAA0B;QAC3D,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,CACvC,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,EAAE,GAAG,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAC1C,CAAC;IACJ,CAAC;IAEO,cAAc,CAAC,QAAkB;QACvC,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;YACrC,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC;gBACnB,MAAM,EAAE,QAAQ,CAAC,QAAQ;gBACzB,OAAO;gBACP,MAAM;aACP,CAAC,CAAC;QACL,CAAC,CAAC,CAAC;IACL,CAAC;IAEO,UAAU,CAAC,MAAe;QAChC,MAAM,MAAM,YAAY,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;IACrE,CAAC;IAEO,eAAe,CAAC,MAAe,EAAE,QAAgB;QACvD,OAAO,MAAM,YAAY,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC;IAC7D,CAAC;IAEO,KAAK,CAAC,aAAa,CAAC,QAAkB;QAC5C,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC7B,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QACtD,MAAM,CAAC,UAAU,EAAE,iBAAiB,CAAC,GAAG,MAAM,OAAO,CAAC,UAAU,CAAC;YAC/D,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC;YAC7B,8BAA8B,CAAC,QAAQ,EAAE,IAAI,CAAC,sBAAsB,CAAC;SACtE,CAAC,CAAC;QAEH,MAAM,YAAY,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;QAE5C,IAAI,UAAU,CAAC,MAAM,KAAK,UAAU,EAAE,CAAC;YACrC,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAC/B,SAAS,EAAE,KAAK;gBAChB,MAAM,EAAE,IAAI;gBACZ,KAAK,EAAE,IAAI,CAAC,eAAe,CAAC,UAAU,CAAC,MAAM,EAAE,eAAe,CAAC;gBAC/D,YAAY;aACb,CAAC,CAAC;YACH,IAAI,CAAC,UAAU,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;QACrC,CAAC;QACD,MAAM,UAAU,GAAG,UAAU,CAAC,KAAK,CAAC;QAEpC,IAAI,iBAAiB,CAAC,MAAM,KAAK,UAAU,EAAE,CAAC;YAC5C,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAC/B,SAAS,EAAE,KAAK;gBAChB,MAAM,EAAE,UAAU;gBAClB,KAAK,EAAE,IAAI,CAAC,eAAe,CACzB,iBAAiB,CAAC,MAAM,EACxB,6CAA6C,CAC9C;gBACD,YAAY;aACb,CAAC,CAAC;YACH,IAAI,CAAC,UAAU,CAAC,iBAAiB,CAAC,MAAM,CAAC,CAAC;QAC5C,CAAC;QACD,MAAM,gBAAgB,GAAG,iBAAiB,CAAC,KAAK,CAAC;QAEjD,MAAM,qBAAqB,GAAa;YACtC,GAAG,gBAAgB;YACnB,MAAM,EAAE,UAAU;YAClB,YAAY;SACb,CAAC;QAEF,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE;YAC/B,SAAS,EAAE,KAAK;YAChB,MAAM,EAAE,UAAU;YAClB,KAAK,EAAE,IAAI;YACX,YAAY;YACZ,eAAe,EAAE,qBAAqB,CAAC,eAAe;SACvD,CAAC,CAAC;QAEH,MAAM,IAAI,CAAC,gBAAgB,CAAC,qBAAqB,CAAC,CAAC;IACrD,CAAC;IAEO,cAAc,CAAC,EAAU;QAC/B,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC;IAC7D,CAAC;IAEO,2BAA2B,GAAG,CACpC,KAA+C,EAC/C,EAAE;QACF,MAAM,EAAE,UAAU,EAAE,GAAG,MAAM,EAAE,GAAG,KAAK,CAAC,MAAM,CAAC;QAE/C,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;YACvC,IAAI,EAAE,CAAC,EAAE,KAAK,UAAU,EAAE,CAAC;gBACzB,OAAO,EAAE,CAAC;YACZ,CAAC;YAED,OAAO,iBAAiB,CAAC,0BAA0B,CAAC,EAAE,EAAE,MAAM,CAAC,CAAC;QAClE,CAAC,CAAC,CAAC;IACL,CAAC,CAAC;IAEM,KAAK,CAAC,gBAAgB,CAAC,QAAkB;QAC/C,MAAM,IAAI,CAAC,iBAAiB,CAAC,gBAAgB,CAC3C,QAAQ,EACR,CAAC,MAAwB,EAAE,EAAE;YAC3B,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAC/B,gBAAgB,EAAE,MAAM;aACzB,CAAC,CAAC;QACL,CAAC,CACF,CAAC;IACJ,CAAC;IAEO,KAAK,CAAC,WAAW;QACvB,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC;QACzB,MAAM,KAAK,GAAG,EAAE,CAAC;QACjB,KAAK,MAAM,QAAQ,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YACtC,IAAI,CAAC,QAAQ,CAAC,SAAS,IAAI,QAAQ,CAAC,QAAQ,CAAC,IAAI,EAAE,EAAE,CAAC;gBACpD,KAAK,CAAC,IAAI,CAAC,GAAG,EAAE,CACd,IAAI,CAAC,aAAa,CAAC,QAAQ,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE;oBACvC,OAAO,CAAC,KAAK,CAAC,gBAAgB,QAAQ,CAAC,EAAE,SAAS,EAAE,GAAG,CAAC,CAAC;gBAC3D,CAAC,CAAC,CACH,CAAC;YACJ,CAAC;QACH,CAAC;QACD,IAAI,CAAC;YACH,MAAM,OAAO,GAAG,IAAI,kBAAkB,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;YACrD,MAAM,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,OAAO,CAAC,KAAK,CAAC,6BAA6B,EAAE,GAAG,CAAC,CAAC;QACpD,CAAC;QACD,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC;IAC5B,CAAC;IAEO,KAAK,CAAC,YAAY,CAAC,IAAU;QACnC,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,KAAK,kBAAkB,CAAC;QACpD,MAAM,eAAe,GAAG,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;QAElE,IAAI,CAAC,UAAU,IAAI,CAAC,eAAe,EAAE,CAAC;YACpC,IAAI,CAAC,KAAK,GAAG,+CAA+C,CAAC;YAC7D,OAAO;QACT,CAAC;QAED,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC;QAEhB,IAAI,CAAC;YACH,MAAM,OAAO,GAAG,MAAM,aAAa,CAAC,IAAI,CAAC,CAAC;YAC1C,MAAM,MAAM,GAAG,eAAe,CAAC,OAAO,CAAC,CAAC;YAExC,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;gBACpB,IAAI,CAAC,KAAK,GAAG,MAAM,CAAC,KAAK,IAAI,uCAAuC,CAAC;gBACrE,OAAO;YACT,CAAC;YAED,IAAI,CAAC,SAAS,GAAG,MAAM,CAAC,SAAS,IAAI,EAAE,CAAC;QAC1C,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,IAAI,CAAC,KAAK;gBACR,GAAG,YAAY,KAAK;oBAClB,CAAC,CAAC,GAAG,CAAC,OAAO;oBACb,CAAC,CAAC,gEAAgE,CAAC;YACvE,OAAO,CAAC,KAAK,CAAC,wBAAwB,EAAE,GAAG,CAAC,CAAC;QAC/C,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,qBAAqB;QACjC,IAAI,CAAC,oBAAoB,GAAG,IAAI,CAAC;QACjC,IAAI,CAAC;YACH,MAAM,WAAW,GAAG,qBAAqB,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YAE1D,gDAAgD;YAChD,MAAM,IAAI,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC,UAAU,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC;YAEvD,YAAY,CAAC,WAAW,EAAE,iBAAiB,EAAE,kBAAkB,CAAC,CAAC;QACnE,CAAC;gBAAS,CAAC;YACT,IAAI,CAAC,oBAAoB,GAAG,KAAK,CAAC;QACpC,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,uBAAuB;QACnC,IAAI,CAAC,sBAAsB,GAAG,IAAI,CAAC;QACnC,IAAI,CAAC;YACH,MAAM,UAAU,GAAG,sBAAsB,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YAE1D,gDAAgD;YAChD,MAAM,IAAI,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC,UAAU,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC;YAEvD,YAAY,CAAC,UAAU,EAAE,kBAAkB,EAAE,UAAU,CAAC,CAAC;QAC3D,CAAC;gBAAS,CAAC;YACT,IAAI,CAAC,sBAAsB,GAAG,KAAK,CAAC;QACtC,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,UAAU;QACtB,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC;QACrB,IAAI,CAAC;YACH,MAAM,OAAO,GAAG;gBACd,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;gBACnC,SAAS,EAAE,IAAI,CAAC,SAAS;aAC1B,CAAC;YACF,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;YAExB,yEAAyE;YACzE,UAAU,CAAC,GAAG,EAAE;gBACd,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;oBAClB,OAAO,CAAC,IAAI,CAAC,2CAA2C,CAAC,CAAC;oBAC1D,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC;gBACxB,CAAC;YACH,CAAC,EAAE,KAAK,CAAC,CAAC;QACZ,CAAC;gBAAS,CAAC;YACT,iEAAiE;YACjE,mEAAmE;QACrE,CAAC;IACH,CAAC;IAED,MAAM;QACJ,OAAO,CACL,4DAAK,KAAK,EAAC,uBAAuB;YAChC,EAAC,mBAAmB,qDAClB,oBAAoB,EAAE,IAAI,CAAC,oBAAoB,EAC/C,sBAAsB,EAAE,IAAI,CAAC,sBAAsB,EACnD,YAAY,EAAE,IAAI,CAAC,YAAY,EAC/B,OAAO,EAAE,IAAI,CAAC,OAAO,EACrB,QAAQ,EAAE,IAAI,CAAC,QAAQ,EACvB,eAAe,EAAE,IAAI,CAAC,eAAe,EACrC,QAAQ,EAAE,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC,EACzC,aAAa,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,qBAAqB,EAAE,EACjD,eAAe,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,uBAAuB,EAAE,EACrD,QAAQ,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,WAAW,EAAE,EAClC,MAAM,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,UAAU,EAAE,GAC/B;YACF,EAAC,YAAY,qDAAC,OAAO,EAAE,IAAI,CAAC,KAAK,EAAE,OAAO,EAAE,GAAG,EAAE,CAAC,CAAC,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC,GAAI;YACvE,4DAAK,KAAK,EAAC,gCAAgC;gBACzC,EAAC,YAAY,qDACX,SAAS,EAAE,IAAI,CAAC,SAAS,EACzB,0BAA0B,EAAE,CAAC,CAAC,IAAI,CAAC,sBAAsB,EACzD,KAAK,EAAE,QAAQ,CAAC,EAAE,CAAC,IAAI,CAAC,aAAa,CAAC,QAAQ,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,EAC/D,QAAQ,EAAE,EAAE,CAAC,EAAE,CAAC,IAAI,CAAC,cAAc,CAAC,EAAE,CAAC,EACvC,aAAa,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,cAAc,EAAE,EAC1C,oBAAoB,EAAE,IAAI,CAAC,oBAAoB,EAC/C,uBAAuB,EAAE,IAAI,CAAC,2BAA2B,GACzD,CACE,CACF,CACP,CAAC;IACJ,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CACF","sourcesContent":["import {\n Component,\n State,\n Prop,\n h,\n EventEmitter,\n Event,\n Method,\n} from '@stencil/core';\nimport { EvaluationResult } from '../../lib/evaluation/types';\nimport { ErrorMessage } from '../error-message/error-message';\nimport { RateLimitedFetcher } from '../../lib/rate-limited-fetcher/rate-limited-fetcher';\nimport {\n ExpectedOutcomeSchema,\n TestCase,\n LLMRequestPayload,\n SavePayload,\n} from '../../types/llm-test-runner';\nimport { readFileAsync } from '../../lib/file/file-reader';\nimport { downloadFile } from '../../lib/file/file-download';\nimport { formatTestSuiteAsJson } from '../../lib/import-export/test-suite-exporter';\nimport { exportTestResultsToCsv } from '../../lib/import-export/test-results-csv';\nimport { importTestSuite } from '../../lib/import-export/test-suite-importer';\nimport {\n createTestCase,\n createTestCaseFromInput,\n DEFAULT_EXPECTED_OUTCOME_SCHEMA,\n} from '../../lib/test-cases/test-case-factory';\nimport {\n type ExpectedOutcomeResolver,\n resolveDynamicExpectedOutcomes,\n} from '../../lib/test-cases/dynamic-expected-outcome-resolver';\nimport * as TestCaseMutations from '../../lib/test-cases/test-case-mutations';\nimport { EvaluationService } from '../../lib/evaluation/evaluation-service';\nimport { validateTestCaseInputArray } from '../../schemas/test-case';\nimport { validateExpectedOutcomeSchema } from '../../schemas/expected-outcome';\nimport { LLMTestRunnerHeader } from './header/llm-test-runner-header';\nimport { LLMTestCases } from './test-cases/llm-test-cases';\nimport { ExpectedOutcomeChangeDetail } from './test-cases/expected-outcome-renderer';\n\n@Component({\n tag: 'llm-test-runner',\n styleUrls: [\n '../../styles/tokens.css',\n 'llm-test-runner.css',\n 'header/llm-test-runner-header.css',\n 'test-cases/llm-test-cases.css',\n 'test-cases/llm-test-case-row.css',\n 'test-cases/actions/row-actions.css',\n 'test-cases/evaluation/evaluation-summary.css',\n 'test-cases/output/response-output.css',\n '../error-message/error-message.css',\n '../../lib/ui/button/button.css',\n '../../lib/ui/icon-button/icon-button.css',\n ],\n shadow: true,\n})\nexport class LLMTestRunner {\n @Event() llmRequest: EventEmitter<LLMRequestPayload>;\n @Event() save: EventEmitter<SavePayload>;\n @Prop() delayMs?: number = 500;\n @Prop() useSave?: boolean = false;\n @Prop() usePromptEditor?: boolean = false;\n @Prop() resolveExpectedOutcome?: ExpectedOutcomeResolver;\n @Prop() initialTestCases?: TestCase[];\n @Prop() defaultExpectedOutcomeSchema?: ExpectedOutcomeSchema;\n @State() testCases: TestCase[] = [\n {\n id: '1',\n question: '',\n expectedOutcome: [\n {\n type: 'textarea',\n label: 'Expected Outcome',\n value: '',\n },\n ],\n isRunning: false,\n },\n ];\n @State() isRunningAll: boolean = false;\n @State() error: string = '';\n @State() isExportingTestSuite: boolean = false;\n @State() isExportingTestResults: boolean = false;\n @State() isSaving: boolean = false;\n\n private evaluationService: EvaluationService;\n\n private getResolvedExpectedOutcomeSchema(): ExpectedOutcomeSchema {\n if (this.defaultExpectedOutcomeSchema === undefined) {\n return DEFAULT_EXPECTED_OUTCOME_SCHEMA;\n }\n\n validateExpectedOutcomeSchema(this.defaultExpectedOutcomeSchema);\n return this.defaultExpectedOutcomeSchema;\n }\n\n componentWillLoad() {\n this.evaluationService = new EvaluationService();\n try {\n // Initialize testCases from prop if provided\n if (this.initialTestCases !== undefined) {\n validateTestCaseInputArray(this.initialTestCases);\n this.testCases = this.initialTestCases.map((rawTestCase, index) => {\n try {\n return createTestCaseFromInput(rawTestCase);\n } catch (err) {\n const message = err instanceof Error ? err.message : 'Unknown error';\n throw new Error(`Invalid initial test case at index ${index}: ${message}`);\n }\n });\n } else {\n const schema = this.getResolvedExpectedOutcomeSchema();\n this.testCases = [createTestCase(schema)];\n }\n } catch (err) {\n this.error =\n err instanceof Error\n ? err.message\n : 'Invalid defaultExpectedOutcomeSchema provided.';\n this.testCases = [];\n }\n }\n\n componentDidLoad() {}\n\n disconnectedCallback() {}\n\n @Method()\n async resetSavingState(): Promise<void> {\n this.isSaving = false;\n }\n\n @Method()\n async getTestCases(): Promise<TestCase[]> {\n return this.testCases;\n }\n\n private handleTestCaseChange = (\n event: CustomEvent<{ testCaseId: string; key: string; value: string }>,\n ) => {\n const { testCaseId, key, value } = event.detail;\n this.testCases = this.testCases.map(tc =>\n tc.id === testCaseId ? { ...tc, [key]: value } : tc,\n );\n };\n\n private addNewTestCase() {\n try {\n const schema = this.getResolvedExpectedOutcomeSchema();\n const newTestCase = createTestCase(schema);\n this.testCases = [...this.testCases, newTestCase];\n } catch (err) {\n this.error =\n err instanceof Error\n ? err.message\n : 'Invalid defaultExpectedOutcomeSchema provided.';\n }\n }\n\n private updateTestCase(id: string, updates: Partial<TestCase>) {\n this.testCases = this.testCases.map(tc =>\n tc.id === id ? { ...tc, ...updates } : tc,\n );\n }\n\n private requestLlmText(testCase: TestCase): Promise<string> {\n return new Promise((resolve, reject) => {\n this.llmRequest.emit({\n prompt: testCase.question,\n resolve,\n reject,\n });\n });\n }\n\n private throwError(reason: unknown): never {\n throw reason instanceof Error ? reason : new Error(String(reason));\n }\n\n private addErrorMessage(reason: unknown, fallback: string): string {\n return reason instanceof Error ? reason.message : fallback;\n }\n\n private async runSingleTest(testCase: TestCase): Promise<void> {\n const startTime = Date.now();\n this.updateTestCase(testCase.id, { isRunning: true });\n const [llmSettled, resolutionSettled] = await Promise.allSettled([\n this.requestLlmText(testCase),\n resolveDynamicExpectedOutcomes(testCase, this.resolveExpectedOutcome),\n ]);\n\n const responseTime = Date.now() - startTime;\n\n if (llmSettled.status === 'rejected') {\n this.updateTestCase(testCase.id, {\n isRunning: false,\n output: null,\n error: this.addErrorMessage(llmSettled.reason, 'Unknown error'),\n responseTime,\n });\n this.throwError(llmSettled.reason);\n }\n const aiResponse = llmSettled.value;\n\n if (resolutionSettled.status === 'rejected') {\n this.updateTestCase(testCase.id, {\n isRunning: false,\n output: aiResponse,\n error: this.addErrorMessage(\n resolutionSettled.reason,\n 'Failed to resolve dynamic expected outcome.',\n ),\n responseTime,\n });\n this.throwError(resolutionSettled.reason);\n }\n const resolvedTestCase = resolutionSettled.value;\n\n const forEvaluationTestCase: TestCase = {\n ...resolvedTestCase,\n output: aiResponse,\n responseTime,\n };\n\n this.updateTestCase(testCase.id, {\n isRunning: false,\n output: aiResponse,\n error: null,\n responseTime,\n expectedOutcome: forEvaluationTestCase.expectedOutcome,\n });\n\n await this.evaluateResponse(forEvaluationTestCase);\n }\n\n private deleteTestCase(id: string) {\n this.testCases = this.testCases.filter(tc => tc.id !== id);\n }\n\n private handleExpectedOutcomeChange = (\n event: CustomEvent<ExpectedOutcomeChangeDetail>,\n ) => {\n const { testCaseId, ...change } = event.detail;\n\n this.testCases = this.testCases.map(tc => {\n if (tc.id !== testCaseId) {\n return tc;\n }\n\n return TestCaseMutations.applyExpectedOutcomeChange(tc, change);\n });\n };\n\n private async evaluateResponse(testCase: TestCase): Promise<void> {\n await this.evaluationService.evaluateTestCase(\n testCase,\n (result: EvaluationResult) => {\n this.updateTestCase(testCase.id, {\n evaluationResult: result,\n });\n },\n );\n }\n\n private async runAllTests() {\n this.isRunningAll = true;\n const tasks = [];\n for (const testCase of this.testCases) {\n if (!testCase.isRunning && testCase.question.trim()) {\n tasks.push(() =>\n this.runSingleTest(testCase).catch(err => {\n console.error(`⚠️ Test case ${testCase.id} failed`, err);\n }),\n );\n }\n }\n try {\n const fetcher = new RateLimitedFetcher(this.delayMs);\n await fetcher.runAll(tasks);\n } catch (err) {\n console.error('⚠️ Error running all tests:', err);\n }\n this.isRunningAll = false;\n }\n\n private async handleImport(file: File): Promise<void> {\n const isJsonType = file.type === 'application/json';\n const isJsonExtension = file.name.toLowerCase().endsWith('.json');\n\n if (!isJsonType && !isJsonExtension) {\n this.error = 'Invalid file type. Please select a JSON file.';\n return;\n }\n\n this.error = '';\n\n try {\n const content = await readFileAsync(file);\n const result = importTestSuite(content);\n\n if (!result.success) {\n this.error = result.error || 'Unknown error occurred during import.';\n return;\n }\n\n this.testCases = result.testCases || [];\n } catch (err) {\n this.error =\n err instanceof Error\n ? err.message\n : 'Error processing file. Please ensure it is a valid JSON array.';\n console.error('File Processing Error:', err);\n }\n }\n\n private async handleExportTestSuite() {\n this.isExportingTestSuite = true;\n try {\n const jsonContent = formatTestSuiteAsJson(this.testCases);\n\n // Added a small delay to show the loading state\n await new Promise(resolve => setTimeout(resolve, 500));\n\n downloadFile(jsonContent, 'test-suite.json', 'application/json');\n } finally {\n this.isExportingTestSuite = false;\n }\n }\n\n private async handleExportTestResults() {\n this.isExportingTestResults = true;\n try {\n const csvContent = exportTestResultsToCsv(this.testCases);\n\n // Added a small delay to show the loading state\n await new Promise(resolve => setTimeout(resolve, 500));\n\n downloadFile(csvContent, 'test-results.csv', 'text/csv');\n } finally {\n this.isExportingTestResults = false;\n }\n }\n\n private async handleSave() {\n this.isSaving = true;\n try {\n const testRun = {\n timestamp: new Date().toISOString(),\n testCases: this.testCases,\n };\n this.save.emit(testRun);\n\n // Failsafe: Auto-reset saving state after 10 seconds to prevent stuck UI\n setTimeout(() => {\n if (this.isSaving) {\n console.warn('Save operation timed out, resetting state');\n this.isSaving = false;\n }\n }, 10000);\n } finally {\n // Parent will call resetSavingState() when actual save completes\n // If not called within 10 seconds, failsafe above will reset state\n }\n }\n\n render() {\n return (\n <div class=\"test-runner-container\">\n <LLMTestRunnerHeader\n isExportingTestSuite={this.isExportingTestSuite}\n isExportingTestResults={this.isExportingTestResults}\n isRunningAll={this.isRunningAll}\n useSave={this.useSave}\n isSaving={this.isSaving}\n usePromptEditor={this.usePromptEditor}\n onImport={file => this.handleImport(file)}\n onExportSuite={() => this.handleExportTestSuite()}\n onExportResults={() => this.handleExportTestResults()}\n onRunAll={() => this.runAllTests()}\n onSave={() => this.handleSave()}\n />\n <ErrorMessage message={this.error} onClear={() => (this.error = '')} />\n <div class=\"test-runner-container__content\">\n <LLMTestCases\n testCases={this.testCases}\n dynamicResolutionSupported={!!this.resolveExpectedOutcome}\n onRun={testCase => this.runSingleTest(testCase).catch(() => {})}\n onDelete={id => this.deleteTestCase(id)}\n onAddTestCase={() => this.addNewTestCase()}\n handleTestCaseChange={this.handleTestCaseChange}\n onExpectedOutcomeChange={this.handleExpectedOutcomeChange}\n />\n </div>\n </div>\n );\n }\n}\n"]}
|
|
1
|
+
{"version":3,"file":"llm-test-runner.js","sourceRoot":"","sources":["../../../src/components/llm-test-runner/llm-test-runner.tsx"],"names":[],"mappings":"AAAA,OAAO,EACL,SAAS,EACT,KAAK,EACL,IAAI,EACJ,CAAC,EAED,KAAK,EACL,MAAM,GACP,MAAM,eAAe,CAAC;AAEvB,OAAO,EAAE,YAAY,EAAE,MAAM,gCAAgC,CAAC;AAC9D,OAAO,EAAE,kBAAkB,EAAE,MAAM,qDAAqD,CAAC;AASzF,OAAO,EAAE,aAAa,EAAE,MAAM,4BAA4B,CAAC;AAC3D,OAAO,EAAE,YAAY,EAAE,MAAM,8BAA8B,CAAC;AAC5D,OAAO,EAAE,qBAAqB,EAAE,MAAM,6CAA6C,CAAC;AACpF,OAAO,EAAE,sBAAsB,EAAE,MAAM,0CAA0C,CAAC;AAClF,OAAO,EAAE,eAAe,EAAE,MAAM,6CAA6C,CAAC;AAC9E,OAAO,EACL,cAAc,EACd,uBAAuB,EACvB,+BAA+B,GAChC,MAAM,wCAAwC,CAAC;AAChD,OAAO,EAEL,8BAA8B,GAC/B,MAAM,wDAAwD,CAAC;AAChE,OAAO,KAAK,iBAAiB,MAAM,0CAA0C,CAAC;AAC9E,OAAO,EAAE,iBAAiB,EAAE,MAAM,yCAAyC,CAAC;AAC5E,OAAO,EAAE,0BAA0B,EAAE,MAAM,yBAAyB,CAAC;AACrE,OAAO,EACL,eAAe,EACf,0CAA0C,EAC1C,6BAA6B,GAC9B,MAAM,gCAAgC,CAAC;AACxC,OAAO,EAAE,mBAAmB,EAAE,MAAM,iCAAiC,CAAC;AACtE,OAAO,EAAE,YAAY,EAAE,MAAM,6BAA6B,CAAC;AAqB3D,MAAM,OAAO,aAAa;IACf,UAAU,CAAkC;IAC5C,IAAI,CAA4B;IACjC,OAAO,GAAY,GAAG,CAAC;IACvB,OAAO,GAAa,KAAK,CAAC;IAC1B,eAAe,GAAa,KAAK,CAAC;IAClC,sBAAsB,CAA2B;IACjD,0BAA0B,CAA8B;IACxD,gBAAgB,CAAc;IAC9B,4BAA4B,CAAyB;IACpD,SAAS,GAAe;QAC/B;YACE,EAAE,EAAE,GAAG;YACP,QAAQ,EAAE,EAAE;YACZ,eAAe,EAAE;gBACf;oBACE,IAAI,EAAE,UAAU;oBAChB,KAAK,EAAE,kBAAkB;oBACzB,KAAK,EAAE,EAAE;iBACV;aACF;YACD,WAAW,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,EAAE,EAAE;YAC1C,SAAS,EAAE,KAAK;SACjB;KACF,CAAC;IACO,YAAY,GAAY,KAAK,CAAC;IAC9B,KAAK,GAAW,EAAE,CAAC;IACnB,oBAAoB,GAAY,KAAK,CAAC;IACtC,sBAAsB,GAAY,KAAK,CAAC;IACxC,QAAQ,GAAY,KAAK,CAAC;IAE3B,iBAAiB,CAAoB;IAErC,gCAAgC;QACtC,IAAI,IAAI,CAAC,4BAA4B,KAAK,SAAS,EAAE,CAAC;YACpD,OAAO,+BAA+B,CAAC;QACzC,CAAC;QAED,6BAA6B,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAC;QACjE,OAAO,IAAI,CAAC,4BAA4B,CAAC;IAC3C,CAAC;IAED,iBAAiB;QACf,IAAI,CAAC,iBAAiB,GAAG,IAAI,iBAAiB,EAAE,CAAC;QACjD,IAAI,CAAC;YACH,6CAA6C;YAC7C,IAAI,IAAI,CAAC,gBAAgB,KAAK,SAAS,EAAE,CAAC;gBACxC,0BAA0B,CAAC,IAAI,CAAC,gBAAgB,CAAC,CAAC;gBAClD,MAAM,YAAY,GAAG,eAAe,CAAC,IAAI,CAAC,0BAA0B,CAAC,CAAC;gBACtE,IAAI,YAAY,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;oBAC5B,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE;wBACvC,0CAA0C,CACxC,QAAQ,CAAC,eAAe,EACxB,YAAY,CACb,CAAC;oBACJ,CAAC,CAAC,CAAC;gBACL,CAAC;gBACD,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,KAAK,EAAE,EAAE;oBAChE,IAAI,CAAC;wBACH,OAAO,uBAAuB,CAAC,WAAW,CAAC,CAAC;oBAC9C,CAAC;oBAAC,OAAO,GAAG,EAAE,CAAC;wBACb,MAAM,OAAO,GAAG,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,eAAe,CAAC;wBACrE,MAAM,IAAI,KAAK,CAAC,sCAAsC,KAAK,KAAK,OAAO,EAAE,CAAC,CAAC;oBAC7E,CAAC;gBACH,CAAC,CAAC,CAAC;YACL,CAAC;iBAAM,CAAC;gBACN,MAAM,MAAM,GAAG,IAAI,CAAC,gCAAgC,EAAE,CAAC;gBACvD,IAAI,CAAC,SAAS,GAAG,CAAC,cAAc,CAAC,MAAM,CAAC,CAAC,CAAC;YAC5C,CAAC;QACH,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,IAAI,CAAC,KAAK;gBACR,GAAG,YAAY,KAAK;oBAClB,CAAC,CAAC,GAAG,CAAC,OAAO;oBACb,CAAC,CAAC,gDAAgD,CAAC;YACvD,IAAI,CAAC,SAAS,GAAG,EAAE,CAAC;QACtB,CAAC;IACH,CAAC;IAGD,KAAK,CAAC,gBAAgB;QACpB,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC;IACxB,CAAC;IAGD,KAAK,CAAC,YAAY;QAChB,OAAO,IAAI,CAAC,SAAS,CAAC;IACxB,CAAC;IAEO,oBAAoB,GAAG,CAC7B,KAAsE,EACtE,EAAE;QACF,MAAM,EAAE,UAAU,EAAE,GAAG,EAAE,KAAK,EAAE,GAAG,KAAK,CAAC,MAAM,CAAC;QAChD,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,CACvC,EAAE,CAAC,EAAE,KAAK,UAAU,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,CACpD,CAAC;IACJ,CAAC,CAAC;IAEM,uBAAuB,GAAG,CAChC,KAA8C,EAC9C,EAAE;QACF,MAAM,EAAE,UAAU,EAAE,OAAO,EAAE,KAAK,EAAE,GAAG,KAAK,CAAC,MAAM,CAAC;QACpD,IAAI,CAAC,cAAc,CAAC,UAAU,EAAE;YAC9B,WAAW,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE;SAChC,CAAC,CAAC;IACL,CAAC,CAAC;IAEM,cAAc;QACpB,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,IAAI,CAAC,gCAAgC,EAAE,CAAC;YACvD,MAAM,WAAW,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC;YAC3C,IAAI,CAAC,SAAS,GAAG,CAAC,GAAG,IAAI,CAAC,SAAS,EAAE,WAAW,CAAC,CAAC;QACpD,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,IAAI,CAAC,KAAK;gBACR,GAAG,YAAY,KAAK;oBAClB,CAAC,CAAC,GAAG,CAAC,OAAO;oBACb,CAAC,CAAC,gDAAgD,CAAC;QACzD,CAAC;IACH,CAAC;IAEO,cAAc,CAAC,EAAU,EAAE,OAA0B;QAC3D,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,CACvC,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,EAAE,GAAG,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAC1C,CAAC;IACJ,CAAC;IAEO,kBAAkB,CAAC,QAAkB;QAC3C,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;YACrC,MAAM,OAAO,GAAsB;gBACjC,MAAM,EAAE,QAAQ,CAAC,QAAQ;gBACzB,OAAO;gBACP,MAAM;aACP,CAAC;YACF,IAAI,QAAQ,CAAC,WAAW,EAAE,OAAO,EAAE,CAAC;gBAClC,OAAO,CAAC,WAAW,GAAG,QAAQ,CAAC,WAAW,CAAC,KAAK,CAAC;YACnD,CAAC;YACD,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;QAChC,CAAC,CAAC,CAAC;IACL,CAAC;IAEO,UAAU,CAAC,MAAe;QAChC,MAAM,MAAM,YAAY,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;IACrE,CAAC;IAEO,eAAe,CAAC,MAAe,EAAE,QAAgB;QACvD,OAAO,MAAM,YAAY,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC;IAC7D,CAAC;IAEO,KAAK,CAAC,aAAa,CAAC,QAAkB;QAC5C,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC7B,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QACtD,MAAM,CAAC,UAAU,EAAE,iBAAiB,CAAC,GAAG,MAAM,OAAO,CAAC,UAAU,CAAC;YAC/D,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC;YACjC,8BAA8B,CAAC,QAAQ,EAAE,IAAI,CAAC,sBAAsB,CAAC;SACtE,CAAC,CAAC;QAEH,MAAM,YAAY,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;QAE5C,IAAI,UAAU,CAAC,MAAM,KAAK,UAAU,EAAE,CAAC;YACrC,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAC/B,SAAS,EAAE,KAAK;gBAChB,MAAM,EAAE,SAAS;gBACjB,KAAK,EAAE,IAAI,CAAC,eAAe,CAAC,UAAU,CAAC,MAAM,EAAE,eAAe,CAAC;gBAC/D,YAAY;aACb,CAAC,CAAC;YACH,IAAI,CAAC,UAAU,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;QACrC,CAAC;QACD,MAAM,UAAU,GAAG,UAAU,CAAC,KAAK,CAAC;QAEpC,IAAI,iBAAiB,CAAC,MAAM,KAAK,UAAU,EAAE,CAAC;YAC5C,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAC/B,SAAS,EAAE,KAAK;gBAChB,MAAM,EAAE,UAAU;gBAClB,KAAK,EAAE,IAAI,CAAC,eAAe,CACzB,iBAAiB,CAAC,MAAM,EACxB,6CAA6C,CAC9C;gBACD,YAAY;aACb,CAAC,CAAC;YACH,IAAI,CAAC,UAAU,CAAC,iBAAiB,CAAC,MAAM,CAAC,CAAC;QAC5C,CAAC;QACD,MAAM,gBAAgB,GAAG,iBAAiB,CAAC,KAAK,CAAC;QAEjD,MAAM,qBAAqB,GAAa;YACtC,GAAG,gBAAgB;YACnB,MAAM,EAAE,UAAU;YAClB,YAAY;SACb,CAAC;QAEF,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE;YAC/B,SAAS,EAAE,KAAK;YAChB,MAAM,EAAE,UAAU;YAClB,KAAK,EAAE,IAAI;YACX,YAAY;YACZ,eAAe,EAAE,qBAAqB,CAAC,eAAe;SACvD,CAAC,CAAC;QAEH,MAAM,IAAI,CAAC,gBAAgB,CAAC,qBAAqB,CAAC,CAAC;IACrD,CAAC;IAEO,cAAc,CAAC,EAAU;QAC/B,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC;IAC7D,CAAC;IAEO,2BAA2B,GAAG,CACpC,KAA+C,EAC/C,EAAE;QACF,MAAM,EAAE,UAAU,EAAE,GAAG,MAAM,EAAE,GAAG,KAAK,CAAC,MAAM,CAAC;QAE/C,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;YACvC,IAAI,EAAE,CAAC,EAAE,KAAK,UAAU,EAAE,CAAC;gBACzB,OAAO,EAAE,CAAC;YACZ,CAAC;YAED,OAAO,iBAAiB,CAAC,0BAA0B,CAAC,EAAE,EAAE,MAAM,CAAC,CAAC;QAClE,CAAC,CAAC,CAAC;IACL,CAAC,CAAC;IAEM,KAAK,CAAC,gBAAgB,CAAC,QAAkB;QAC/C,MAAM,IAAI,CAAC,iBAAiB,CAAC,gBAAgB,CAC3C,QAAQ,EACR,CAAC,MAAwB,EAAE,EAAE;YAC3B,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAC/B,gBAAgB,EAAE,MAAM;aACzB,CAAC,CAAC;QACL,CAAC,EACD,IAAI,CAAC,0BAA0B,CAChC,CAAC;IACJ,CAAC;IAEO,KAAK,CAAC,WAAW;QACvB,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC;QACzB,MAAM,KAAK,GAAG,EAAE,CAAC;QACjB,KAAK,MAAM,QAAQ,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YACtC,IAAI,CAAC,QAAQ,CAAC,SAAS,IAAI,QAAQ,CAAC,QAAQ,CAAC,IAAI,EAAE,EAAE,CAAC;gBACpD,KAAK,CAAC,IAAI,CAAC,GAAG,EAAE,CACd,IAAI,CAAC,aAAa,CAAC,QAAQ,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE;oBACvC,OAAO,CAAC,KAAK,CAAC,gBAAgB,QAAQ,CAAC,EAAE,SAAS,EAAE,GAAG,CAAC,CAAC;gBAC3D,CAAC,CAAC,CACH,CAAC;YACJ,CAAC;QACH,CAAC;QACD,IAAI,CAAC;YACH,MAAM,OAAO,GAAG,IAAI,kBAAkB,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;YACrD,MAAM,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,OAAO,CAAC,KAAK,CAAC,6BAA6B,EAAE,GAAG,CAAC,CAAC;QACpD,CAAC;QACD,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC;IAC5B,CAAC;IAEO,KAAK,CAAC,YAAY,CAAC,IAAU;QACnC,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,KAAK,kBAAkB,CAAC;QACpD,MAAM,eAAe,GAAG,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;QAElE,IAAI,CAAC,UAAU,IAAI,CAAC,eAAe,EAAE,CAAC;YACpC,IAAI,CAAC,KAAK,GAAG,+CAA+C,CAAC;YAC7D,OAAO;QACT,CAAC;QAED,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC;QAEhB,IAAI,CAAC;YACH,MAAM,OAAO,GAAG,MAAM,aAAa,CAAC,IAAI,CAAC,CAAC;YAC1C,MAAM,MAAM,GAAG,eAAe,CAC5B,OAAO,EACP,eAAe,CAAC,IAAI,CAAC,0BAA0B,CAAC,CACjD,CAAC;YAEF,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;gBACpB,IAAI,CAAC,KAAK,GAAG,MAAM,CAAC,KAAK,IAAI,uCAAuC,CAAC;gBACrE,OAAO;YACT,CAAC;YAED,IAAI,CAAC,SAAS,GAAG,MAAM,CAAC,SAAS,IAAI,EAAE,CAAC;QAC1C,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,IAAI,CAAC,KAAK;gBACR,GAAG,YAAY,KAAK;oBAClB,CAAC,CAAC,GAAG,CAAC,OAAO;oBACb,CAAC,CAAC,gEAAgE,CAAC;YACvE,OAAO,CAAC,KAAK,CAAC,wBAAwB,EAAE,GAAG,CAAC,CAAC;QAC/C,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,qBAAqB;QACjC,IAAI,CAAC,oBAAoB,GAAG,IAAI,CAAC;QACjC,IAAI,CAAC;YACH,MAAM,WAAW,GAAG,qBAAqB,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YAE1D,gDAAgD;YAChD,MAAM,IAAI,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC,UAAU,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC;YAEvD,YAAY,CAAC,WAAW,EAAE,iBAAiB,EAAE,kBAAkB,CAAC,CAAC;QACnE,CAAC;gBAAS,CAAC;YACT,IAAI,CAAC,oBAAoB,GAAG,KAAK,CAAC;QACpC,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,uBAAuB;QACnC,IAAI,CAAC,sBAAsB,GAAG,IAAI,CAAC;QACnC,IAAI,CAAC;YACH,MAAM,UAAU,GAAG,sBAAsB,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YAE1D,gDAAgD;YAChD,MAAM,IAAI,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC,UAAU,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC;YAEvD,YAAY,CAAC,UAAU,EAAE,kBAAkB,EAAE,UAAU,CAAC,CAAC;QAC3D,CAAC;gBAAS,CAAC;YACT,IAAI,CAAC,sBAAsB,GAAG,KAAK,CAAC;QACtC,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,UAAU;QACtB,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC;QACrB,IAAI,CAAC;YACH,MAAM,OAAO,GAAG;gBACd,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;gBACnC,SAAS,EAAE,IAAI,CAAC,SAAS;aAC1B,CAAC;YACF,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;YAExB,yEAAyE;YACzE,UAAU,CAAC,GAAG,EAAE;gBACd,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;oBAClB,OAAO,CAAC,IAAI,CAAC,2CAA2C,CAAC,CAAC;oBAC1D,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC;gBACxB,CAAC;YACH,CAAC,EAAE,KAAK,CAAC,CAAC;QACZ,CAAC;gBAAS,CAAC;YACT,iEAAiE;YACjE,mEAAmE;QACrE,CAAC;IACH,CAAC;IAED,MAAM;QACJ,OAAO,CACL,4DAAK,KAAK,EAAC,uBAAuB;YAChC,EAAC,mBAAmB,qDAClB,oBAAoB,EAAE,IAAI,CAAC,oBAAoB,EAC/C,sBAAsB,EAAE,IAAI,CAAC,sBAAsB,EACnD,YAAY,EAAE,IAAI,CAAC,YAAY,EAC/B,OAAO,EAAE,IAAI,CAAC,OAAO,EACrB,QAAQ,EAAE,IAAI,CAAC,QAAQ,EACvB,eAAe,EAAE,IAAI,CAAC,eAAe,EACrC,QAAQ,EAAE,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC,EACzC,aAAa,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,qBAAqB,EAAE,EACjD,eAAe,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,uBAAuB,EAAE,EACrD,QAAQ,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,WAAW,EAAE,EAClC,MAAM,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,UAAU,EAAE,GAC/B;YACF,EAAC,YAAY,qDAAC,OAAO,EAAE,IAAI,CAAC,KAAK,EAAE,OAAO,EAAE,GAAG,EAAE,CAAC,CAAC,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC,GAAI;YACvE,4DAAK,KAAK,EAAC,gCAAgC;gBACzC,EAAC,YAAY,qDACX,SAAS,EAAE,IAAI,CAAC,SAAS,EACzB,0BAA0B,EAAE,CAAC,CAAC,IAAI,CAAC,sBAAsB,EACzD,YAAY,EAAE,eAAe,CAAC,IAAI,CAAC,0BAA0B,CAAC,EAC9D,KAAK,EAAE,QAAQ,CAAC,EAAE,CAAC,IAAI,CAAC,aAAa,CAAC,QAAQ,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,EAC/D,QAAQ,EAAE,EAAE,CAAC,EAAE,CAAC,IAAI,CAAC,cAAc,CAAC,EAAE,CAAC,EACvC,aAAa,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,cAAc,EAAE,EAC1C,oBAAoB,EAAE,IAAI,CAAC,oBAAoB,EAC/C,uBAAuB,EAAE,IAAI,CAAC,2BAA2B,EACzD,mBAAmB,EAAE,IAAI,CAAC,uBAAuB,GACjD,CACE,CACF,CACP,CAAC;IACJ,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CACF","sourcesContent":["import {\n Component,\n State,\n Prop,\n h,\n EventEmitter,\n Event,\n Method,\n} from '@stencil/core';\nimport { EvaluationResult } from '../../lib/evaluation/types';\nimport { ErrorMessage } from '../error-message/error-message';\nimport { RateLimitedFetcher } from '../../lib/rate-limited-fetcher/rate-limited-fetcher';\nimport {\n ExpectedOutcomeSchema,\n TestCase,\n LLMRequestPayload,\n SavePayload,\n ModelResponsePayload,\n EvaluationSourceExtractors,\n} from '../../types/llm-test-runner';\nimport { readFileAsync } from '../../lib/file/file-reader';\nimport { downloadFile } from '../../lib/file/file-download';\nimport { formatTestSuiteAsJson } from '../../lib/import-export/test-suite-exporter';\nimport { exportTestResultsToCsv } from '../../lib/import-export/test-results-csv';\nimport { importTestSuite } from '../../lib/import-export/test-suite-importer';\nimport {\n createTestCase,\n createTestCaseFromInput,\n DEFAULT_EXPECTED_OUTCOME_SCHEMA,\n} from '../../lib/test-cases/test-case-factory';\nimport {\n type ExpectedOutcomeResolver,\n resolveDynamicExpectedOutcomes,\n} from '../../lib/test-cases/dynamic-expected-outcome-resolver';\nimport * as TestCaseMutations from '../../lib/test-cases/test-case-mutations';\nimport { EvaluationService } from '../../lib/evaluation/evaluation-service';\nimport { validateTestCaseInputArray } from '../../schemas/test-case';\nimport {\n getExtractorIds,\n validateExpectedOutcomeArrayWithExtractors,\n validateExpectedOutcomeSchema,\n} from '../../schemas/expected-outcome';\nimport { LLMTestRunnerHeader } from './header/llm-test-runner-header';\nimport { LLMTestCases } from './test-cases/llm-test-cases';\nimport { ExpectedOutcomeChangeDetail } from './test-cases/expected-outcome-renderer';\nimport type { ChatHistoryRowChangeDetail } from './test-cases/llm-test-case-row';\n\n@Component({\n tag: 'llm-test-runner',\n styleUrls: [\n '../../styles/tokens.css',\n 'llm-test-runner.css',\n 'header/llm-test-runner-header.css',\n 'test-cases/llm-test-cases.css',\n 'test-cases/llm-test-case-row.css',\n 'test-cases/actions/row-actions.css',\n 'test-cases/evaluation/evaluation-summary.css',\n 'test-cases/output/response-output.css',\n '../error-message/error-message.css',\n '../../lib/ui/button/button.css',\n '../../lib/ui/icon-button/icon-button.css',\n ],\n shadow: true,\n})\nexport class LLMTestRunner {\n @Event() llmRequest: EventEmitter<LLMRequestPayload>;\n @Event() save: EventEmitter<SavePayload>;\n @Prop() delayMs?: number = 500;\n @Prop() useSave?: boolean = false;\n @Prop() usePromptEditor?: boolean = false;\n @Prop() resolveExpectedOutcome?: ExpectedOutcomeResolver;\n @Prop() evaluationSourceExtractors?: EvaluationSourceExtractors;\n @Prop() initialTestCases?: TestCase[];\n @Prop() defaultExpectedOutcomeSchema?: ExpectedOutcomeSchema;\n @State() testCases: TestCase[] = [\n {\n id: '1',\n question: '',\n expectedOutcome: [\n {\n type: 'textarea',\n label: 'Expected Outcome',\n value: '',\n },\n ],\n chatHistory: { enabled: false, value: '' },\n isRunning: false,\n },\n ];\n @State() isRunningAll: boolean = false;\n @State() error: string = '';\n @State() isExportingTestSuite: boolean = false;\n @State() isExportingTestResults: boolean = false;\n @State() isSaving: boolean = false;\n\n private evaluationService: EvaluationService;\n\n private getResolvedExpectedOutcomeSchema(): ExpectedOutcomeSchema {\n if (this.defaultExpectedOutcomeSchema === undefined) {\n return DEFAULT_EXPECTED_OUTCOME_SCHEMA;\n }\n\n validateExpectedOutcomeSchema(this.defaultExpectedOutcomeSchema);\n return this.defaultExpectedOutcomeSchema;\n }\n\n componentWillLoad() {\n this.evaluationService = new EvaluationService();\n try {\n // Initialize testCases from prop if provided\n if (this.initialTestCases !== undefined) {\n validateTestCaseInputArray(this.initialTestCases);\n const extractorIds = getExtractorIds(this.evaluationSourceExtractors);\n if (extractorIds.length > 0) {\n this.initialTestCases.forEach(testCase => {\n validateExpectedOutcomeArrayWithExtractors(\n testCase.expectedOutcome,\n extractorIds,\n );\n });\n }\n this.testCases = this.initialTestCases.map((rawTestCase, index) => {\n try {\n return createTestCaseFromInput(rawTestCase);\n } catch (err) {\n const message = err instanceof Error ? err.message : 'Unknown error';\n throw new Error(`Invalid initial test case at index ${index}: ${message}`);\n }\n });\n } else {\n const schema = this.getResolvedExpectedOutcomeSchema();\n this.testCases = [createTestCase(schema)];\n }\n } catch (err) {\n this.error =\n err instanceof Error\n ? err.message\n : 'Invalid defaultExpectedOutcomeSchema provided.';\n this.testCases = [];\n }\n }\n\n @Method()\n async resetSavingState(): Promise<void> {\n this.isSaving = false;\n }\n\n @Method()\n async getTestCases(): Promise<TestCase[]> {\n return this.testCases;\n }\n\n private handleTestCaseChange = (\n event: CustomEvent<{ testCaseId: string; key: string; value: string }>,\n ) => {\n const { testCaseId, key, value } = event.detail;\n this.testCases = this.testCases.map(tc =>\n tc.id === testCaseId ? { ...tc, [key]: value } : tc,\n );\n };\n\n private handleChatHistoryChange = (\n event: CustomEvent<ChatHistoryRowChangeDetail>,\n ) => {\n const { testCaseId, enabled, value } = event.detail;\n this.updateTestCase(testCaseId, {\n chatHistory: { enabled, value },\n });\n };\n\n private addNewTestCase() {\n try {\n const schema = this.getResolvedExpectedOutcomeSchema();\n const newTestCase = createTestCase(schema);\n this.testCases = [...this.testCases, newTestCase];\n } catch (err) {\n this.error =\n err instanceof Error\n ? err.message\n : 'Invalid defaultExpectedOutcomeSchema provided.';\n }\n }\n\n private updateTestCase(id: string, updates: Partial<TestCase>) {\n this.testCases = this.testCases.map(tc =>\n tc.id === id ? { ...tc, ...updates } : tc,\n );\n }\n\n private requestLlmResponse(testCase: TestCase): Promise<ModelResponsePayload> {\n return new Promise((resolve, reject) => {\n const payload: LLMRequestPayload = {\n prompt: testCase.question,\n resolve,\n reject,\n };\n if (testCase.chatHistory?.enabled) {\n payload.chatHistory = testCase.chatHistory.value;\n }\n this.llmRequest.emit(payload);\n });\n }\n\n private throwError(reason: unknown): never {\n throw reason instanceof Error ? reason : new Error(String(reason));\n }\n\n private addErrorMessage(reason: unknown, fallback: string): string {\n return reason instanceof Error ? reason.message : fallback;\n }\n\n private async runSingleTest(testCase: TestCase): Promise<void> {\n const startTime = Date.now();\n this.updateTestCase(testCase.id, { isRunning: true });\n const [llmSettled, resolutionSettled] = await Promise.allSettled([\n this.requestLlmResponse(testCase),\n resolveDynamicExpectedOutcomes(testCase, this.resolveExpectedOutcome),\n ]);\n\n const responseTime = Date.now() - startTime;\n\n if (llmSettled.status === 'rejected') {\n this.updateTestCase(testCase.id, {\n isRunning: false,\n output: undefined,\n error: this.addErrorMessage(llmSettled.reason, 'Unknown error'),\n responseTime,\n });\n this.throwError(llmSettled.reason);\n }\n const aiResponse = llmSettled.value;\n\n if (resolutionSettled.status === 'rejected') {\n this.updateTestCase(testCase.id, {\n isRunning: false,\n output: aiResponse,\n error: this.addErrorMessage(\n resolutionSettled.reason,\n 'Failed to resolve dynamic expected outcome.',\n ),\n responseTime,\n });\n this.throwError(resolutionSettled.reason);\n }\n const resolvedTestCase = resolutionSettled.value;\n\n const forEvaluationTestCase: TestCase = {\n ...resolvedTestCase,\n output: aiResponse,\n responseTime,\n };\n\n this.updateTestCase(testCase.id, {\n isRunning: false,\n output: aiResponse,\n error: null,\n responseTime,\n expectedOutcome: forEvaluationTestCase.expectedOutcome,\n });\n\n await this.evaluateResponse(forEvaluationTestCase);\n }\n\n private deleteTestCase(id: string) {\n this.testCases = this.testCases.filter(tc => tc.id !== id);\n }\n\n private handleExpectedOutcomeChange = (\n event: CustomEvent<ExpectedOutcomeChangeDetail>,\n ) => {\n const { testCaseId, ...change } = event.detail;\n\n this.testCases = this.testCases.map(tc => {\n if (tc.id !== testCaseId) {\n return tc;\n }\n\n return TestCaseMutations.applyExpectedOutcomeChange(tc, change);\n });\n };\n\n private async evaluateResponse(testCase: TestCase): Promise<void> {\n await this.evaluationService.evaluateTestCase(\n testCase,\n (result: EvaluationResult) => {\n this.updateTestCase(testCase.id, {\n evaluationResult: result,\n });\n },\n this.evaluationSourceExtractors,\n );\n }\n\n private async runAllTests() {\n this.isRunningAll = true;\n const tasks = [];\n for (const testCase of this.testCases) {\n if (!testCase.isRunning && testCase.question.trim()) {\n tasks.push(() =>\n this.runSingleTest(testCase).catch(err => {\n console.error(`⚠️ Test case ${testCase.id} failed`, err);\n }),\n );\n }\n }\n try {\n const fetcher = new RateLimitedFetcher(this.delayMs);\n await fetcher.runAll(tasks);\n } catch (err) {\n console.error('⚠️ Error running all tests:', err);\n }\n this.isRunningAll = false;\n }\n\n private async handleImport(file: File): Promise<void> {\n const isJsonType = file.type === 'application/json';\n const isJsonExtension = file.name.toLowerCase().endsWith('.json');\n\n if (!isJsonType && !isJsonExtension) {\n this.error = 'Invalid file type. Please select a JSON file.';\n return;\n }\n\n this.error = '';\n\n try {\n const content = await readFileAsync(file);\n const result = importTestSuite(\n content,\n getExtractorIds(this.evaluationSourceExtractors),\n );\n\n if (!result.success) {\n this.error = result.error || 'Unknown error occurred during import.';\n return;\n }\n\n this.testCases = result.testCases || [];\n } catch (err) {\n this.error =\n err instanceof Error\n ? err.message\n : 'Error processing file. Please ensure it is a valid JSON array.';\n console.error('File Processing Error:', err);\n }\n }\n\n private async handleExportTestSuite() {\n this.isExportingTestSuite = true;\n try {\n const jsonContent = formatTestSuiteAsJson(this.testCases);\n\n // Added a small delay to show the loading state\n await new Promise(resolve => setTimeout(resolve, 500));\n\n downloadFile(jsonContent, 'test-suite.json', 'application/json');\n } finally {\n this.isExportingTestSuite = false;\n }\n }\n\n private async handleExportTestResults() {\n this.isExportingTestResults = true;\n try {\n const csvContent = exportTestResultsToCsv(this.testCases);\n\n // Added a small delay to show the loading state\n await new Promise(resolve => setTimeout(resolve, 500));\n\n downloadFile(csvContent, 'test-results.csv', 'text/csv');\n } finally {\n this.isExportingTestResults = false;\n }\n }\n\n private async handleSave() {\n this.isSaving = true;\n try {\n const testRun = {\n timestamp: new Date().toISOString(),\n testCases: this.testCases,\n };\n this.save.emit(testRun);\n\n // Failsafe: Auto-reset saving state after 10 seconds to prevent stuck UI\n setTimeout(() => {\n if (this.isSaving) {\n console.warn('Save operation timed out, resetting state');\n this.isSaving = false;\n }\n }, 10000);\n } finally {\n // Parent will call resetSavingState() when actual save completes\n // If not called within 10 seconds, failsafe above will reset state\n }\n }\n\n render() {\n return (\n <div class=\"test-runner-container\">\n <LLMTestRunnerHeader\n isExportingTestSuite={this.isExportingTestSuite}\n isExportingTestResults={this.isExportingTestResults}\n isRunningAll={this.isRunningAll}\n useSave={this.useSave}\n isSaving={this.isSaving}\n usePromptEditor={this.usePromptEditor}\n onImport={file => this.handleImport(file)}\n onExportSuite={() => this.handleExportTestSuite()}\n onExportResults={() => this.handleExportTestResults()}\n onRunAll={() => this.runAllTests()}\n onSave={() => this.handleSave()}\n />\n <ErrorMessage message={this.error} onClear={() => (this.error = '')} />\n <div class=\"test-runner-container__content\">\n <LLMTestCases\n testCases={this.testCases}\n dynamicResolutionSupported={!!this.resolveExpectedOutcome}\n extractorIds={getExtractorIds(this.evaluationSourceExtractors)}\n onRun={testCase => this.runSingleTest(testCase).catch(() => {})}\n onDelete={id => this.deleteTestCase(id)}\n onAddTestCase={() => this.addNewTestCase()}\n handleTestCaseChange={this.handleTestCaseChange}\n onExpectedOutcomeChange={this.handleExpectedOutcomeChange}\n onChatHistoryChange={this.handleChatHistoryChange}\n />\n </div>\n </div>\n );\n }\n}\n"]}
|