@midscene/core 0.26.7-beta-20250821033353.0 → 0.26.7-beta-20250821074936.0
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/dist/es/ai-model/common.mjs +79 -13
- package/dist/es/ai-model/common.mjs.map +1 -1
- package/dist/es/ai-model/index.mjs +2 -2
- package/dist/es/ai-model/inspect.mjs.map +1 -1
- package/dist/es/ai-model/llm-planning.mjs +0 -1
- package/dist/es/ai-model/llm-planning.mjs.map +1 -1
- package/dist/es/ai-model/prompt/llm-planning.mjs +43 -21
- package/dist/es/ai-model/prompt/llm-planning.mjs.map +1 -1
- package/dist/es/index.mjs +2 -2
- package/dist/es/index.mjs.map +1 -1
- package/dist/es/insight/index.mjs +61 -3
- package/dist/es/insight/index.mjs.map +1 -1
- package/dist/es/types.mjs.map +1 -1
- package/dist/es/utils.mjs +2 -2
- package/dist/es/utils.mjs.map +1 -1
- package/dist/lib/ai-model/common.js +97 -19
- package/dist/lib/ai-model/common.js.map +1 -1
- package/dist/lib/ai-model/index.js +22 -10
- package/dist/lib/ai-model/inspect.js.map +1 -1
- package/dist/lib/ai-model/llm-planning.js +0 -1
- package/dist/lib/ai-model/llm-planning.js.map +1 -1
- package/dist/lib/ai-model/prompt/llm-planning.js +43 -21
- package/dist/lib/ai-model/prompt/llm-planning.js.map +1 -1
- package/dist/lib/index.js +14 -8
- package/dist/lib/index.js.map +1 -1
- package/dist/lib/insight/index.js +58 -0
- package/dist/lib/insight/index.js.map +1 -1
- package/dist/lib/types.js.map +1 -1
- package/dist/lib/utils.js +2 -2
- package/dist/lib/utils.js.map +1 -1
- package/dist/types/ai-model/common.d.ts +363 -7
- package/dist/types/ai-model/index.d.ts +1 -1
- package/dist/types/ai-model/inspect.d.ts +2 -1
- package/dist/types/index.d.ts +1 -1
- package/dist/types/insight/index.d.ts +5 -1
- package/dist/types/types.d.ts +2 -26
- package/dist/types/yaml.d.ts +4 -2
- package/package.json +3 -3
|
@@ -14,7 +14,6 @@ var common_AIActionType = /*#__PURE__*/ function(AIActionType) {
|
|
|
14
14
|
AIActionType[AIActionType["DESCRIBE_ELEMENT"] = 4] = "DESCRIBE_ELEMENT";
|
|
15
15
|
return AIActionType;
|
|
16
16
|
}({});
|
|
17
|
-
const actionSpaceTypePrefix = 'action_space_';
|
|
18
17
|
async function callAiFn(msgs, AIActionTypeValue, modelPreferences) {
|
|
19
18
|
const jsonObject = await callToGetJSONObject(msgs, AIActionTypeValue, modelPreferences);
|
|
20
19
|
return {
|
|
@@ -174,18 +173,17 @@ async function markupImageForLLM(screenshotBase64, tree, size) {
|
|
|
174
173
|
function buildYamlFlowFromPlans(plans, actionSpace, sleep) {
|
|
175
174
|
const flow = [];
|
|
176
175
|
for (const plan of plans){
|
|
177
|
-
var _plan_locate;
|
|
178
176
|
const verb = plan.type;
|
|
179
177
|
const action = actionSpace.find((action)=>action.name === verb);
|
|
180
178
|
if (!action) {
|
|
181
179
|
console.warn(`Cannot convert action ${verb} to yaml flow. Will ignore it.`);
|
|
182
180
|
continue;
|
|
183
181
|
}
|
|
184
|
-
const
|
|
185
|
-
const
|
|
182
|
+
const flowKey = action.interfaceAlias || verb;
|
|
183
|
+
const flowParam = action.paramSchema ? dumpActionParam(plan.param || {}, action.paramSchema) : {};
|
|
186
184
|
const flowItem = {
|
|
187
|
-
[flowKey]:
|
|
188
|
-
...
|
|
185
|
+
[flowKey]: '',
|
|
186
|
+
...flowParam
|
|
189
187
|
};
|
|
190
188
|
flow.push(flowItem);
|
|
191
189
|
}
|
|
@@ -206,35 +204,103 @@ const SizeSchema = z.object({
|
|
|
206
204
|
const RectSchema = PointSchema.and(SizeSchema).and(z.object({
|
|
207
205
|
zoom: z.number().optional()
|
|
208
206
|
}));
|
|
209
|
-
const
|
|
210
|
-
|
|
211
|
-
|
|
207
|
+
const TMultimodalPromptSchema = z.object({
|
|
208
|
+
images: z.array(z.object({
|
|
209
|
+
name: z.string(),
|
|
210
|
+
url: z.string()
|
|
211
|
+
})).optional(),
|
|
212
|
+
convertHttpImage2Base64: z.boolean().optional()
|
|
213
|
+
});
|
|
214
|
+
const TUserPromptSchema = z.union([
|
|
215
|
+
z.string(),
|
|
216
|
+
z.object({
|
|
217
|
+
prompt: z.string()
|
|
218
|
+
}).and(TMultimodalPromptSchema.partial())
|
|
219
|
+
]);
|
|
220
|
+
const locateFieldFlagName = 'midscene_location_field_flag';
|
|
221
|
+
const MidsceneLocationResult = z.object({
|
|
222
|
+
[locateFieldFlagName]: z.literal(true),
|
|
223
|
+
prompt: TUserPromptSchema,
|
|
224
|
+
deepThink: z.boolean().optional(),
|
|
225
|
+
cacheable: z.boolean().optional(),
|
|
226
|
+
xpath: z.boolean().optional(),
|
|
212
227
|
center: z.tuple([
|
|
213
228
|
z.number(),
|
|
214
229
|
z.number()
|
|
215
230
|
]),
|
|
216
231
|
rect: RectSchema
|
|
217
232
|
}).passthrough();
|
|
233
|
+
const getMidsceneLocationSchema = ()=>MidsceneLocationResult;
|
|
218
234
|
const ifMidsceneLocatorField = (field)=>{
|
|
219
235
|
var _actualField__def, _actualField__def1;
|
|
220
236
|
let actualField = field;
|
|
221
237
|
if ((null == (_actualField__def = actualField._def) ? void 0 : _actualField__def.typeName) === 'ZodOptional') actualField = actualField._def.innerType;
|
|
222
238
|
if ((null == (_actualField__def1 = actualField._def) ? void 0 : _actualField__def1.typeName) === 'ZodObject') {
|
|
223
239
|
const shape = actualField._def.shape();
|
|
224
|
-
return
|
|
240
|
+
return locateFieldFlagName in shape;
|
|
225
241
|
}
|
|
226
242
|
return false;
|
|
227
243
|
};
|
|
228
|
-
const
|
|
244
|
+
const dumpMidsceneLocatorField = (field)=>{
|
|
245
|
+
assert(ifMidsceneLocatorField(field), 'field is not a midscene locator field');
|
|
246
|
+
if ('string' == typeof field) return field;
|
|
247
|
+
if (field && 'object' == typeof field && field.prompt) {
|
|
248
|
+
if ('string' == typeof field.prompt) return field.prompt;
|
|
249
|
+
if ('object' == typeof field.prompt && field.prompt.prompt) return field.prompt.prompt;
|
|
250
|
+
}
|
|
251
|
+
return String(field);
|
|
252
|
+
};
|
|
253
|
+
const findAllMidsceneLocatorField = (zodType, requiredOnly)=>{
|
|
229
254
|
var _zodObject__def;
|
|
230
255
|
if (!zodType) return [];
|
|
231
256
|
const zodObject = zodType;
|
|
232
257
|
if ((null == (_zodObject__def = zodObject._def) ? void 0 : _zodObject__def.typeName) === 'ZodObject' && zodObject.shape) {
|
|
233
258
|
const keys = Object.keys(zodObject.shape);
|
|
234
|
-
return keys.filter((key)=>
|
|
259
|
+
return keys.filter((key)=>{
|
|
260
|
+
const field = zodObject.shape[key];
|
|
261
|
+
if (!ifMidsceneLocatorField(field)) return false;
|
|
262
|
+
if (requiredOnly) {
|
|
263
|
+
var _field__def;
|
|
264
|
+
return (null == (_field__def = field._def) ? void 0 : _field__def.typeName) !== 'ZodOptional';
|
|
265
|
+
}
|
|
266
|
+
return true;
|
|
267
|
+
});
|
|
235
268
|
}
|
|
236
269
|
return [];
|
|
237
270
|
};
|
|
238
|
-
|
|
271
|
+
const dumpActionParam = (jsonObject, zodSchema)=>{
|
|
272
|
+
const locatorFields = findAllMidsceneLocatorField(zodSchema);
|
|
273
|
+
const result = {
|
|
274
|
+
...jsonObject
|
|
275
|
+
};
|
|
276
|
+
for (const fieldName of locatorFields){
|
|
277
|
+
const fieldValue = result[fieldName];
|
|
278
|
+
if (fieldValue) {
|
|
279
|
+
if ('string' == typeof fieldValue) result[fieldName] = fieldValue;
|
|
280
|
+
else if ('object' == typeof fieldValue) {
|
|
281
|
+
if (fieldValue.prompt) {
|
|
282
|
+
if ('string' == typeof fieldValue.prompt) result[fieldName] = fieldValue.prompt;
|
|
283
|
+
else if ('object' == typeof fieldValue.prompt && fieldValue.prompt.prompt) result[fieldName] = fieldValue.prompt.prompt;
|
|
284
|
+
}
|
|
285
|
+
}
|
|
286
|
+
}
|
|
287
|
+
}
|
|
288
|
+
return result;
|
|
289
|
+
};
|
|
290
|
+
const loadActionParam = (jsonObject, zodSchema)=>{
|
|
291
|
+
const locatorFields = findAllMidsceneLocatorField(zodSchema);
|
|
292
|
+
const result = {
|
|
293
|
+
...jsonObject
|
|
294
|
+
};
|
|
295
|
+
for (const fieldName of locatorFields){
|
|
296
|
+
const fieldValue = result[fieldName];
|
|
297
|
+
if (fieldValue && 'string' == typeof fieldValue) result[fieldName] = {
|
|
298
|
+
[locateFieldFlagName]: true,
|
|
299
|
+
prompt: fieldValue
|
|
300
|
+
};
|
|
301
|
+
}
|
|
302
|
+
return result;
|
|
303
|
+
};
|
|
304
|
+
export { common_AIActionType as AIActionType, PointSchema, RectSchema, SizeSchema, TMultimodalPromptSchema, TUserPromptSchema, adaptBbox, adaptBboxToRect, adaptDoubaoBbox, adaptGeminiBbox, adaptQwenBbox, buildYamlFlowFromPlans, callAiFn, dumpActionParam, dumpMidsceneLocatorField, expandSearchArea, fillBboxParam, findAllMidsceneLocatorField, getMidsceneLocationSchema, ifMidsceneLocatorField, loadActionParam, markupImageForLLM, mergeRects, warnGPT4oSizeLimit };
|
|
239
305
|
|
|
240
306
|
//# sourceMappingURL=common.mjs.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"ai-model/common.mjs","sources":["webpack://@midscene/core/./src/ai-model/common.ts"],"sourcesContent":["import type {\n AIUsageInfo,\n BaseElement,\n DeviceAction,\n ElementTreeNode,\n MidsceneYamlFlowItem,\n PlanningAction,\n Rect,\n Size,\n} from '@/types';\nimport { assert } from '@midscene/shared/utils';\n\nimport type {\n ChatCompletionSystemMessageParam,\n ChatCompletionUserMessageParam,\n} from 'openai/resources/index';\nimport { callToGetJSONObject } from './service-caller/index';\n\nimport type { PlanningLocateParam } from '@/types';\nimport { NodeType } from '@midscene/shared/constants';\nimport {\n type IModelPreferences,\n getModelName,\n vlLocateMode,\n} from '@midscene/shared/env';\nimport { treeToList } from '@midscene/shared/extractor';\nimport { compositeElementInfoImg } from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { z } from 'zod';\n\nexport type AIArgs = [\n ChatCompletionSystemMessageParam,\n ...ChatCompletionUserMessageParam[],\n];\n\nexport enum AIActionType {\n ASSERT = 0,\n INSPECT_ELEMENT = 1,\n EXTRACT_DATA = 2,\n PLAN = 3,\n DESCRIBE_ELEMENT = 4,\n}\n\nexport const actionSpaceTypePrefix = 'action_space_';\n\nexport async function callAiFn<T>(\n msgs: AIArgs,\n AIActionTypeValue: AIActionType,\n modelPreferences: IModelPreferences,\n): Promise<{ content: T; usage?: AIUsageInfo }> {\n const jsonObject = await callToGetJSONObject<T>(\n msgs,\n AIActionTypeValue,\n modelPreferences,\n );\n\n return {\n content: jsonObject.content,\n usage: jsonObject.usage,\n };\n}\n\nconst defaultBboxSize = 20; // must be even number\nconst debugInspectUtils = getDebug('ai:common');\n\n// transform the param of locate from qwen mode\nexport function fillBboxParam(\n locate: PlanningLocateParam,\n width: number,\n height: number,\n modelPreferences: IModelPreferences,\n) {\n // The Qwen model might have hallucinations of naming bbox as bbox_2d.\n if ((locate as any).bbox_2d && !locate?.bbox) {\n locate.bbox = (locate as any).bbox_2d;\n // biome-ignore lint/performance/noDelete: <explanation>\n delete (locate as any).bbox_2d;\n }\n\n if (locate?.bbox) {\n locate.bbox = adaptBbox(locate.bbox, width, height, modelPreferences);\n }\n\n return locate;\n}\n\nexport function adaptQwenBbox(\n bbox: number[],\n): [number, number, number, number] {\n if (bbox.length < 2) {\n const msg = `invalid bbox data for qwen-vl mode: ${JSON.stringify(bbox)} `;\n throw new Error(msg);\n }\n\n const result: [number, number, number, number] = [\n Math.round(bbox[0]),\n Math.round(bbox[1]),\n typeof bbox[2] === 'number'\n ? Math.round(bbox[2])\n : Math.round(bbox[0] + defaultBboxSize),\n typeof bbox[3] === 'number'\n ? Math.round(bbox[3])\n : Math.round(bbox[1] + defaultBboxSize),\n ];\n return result;\n}\n\nexport function adaptDoubaoBbox(\n bbox: string[] | number[] | string,\n width: number,\n height: number,\n): [number, number, number, number] {\n assert(\n width > 0 && height > 0,\n 'width and height must be greater than 0 in doubao mode',\n );\n\n if (typeof bbox === 'string') {\n assert(\n /^(\\d+)\\s(\\d+)\\s(\\d+)\\s(\\d+)$/.test(bbox.trim()),\n `invalid bbox data string for doubao-vision mode: ${bbox}`,\n );\n const splitted = bbox.split(' ');\n if (splitted.length === 4) {\n return [\n Math.round((Number(splitted[0]) * width) / 1000),\n Math.round((Number(splitted[1]) * height) / 1000),\n Math.round((Number(splitted[2]) * width) / 1000),\n Math.round((Number(splitted[3]) * height) / 1000),\n ];\n }\n throw new Error(`invalid bbox data string for doubao-vision mode: ${bbox}`);\n }\n\n if (Array.isArray(bbox) && Array.isArray(bbox[0])) {\n bbox = bbox[0];\n }\n\n let bboxList: number[] = [];\n if (Array.isArray(bbox) && typeof bbox[0] === 'string') {\n bbox.forEach((item) => {\n if (typeof item === 'string' && item.includes(',')) {\n const [x, y] = item.split(',');\n bboxList.push(Number(x.trim()), Number(y.trim()));\n } else if (typeof item === 'string' && item.includes(' ')) {\n const [x, y] = item.split(' ');\n bboxList.push(Number(x.trim()), Number(y.trim()));\n } else {\n bboxList.push(Number(item));\n }\n });\n } else {\n bboxList = bbox as any;\n }\n\n if (bboxList.length === 4 || bboxList.length === 5) {\n return [\n Math.round((bboxList[0] * width) / 1000),\n Math.round((bboxList[1] * height) / 1000),\n Math.round((bboxList[2] * width) / 1000),\n Math.round((bboxList[3] * height) / 1000),\n ];\n }\n\n // treat the bbox as a center point\n if (\n bboxList.length === 6 ||\n bboxList.length === 2 ||\n bboxList.length === 3 ||\n bboxList.length === 7\n ) {\n return [\n Math.max(\n 0,\n Math.round((bboxList[0] * width) / 1000) - defaultBboxSize / 2,\n ),\n Math.max(\n 0,\n Math.round((bboxList[1] * height) / 1000) - defaultBboxSize / 2,\n ),\n Math.min(\n width,\n Math.round((bboxList[0] * width) / 1000) + defaultBboxSize / 2,\n ),\n Math.min(\n height,\n Math.round((bboxList[1] * height) / 1000) + defaultBboxSize / 2,\n ),\n ];\n }\n\n if (bbox.length === 8) {\n return [\n Math.round((bboxList[0] * width) / 1000),\n Math.round((bboxList[1] * height) / 1000),\n Math.round((bboxList[4] * width) / 1000),\n Math.round((bboxList[5] * height) / 1000),\n ];\n }\n\n const msg = `invalid bbox data for doubao-vision mode: ${JSON.stringify(bbox)} `;\n throw new Error(msg);\n}\n\nexport function adaptBbox(\n bbox: number[],\n width: number,\n height: number,\n modelPreferences: IModelPreferences,\n): [number, number, number, number] {\n if (\n vlLocateMode(modelPreferences) === 'doubao-vision' ||\n vlLocateMode(modelPreferences) === 'vlm-ui-tars'\n ) {\n return adaptDoubaoBbox(bbox, width, height);\n }\n\n if (vlLocateMode(modelPreferences) === 'gemini') {\n return adaptGeminiBbox(bbox, width, height);\n }\n\n return adaptQwenBbox(bbox);\n}\n\nexport function adaptGeminiBbox(\n bbox: number[],\n width: number,\n height: number,\n): [number, number, number, number] {\n const left = Math.round((bbox[1] * width) / 1000);\n const top = Math.round((bbox[0] * height) / 1000);\n const right = Math.round((bbox[3] * width) / 1000);\n const bottom = Math.round((bbox[2] * height) / 1000);\n return [left, top, right, bottom];\n}\n\nexport function adaptBboxToRect(\n bbox: number[],\n width: number,\n height: number,\n modelPreferences: IModelPreferences,\n offsetX = 0,\n offsetY = 0,\n): Rect {\n debugInspectUtils('adaptBboxToRect', bbox, width, height, offsetX, offsetY);\n const [left, top, right, bottom] = adaptBbox(\n bbox,\n width,\n height,\n modelPreferences,\n );\n const rect = {\n left: left + offsetX,\n top: top + offsetY,\n width: right - left,\n height: bottom - top,\n };\n debugInspectUtils('adaptBboxToRect, result=', rect);\n return rect;\n}\n\nlet warned = false;\nexport function warnGPT4oSizeLimit(\n size: Size,\n modelPreferences: IModelPreferences,\n) {\n if (warned) return;\n if (getModelName(modelPreferences)?.toLowerCase().includes('gpt-4o')) {\n const warningMsg = `GPT-4o has a maximum image input size of 2000x768 or 768x2000, but got ${size.width}x${size.height}. Please set your page to a smaller resolution. Otherwise, the result may be inaccurate.`;\n\n if (\n Math.max(size.width, size.height) > 2000 ||\n Math.min(size.width, size.height) > 768\n ) {\n console.warn(warningMsg);\n warned = true;\n }\n } else if (size.width > 1800 || size.height > 1800) {\n console.warn(\n `The image size seems too large (${size.width}x${size.height}). It may lead to more token usage, slower response, and inaccurate result.`,\n );\n warned = true;\n }\n}\n\nexport function mergeRects(rects: Rect[]) {\n const minLeft = Math.min(...rects.map((r) => r.left));\n const minTop = Math.min(...rects.map((r) => r.top));\n const maxRight = Math.max(...rects.map((r) => r.left + r.width));\n const maxBottom = Math.max(...rects.map((r) => r.top + r.height));\n return {\n left: minLeft,\n top: minTop,\n width: maxRight - minLeft,\n height: maxBottom - minTop,\n };\n}\n\n// expand the search area to at least 300 x 300, or add a default padding\nexport function expandSearchArea(\n rect: Rect,\n screenSize: Size,\n modelPreferences: IModelPreferences,\n) {\n const minEdgeSize =\n vlLocateMode(modelPreferences) === 'doubao-vision' ? 500 : 300;\n const defaultPadding = 160;\n\n const paddingSizeHorizontal =\n rect.width < minEdgeSize\n ? Math.ceil((minEdgeSize - rect.width) / 2)\n : defaultPadding;\n const paddingSizeVertical =\n rect.height < minEdgeSize\n ? Math.ceil((minEdgeSize - rect.height) / 2)\n : defaultPadding;\n rect.left = Math.max(0, rect.left - paddingSizeHorizontal);\n rect.width = Math.min(\n rect.width + paddingSizeHorizontal * 2,\n screenSize.width - rect.left,\n );\n rect.top = Math.max(0, rect.top - paddingSizeVertical);\n rect.height = Math.min(\n rect.height + paddingSizeVertical * 2,\n screenSize.height - rect.top,\n );\n return rect;\n}\n\nexport async function markupImageForLLM(\n screenshotBase64: string,\n tree: ElementTreeNode<BaseElement>,\n size: Size,\n) {\n const elementsInfo = treeToList(tree);\n const elementsPositionInfoWithoutText = elementsInfo!.filter(\n (elementInfo) => {\n if (elementInfo.attributes.nodeType === NodeType.TEXT) {\n return false;\n }\n return true;\n },\n );\n\n const imagePayload = await compositeElementInfoImg({\n inputImgBase64: screenshotBase64,\n elementsPositionInfo: elementsPositionInfoWithoutText,\n size,\n });\n return imagePayload;\n}\n\nexport function buildYamlFlowFromPlans(\n plans: PlanningAction[],\n actionSpace: DeviceAction<any>[],\n sleep?: number,\n): MidsceneYamlFlowItem[] {\n const flow: MidsceneYamlFlowItem[] = [];\n\n for (const plan of plans) {\n const verb = plan.type;\n\n const action = actionSpace.find((action) => action.name === verb);\n if (!action) {\n console.warn(\n `Cannot convert action ${verb} to yaml flow. Will ignore it.`,\n );\n continue;\n }\n\n const locate = plan.locate?.prompt;\n const flowKey = action.interfaceAlias || `${actionSpaceTypePrefix}${verb}`;\n\n const flowItem: MidsceneYamlFlowItem = {\n [flowKey]: locate || '',\n ...(plan.param || {}),\n };\n\n flow.push(flowItem);\n }\n\n if (sleep) {\n flow.push({\n sleep,\n });\n }\n\n return flow;\n}\n\n// Zod schemas for shared types\nexport const PointSchema = z.object({\n left: z.number(),\n top: z.number(),\n});\n\nexport const SizeSchema = z.object({\n width: z.number(),\n height: z.number(),\n dpr: z.number().optional(),\n});\n\nexport const RectSchema = PointSchema.and(SizeSchema).and(\n z.object({\n zoom: z.number().optional(),\n }),\n);\n\nexport const MidsceneLocation = z\n .object({\n midscene_location_field_flag: z.literal(true),\n prompt: z.string(),\n center: z.tuple([z.number(), z.number()]),\n rect: RectSchema,\n })\n .passthrough();\n\nexport type MidsceneLocationType = z.infer<typeof MidsceneLocation>;\n\nexport const ifMidsceneLocatorField = (field: any): boolean => {\n // Handle optional fields by getting the inner type\n let actualField = field;\n if (actualField._def?.typeName === 'ZodOptional') {\n actualField = actualField._def.innerType;\n }\n\n // Check if this is a ZodObject with midscene_location_field_flag\n if (actualField._def?.typeName === 'ZodObject') {\n const shape = actualField._def.shape();\n return 'midscene_location_field_flag' in shape;\n }\n\n return false;\n};\n\nexport const findAllMidsceneLocatorField = (\n zodType?: z.ZodType<any>,\n): string[] => {\n if (!zodType) {\n return [];\n }\n\n // Check if this is a ZodObject by checking if it has a shape property\n const zodObject = zodType as any;\n if (zodObject._def?.typeName === 'ZodObject' && zodObject.shape) {\n const keys = Object.keys(zodObject.shape);\n return keys.filter((key) => ifMidsceneLocatorField(zodObject.shape[key]));\n }\n\n // For other ZodType instances, we can't extract field names\n return [];\n};\n"],"names":["AIActionType","actionSpaceTypePrefix","callAiFn","msgs","AIActionTypeValue","modelPreferences","jsonObject","callToGetJSONObject","defaultBboxSize","debugInspectUtils","getDebug","fillBboxParam","locate","width","height","adaptBbox","adaptQwenBbox","bbox","msg","JSON","Error","result","Math","adaptDoubaoBbox","assert","splitted","Number","Array","bboxList","item","x","y","vlLocateMode","adaptGeminiBbox","left","top","right","bottom","adaptBboxToRect","offsetX","offsetY","rect","warned","warnGPT4oSizeLimit","size","_getModelName","warningMsg","console","mergeRects","rects","minLeft","r","minTop","maxRight","maxBottom","expandSearchArea","screenSize","minEdgeSize","defaultPadding","paddingSizeHorizontal","paddingSizeVertical","markupImageForLLM","screenshotBase64","tree","elementsInfo","treeToList","elementsPositionInfoWithoutText","elementInfo","NodeType","imagePayload","compositeElementInfoImg","buildYamlFlowFromPlans","plans","actionSpace","sleep","flow","plan","_plan_locate","verb","action","flowKey","flowItem","PointSchema","z","SizeSchema","RectSchema","MidsceneLocation","ifMidsceneLocatorField","field","_actualField__def","_actualField__def1","actualField","shape","findAllMidsceneLocatorField","zodType","_zodObject__def","zodObject","keys","Object","key"],"mappings":";;;;;;;;AAmCO,IAAKA,sBAAYA,WAAAA,GAAAA,SAAZA,YAAY;;;;;;WAAZA;;AAQL,MAAMC,wBAAwB;AAE9B,eAAeC,SACpBC,IAAY,EACZC,iBAA+B,EAC/BC,gBAAmC;IAEnC,MAAMC,aAAa,MAAMC,oBACvBJ,MACAC,mBACAC;IAGF,OAAO;QACL,SAASC,WAAW,OAAO;QAC3B,OAAOA,WAAW,KAAK;IACzB;AACF;AAEA,MAAME,kBAAkB;AACxB,MAAMC,oBAAoBC,SAAS;AAG5B,SAASC,cACdC,MAA2B,EAC3BC,KAAa,EACbC,MAAc,EACdT,gBAAmC;IAGnC,IAAKO,OAAe,OAAO,IAAI,CAACA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,IAAI,AAAD,GAAG;QAC5CA,OAAO,IAAI,GAAIA,OAAe,OAAO;QAErC,OAAQA,OAAe,OAAO;IAChC;IAEA,IAAIA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,IAAI,EACdA,OAAO,IAAI,GAAGG,UAAUH,OAAO,IAAI,EAAEC,OAAOC,QAAQT;IAGtD,OAAOO;AACT;AAEO,SAASI,cACdC,IAAc;IAEd,IAAIA,KAAK,MAAM,GAAG,GAAG;QACnB,MAAMC,MAAM,CAAC,oCAAoC,EAAEC,KAAK,SAAS,CAACF,MAAM,CAAC,CAAC;QAC1E,MAAM,IAAIG,MAAMF;IAClB;IAEA,MAAMG,SAA2C;QAC/CC,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE;QAClBK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE;QACC,YAAnB,OAAOA,IAAI,CAAC,EAAE,GACVK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE,IAClBK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE,GAAGT;QACN,YAAnB,OAAOS,IAAI,CAAC,EAAE,GACVK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE,IAClBK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE,GAAGT;KAC1B;IACD,OAAOa;AACT;AAEO,SAASE,gBACdN,IAAkC,EAClCJ,KAAa,EACbC,MAAc;IAEdU,OACEX,QAAQ,KAAKC,SAAS,GACtB;IAGF,IAAI,AAAgB,YAAhB,OAAOG,MAAmB;QAC5BO,OACE,+BAA+B,IAAI,CAACP,KAAK,IAAI,KAC7C,CAAC,iDAAiD,EAAEA,MAAM;QAE5D,MAAMQ,WAAWR,KAAK,KAAK,CAAC;QAC5B,IAAIQ,AAAoB,MAApBA,SAAS,MAAM,EACjB,OAAO;YACLH,KAAK,KAAK,CAAEI,OAAOD,QAAQ,CAAC,EAAE,IAAIZ,QAAS;YAC3CS,KAAK,KAAK,CAAEI,OAAOD,QAAQ,CAAC,EAAE,IAAIX,SAAU;YAC5CQ,KAAK,KAAK,CAAEI,OAAOD,QAAQ,CAAC,EAAE,IAAIZ,QAAS;YAC3CS,KAAK,KAAK,CAAEI,OAAOD,QAAQ,CAAC,EAAE,IAAIX,SAAU;SAC7C;QAEH,MAAM,IAAIM,MAAM,CAAC,iDAAiD,EAAEH,MAAM;IAC5E;IAEA,IAAIU,MAAM,OAAO,CAACV,SAASU,MAAM,OAAO,CAACV,IAAI,CAAC,EAAE,GAC9CA,OAAOA,IAAI,CAAC,EAAE;IAGhB,IAAIW,WAAqB,EAAE;IAC3B,IAAID,MAAM,OAAO,CAACV,SAAS,AAAmB,YAAnB,OAAOA,IAAI,CAAC,EAAE,EACvCA,KAAK,OAAO,CAAC,CAACY;QACZ,IAAI,AAAgB,YAAhB,OAAOA,QAAqBA,KAAK,QAAQ,CAAC,MAAM;YAClD,MAAM,CAACC,GAAGC,EAAE,GAAGF,KAAK,KAAK,CAAC;YAC1BD,SAAS,IAAI,CAACF,OAAOI,EAAE,IAAI,KAAKJ,OAAOK,EAAE,IAAI;QAC/C,OAAO,IAAI,AAAgB,YAAhB,OAAOF,QAAqBA,KAAK,QAAQ,CAAC,MAAM;YACzD,MAAM,CAACC,GAAGC,EAAE,GAAGF,KAAK,KAAK,CAAC;YAC1BD,SAAS,IAAI,CAACF,OAAOI,EAAE,IAAI,KAAKJ,OAAOK,EAAE,IAAI;QAC/C,OACEH,SAAS,IAAI,CAACF,OAAOG;IAEzB;SAEAD,WAAWX;IAGb,IAAIW,AAAoB,MAApBA,SAAS,MAAM,IAAUA,AAAoB,MAApBA,SAAS,MAAM,EAC1C,OAAO;QACLN,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS;QACnCS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU;QACpCQ,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS;QACnCS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU;KACrC;IAIH,IACEc,AAAoB,MAApBA,SAAS,MAAM,IACfA,AAAoB,MAApBA,SAAS,MAAM,IACfA,AAAoB,MAApBA,SAAS,MAAM,IACfA,AAAoB,MAApBA,SAAS,MAAM,EAEf,OAAO;QACLN,KAAK,GAAG,CACN,GACAA,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS,QAAQL,kBAAkB;QAE/Dc,KAAK,GAAG,CACN,GACAA,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU,QAAQN,kBAAkB;QAEhEc,KAAK,GAAG,CACNT,OACAS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS,QAAQL,kBAAkB;QAE/Dc,KAAK,GAAG,CACNR,QACAQ,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU,QAAQN,kBAAkB;KAEjE;IAGH,IAAIS,AAAgB,MAAhBA,KAAK,MAAM,EACb,OAAO;QACLK,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS;QACnCS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU;QACpCQ,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS;QACnCS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU;KACrC;IAGH,MAAMI,MAAM,CAAC,0CAA0C,EAAEC,KAAK,SAAS,CAACF,MAAM,CAAC,CAAC;IAChF,MAAM,IAAIG,MAAMF;AAClB;AAEO,SAASH,UACdE,IAAc,EACdJ,KAAa,EACbC,MAAc,EACdT,gBAAmC;IAEnC,IACE2B,AAAmC,oBAAnCA,aAAa3B,qBACb2B,AAAmC,kBAAnCA,aAAa3B,mBAEb,OAAOkB,gBAAgBN,MAAMJ,OAAOC;IAGtC,IAAIkB,AAAmC,aAAnCA,aAAa3B,mBACf,OAAO4B,gBAAgBhB,MAAMJ,OAAOC;IAGtC,OAAOE,cAAcC;AACvB;AAEO,SAASgB,gBACdhB,IAAc,EACdJ,KAAa,EACbC,MAAc;IAEd,MAAMoB,OAAOZ,KAAK,KAAK,CAAEL,IAAI,CAAC,EAAE,GAAGJ,QAAS;IAC5C,MAAMsB,MAAMb,KAAK,KAAK,CAAEL,IAAI,CAAC,EAAE,GAAGH,SAAU;IAC5C,MAAMsB,QAAQd,KAAK,KAAK,CAAEL,IAAI,CAAC,EAAE,GAAGJ,QAAS;IAC7C,MAAMwB,SAASf,KAAK,KAAK,CAAEL,IAAI,CAAC,EAAE,GAAGH,SAAU;IAC/C,OAAO;QAACoB;QAAMC;QAAKC;QAAOC;KAAO;AACnC;AAEO,SAASC,gBACdrB,IAAc,EACdJ,KAAa,EACbC,MAAc,EACdT,gBAAmC,EACnCkC,UAAU,CAAC,EACXC,UAAU,CAAC;IAEX/B,kBAAkB,mBAAmBQ,MAAMJ,OAAOC,QAAQyB,SAASC;IACnE,MAAM,CAACN,MAAMC,KAAKC,OAAOC,OAAO,GAAGtB,UACjCE,MACAJ,OACAC,QACAT;IAEF,MAAMoC,OAAO;QACX,MAAMP,OAAOK;QACb,KAAKJ,MAAMK;QACX,OAAOJ,QAAQF;QACf,QAAQG,SAASF;IACnB;IACA1B,kBAAkB,4BAA4BgC;IAC9C,OAAOA;AACT;AAEA,IAAIC,SAAS;AACN,SAASC,mBACdC,IAAU,EACVvC,gBAAmC;QAG/BwC;IADJ,IAAIH,QAAQ;IACZ,IAAI,QAAAG,CAAAA,gBAAAA,aAAaxC,iBAAgB,IAA7BwC,KAAAA,IAAAA,cAAgC,WAAW,GAAG,QAAQ,CAAC,WAAW;QACpE,MAAMC,aAAa,CAAC,uEAAuE,EAAEF,KAAK,KAAK,CAAC,CAAC,EAAEA,KAAK,MAAM,CAAC,wFAAwF,CAAC;QAEhN,IACEtB,KAAK,GAAG,CAACsB,KAAK,KAAK,EAAEA,KAAK,MAAM,IAAI,QACpCtB,KAAK,GAAG,CAACsB,KAAK,KAAK,EAAEA,KAAK,MAAM,IAAI,KACpC;YACAG,QAAQ,IAAI,CAACD;YACbJ,SAAS;QACX;IACF,OAAO,IAAIE,KAAK,KAAK,GAAG,QAAQA,KAAK,MAAM,GAAG,MAAM;QAClDG,QAAQ,IAAI,CACV,CAAC,gCAAgC,EAAEH,KAAK,KAAK,CAAC,CAAC,EAAEA,KAAK,MAAM,CAAC,2EAA2E,CAAC;QAE3IF,SAAS;IACX;AACF;AAEO,SAASM,WAAWC,KAAa;IACtC,MAAMC,UAAU5B,KAAK,GAAG,IAAI2B,MAAM,GAAG,CAAC,CAACE,IAAMA,EAAE,IAAI;IACnD,MAAMC,SAAS9B,KAAK,GAAG,IAAI2B,MAAM,GAAG,CAAC,CAACE,IAAMA,EAAE,GAAG;IACjD,MAAME,WAAW/B,KAAK,GAAG,IAAI2B,MAAM,GAAG,CAAC,CAACE,IAAMA,EAAE,IAAI,GAAGA,EAAE,KAAK;IAC9D,MAAMG,YAAYhC,KAAK,GAAG,IAAI2B,MAAM,GAAG,CAAC,CAACE,IAAMA,EAAE,GAAG,GAAGA,EAAE,MAAM;IAC/D,OAAO;QACL,MAAMD;QACN,KAAKE;QACL,OAAOC,WAAWH;QAClB,QAAQI,YAAYF;IACtB;AACF;AAGO,SAASG,iBACdd,IAAU,EACVe,UAAgB,EAChBnD,gBAAmC;IAEnC,MAAMoD,cACJzB,AAAmC,oBAAnCA,aAAa3B,oBAAwC,MAAM;IAC7D,MAAMqD,iBAAiB;IAEvB,MAAMC,wBACJlB,KAAK,KAAK,GAAGgB,cACTnC,KAAK,IAAI,CAAEmC,AAAAA,CAAAA,cAAchB,KAAK,KAAI,IAAK,KACvCiB;IACN,MAAME,sBACJnB,KAAK,MAAM,GAAGgB,cACVnC,KAAK,IAAI,CAAEmC,AAAAA,CAAAA,cAAchB,KAAK,MAAK,IAAK,KACxCiB;IACNjB,KAAK,IAAI,GAAGnB,KAAK,GAAG,CAAC,GAAGmB,KAAK,IAAI,GAAGkB;IACpClB,KAAK,KAAK,GAAGnB,KAAK,GAAG,CACnBmB,KAAK,KAAK,GAAGkB,AAAwB,IAAxBA,uBACbH,WAAW,KAAK,GAAGf,KAAK,IAAI;IAE9BA,KAAK,GAAG,GAAGnB,KAAK,GAAG,CAAC,GAAGmB,KAAK,GAAG,GAAGmB;IAClCnB,KAAK,MAAM,GAAGnB,KAAK,GAAG,CACpBmB,KAAK,MAAM,GAAGmB,AAAsB,IAAtBA,qBACdJ,WAAW,MAAM,GAAGf,KAAK,GAAG;IAE9B,OAAOA;AACT;AAEO,eAAeoB,kBACpBC,gBAAwB,EACxBC,IAAkC,EAClCnB,IAAU;IAEV,MAAMoB,eAAeC,WAAWF;IAChC,MAAMG,kCAAkCF,aAAc,MAAM,CAC1D,CAACG;QACC,IAAIA,YAAY,UAAU,CAAC,QAAQ,KAAKC,SAAS,IAAI,EACnD,OAAO;QAET,OAAO;IACT;IAGF,MAAMC,eAAe,MAAMC,wBAAwB;QACjD,gBAAgBR;QAChB,sBAAsBI;QACtBtB;IACF;IACA,OAAOyB;AACT;AAEO,SAASE,uBACdC,KAAuB,EACvBC,WAAgC,EAChCC,KAAc;IAEd,MAAMC,OAA+B,EAAE;IAEvC,KAAK,MAAMC,QAAQJ,MAAO;YAWTK;QAVf,MAAMC,OAAOF,KAAK,IAAI;QAEtB,MAAMG,SAASN,YAAY,IAAI,CAAC,CAACM,SAAWA,OAAO,IAAI,KAAKD;QAC5D,IAAI,CAACC,QAAQ;YACXhC,QAAQ,IAAI,CACV,CAAC,sBAAsB,EAAE+B,KAAK,8BAA8B,CAAC;YAE/D;QACF;QAEA,MAAMlE,SAAS,QAAAiE,CAAAA,eAAAA,KAAK,MAAM,AAAD,IAAVA,KAAAA,IAAAA,aAAa,MAAM;QAClC,MAAMG,UAAUD,OAAO,cAAc,IAAI,GAAG9E,wBAAwB6E,MAAM;QAE1E,MAAMG,WAAiC;YACrC,CAACD,QAAQ,EAAEpE,UAAU;YACrB,GAAIgE,KAAK,KAAK,IAAI,CAAC,CAAC;QACtB;QAEAD,KAAK,IAAI,CAACM;IACZ;IAEA,IAAIP,OACFC,KAAK,IAAI,CAAC;QACRD;IACF;IAGF,OAAOC;AACT;AAGO,MAAMO,cAAcC,EAAE,MAAM,CAAC;IAClC,MAAMA,EAAE,MAAM;IACd,KAAKA,EAAE,MAAM;AACf;AAEO,MAAMC,aAAaD,EAAE,MAAM,CAAC;IACjC,OAAOA,EAAE,MAAM;IACf,QAAQA,EAAE,MAAM;IAChB,KAAKA,EAAE,MAAM,GAAG,QAAQ;AAC1B;AAEO,MAAME,aAAaH,YAAY,GAAG,CAACE,YAAY,GAAG,CACvDD,EAAE,MAAM,CAAC;IACP,MAAMA,EAAE,MAAM,GAAG,QAAQ;AAC3B;AAGK,MAAMG,mBAAmBH,EAAAA,MACvB,CAAC;IACN,8BAA8BA,EAAE,OAAO,CAAC;IACxC,QAAQA,EAAE,MAAM;IAChB,QAAQA,EAAE,KAAK,CAAC;QAACA,EAAE,MAAM;QAAIA,EAAE,MAAM;KAAG;IACxC,MAAME;AACR,GACC,WAAW;AAIP,MAAME,yBAAyB,CAACC;QAGjCC,mBAKAC;IANJ,IAAIC,cAAcH;IAClB,IAAIC,AAAAA,SAAAA,CAAAA,oBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,kBAAkB,QAAQ,AAAD,MAAM,eACjCE,cAAcA,YAAY,IAAI,CAAC,SAAS;IAI1C,IAAID,AAAAA,SAAAA,CAAAA,qBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,mBAAkB,QAAQ,AAAD,MAAM,aAAa;QAC9C,MAAME,QAAQD,YAAY,IAAI,CAAC,KAAK;QACpC,OAAO,kCAAkCC;IAC3C;IAEA,OAAO;AACT;AAEO,MAAMC,8BAA8B,CACzCC;QAQIC;IANJ,IAAI,CAACD,SACH,OAAO,EAAE;IAIX,MAAME,YAAYF;IAClB,IAAIC,AAAAA,SAAAA,CAAAA,kBAAAA,UAAU,IAAI,AAAD,IAAbA,KAAAA,IAAAA,gBAAgB,QAAQ,AAAD,MAAM,eAAeC,UAAU,KAAK,EAAE;QAC/D,MAAMC,OAAOC,OAAO,IAAI,CAACF,UAAU,KAAK;QACxC,OAAOC,KAAK,MAAM,CAAC,CAACE,MAAQZ,uBAAuBS,UAAU,KAAK,CAACG,IAAI;IACzE;IAGA,OAAO,EAAE;AACX"}
|
|
1
|
+
{"version":3,"file":"ai-model/common.mjs","sources":["webpack://@midscene/core/./src/ai-model/common.ts"],"sourcesContent":["import type {\n AIUsageInfo,\n BaseElement,\n DeviceAction,\n ElementTreeNode,\n MidsceneYamlFlowItem,\n PlanningAction,\n Rect,\n Size,\n} from '@/types';\nimport { assert } from '@midscene/shared/utils';\n\nimport type {\n ChatCompletionSystemMessageParam,\n ChatCompletionUserMessageParam,\n} from 'openai/resources/index';\nimport { callToGetJSONObject } from './service-caller/index';\n\nimport type { PlanningLocateParam } from '@/types';\nimport { NodeType } from '@midscene/shared/constants';\nimport {\n type IModelPreferences,\n getModelName,\n vlLocateMode,\n} from '@midscene/shared/env';\nimport { treeToList } from '@midscene/shared/extractor';\nimport { compositeElementInfoImg } from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { z } from 'zod';\n\nexport type AIArgs = [\n ChatCompletionSystemMessageParam,\n ...ChatCompletionUserMessageParam[],\n];\n\nexport enum AIActionType {\n ASSERT = 0,\n INSPECT_ELEMENT = 1,\n EXTRACT_DATA = 2,\n PLAN = 3,\n DESCRIBE_ELEMENT = 4,\n}\n\nexport async function callAiFn<T>(\n msgs: AIArgs,\n AIActionTypeValue: AIActionType,\n modelPreferences: IModelPreferences,\n): Promise<{ content: T; usage?: AIUsageInfo }> {\n const jsonObject = await callToGetJSONObject<T>(\n msgs,\n AIActionTypeValue,\n modelPreferences,\n );\n\n return {\n content: jsonObject.content,\n usage: jsonObject.usage,\n };\n}\n\nconst defaultBboxSize = 20; // must be even number\nconst debugInspectUtils = getDebug('ai:common');\n\n// transform the param of locate from qwen mode\nexport function fillBboxParam(\n locate: PlanningLocateParam,\n width: number,\n height: number,\n modelPreferences: IModelPreferences,\n) {\n // The Qwen model might have hallucinations of naming bbox as bbox_2d.\n if ((locate as any).bbox_2d && !locate?.bbox) {\n locate.bbox = (locate as any).bbox_2d;\n // biome-ignore lint/performance/noDelete: <explanation>\n delete (locate as any).bbox_2d;\n }\n\n if (locate?.bbox) {\n locate.bbox = adaptBbox(locate.bbox, width, height, modelPreferences);\n }\n\n return locate;\n}\n\nexport function adaptQwenBbox(\n bbox: number[],\n): [number, number, number, number] {\n if (bbox.length < 2) {\n const msg = `invalid bbox data for qwen-vl mode: ${JSON.stringify(bbox)} `;\n throw new Error(msg);\n }\n\n const result: [number, number, number, number] = [\n Math.round(bbox[0]),\n Math.round(bbox[1]),\n typeof bbox[2] === 'number'\n ? Math.round(bbox[2])\n : Math.round(bbox[0] + defaultBboxSize),\n typeof bbox[3] === 'number'\n ? Math.round(bbox[3])\n : Math.round(bbox[1] + defaultBboxSize),\n ];\n return result;\n}\n\nexport function adaptDoubaoBbox(\n bbox: string[] | number[] | string,\n width: number,\n height: number,\n): [number, number, number, number] {\n assert(\n width > 0 && height > 0,\n 'width and height must be greater than 0 in doubao mode',\n );\n\n if (typeof bbox === 'string') {\n assert(\n /^(\\d+)\\s(\\d+)\\s(\\d+)\\s(\\d+)$/.test(bbox.trim()),\n `invalid bbox data string for doubao-vision mode: ${bbox}`,\n );\n const splitted = bbox.split(' ');\n if (splitted.length === 4) {\n return [\n Math.round((Number(splitted[0]) * width) / 1000),\n Math.round((Number(splitted[1]) * height) / 1000),\n Math.round((Number(splitted[2]) * width) / 1000),\n Math.round((Number(splitted[3]) * height) / 1000),\n ];\n }\n throw new Error(`invalid bbox data string for doubao-vision mode: ${bbox}`);\n }\n\n if (Array.isArray(bbox) && Array.isArray(bbox[0])) {\n bbox = bbox[0];\n }\n\n let bboxList: number[] = [];\n if (Array.isArray(bbox) && typeof bbox[0] === 'string') {\n bbox.forEach((item) => {\n if (typeof item === 'string' && item.includes(',')) {\n const [x, y] = item.split(',');\n bboxList.push(Number(x.trim()), Number(y.trim()));\n } else if (typeof item === 'string' && item.includes(' ')) {\n const [x, y] = item.split(' ');\n bboxList.push(Number(x.trim()), Number(y.trim()));\n } else {\n bboxList.push(Number(item));\n }\n });\n } else {\n bboxList = bbox as any;\n }\n\n if (bboxList.length === 4 || bboxList.length === 5) {\n return [\n Math.round((bboxList[0] * width) / 1000),\n Math.round((bboxList[1] * height) / 1000),\n Math.round((bboxList[2] * width) / 1000),\n Math.round((bboxList[3] * height) / 1000),\n ];\n }\n\n // treat the bbox as a center point\n if (\n bboxList.length === 6 ||\n bboxList.length === 2 ||\n bboxList.length === 3 ||\n bboxList.length === 7\n ) {\n return [\n Math.max(\n 0,\n Math.round((bboxList[0] * width) / 1000) - defaultBboxSize / 2,\n ),\n Math.max(\n 0,\n Math.round((bboxList[1] * height) / 1000) - defaultBboxSize / 2,\n ),\n Math.min(\n width,\n Math.round((bboxList[0] * width) / 1000) + defaultBboxSize / 2,\n ),\n Math.min(\n height,\n Math.round((bboxList[1] * height) / 1000) + defaultBboxSize / 2,\n ),\n ];\n }\n\n if (bbox.length === 8) {\n return [\n Math.round((bboxList[0] * width) / 1000),\n Math.round((bboxList[1] * height) / 1000),\n Math.round((bboxList[4] * width) / 1000),\n Math.round((bboxList[5] * height) / 1000),\n ];\n }\n\n const msg = `invalid bbox data for doubao-vision mode: ${JSON.stringify(bbox)} `;\n throw new Error(msg);\n}\n\nexport function adaptBbox(\n bbox: number[],\n width: number,\n height: number,\n modelPreferences: IModelPreferences,\n): [number, number, number, number] {\n if (\n vlLocateMode(modelPreferences) === 'doubao-vision' ||\n vlLocateMode(modelPreferences) === 'vlm-ui-tars'\n ) {\n return adaptDoubaoBbox(bbox, width, height);\n }\n\n if (vlLocateMode(modelPreferences) === 'gemini') {\n return adaptGeminiBbox(bbox, width, height);\n }\n\n return adaptQwenBbox(bbox);\n}\n\nexport function adaptGeminiBbox(\n bbox: number[],\n width: number,\n height: number,\n): [number, number, number, number] {\n const left = Math.round((bbox[1] * width) / 1000);\n const top = Math.round((bbox[0] * height) / 1000);\n const right = Math.round((bbox[3] * width) / 1000);\n const bottom = Math.round((bbox[2] * height) / 1000);\n return [left, top, right, bottom];\n}\n\nexport function adaptBboxToRect(\n bbox: number[],\n width: number,\n height: number,\n modelPreferences: IModelPreferences,\n offsetX = 0,\n offsetY = 0,\n): Rect {\n debugInspectUtils('adaptBboxToRect', bbox, width, height, offsetX, offsetY);\n const [left, top, right, bottom] = adaptBbox(\n bbox,\n width,\n height,\n modelPreferences,\n );\n const rect = {\n left: left + offsetX,\n top: top + offsetY,\n width: right - left,\n height: bottom - top,\n };\n debugInspectUtils('adaptBboxToRect, result=', rect);\n return rect;\n}\n\nlet warned = false;\nexport function warnGPT4oSizeLimit(\n size: Size,\n modelPreferences: IModelPreferences,\n) {\n if (warned) return;\n if (getModelName(modelPreferences)?.toLowerCase().includes('gpt-4o')) {\n const warningMsg = `GPT-4o has a maximum image input size of 2000x768 or 768x2000, but got ${size.width}x${size.height}. Please set your page to a smaller resolution. Otherwise, the result may be inaccurate.`;\n\n if (\n Math.max(size.width, size.height) > 2000 ||\n Math.min(size.width, size.height) > 768\n ) {\n console.warn(warningMsg);\n warned = true;\n }\n } else if (size.width > 1800 || size.height > 1800) {\n console.warn(\n `The image size seems too large (${size.width}x${size.height}). It may lead to more token usage, slower response, and inaccurate result.`,\n );\n warned = true;\n }\n}\n\nexport function mergeRects(rects: Rect[]) {\n const minLeft = Math.min(...rects.map((r) => r.left));\n const minTop = Math.min(...rects.map((r) => r.top));\n const maxRight = Math.max(...rects.map((r) => r.left + r.width));\n const maxBottom = Math.max(...rects.map((r) => r.top + r.height));\n return {\n left: minLeft,\n top: minTop,\n width: maxRight - minLeft,\n height: maxBottom - minTop,\n };\n}\n\n// expand the search area to at least 300 x 300, or add a default padding\nexport function expandSearchArea(\n rect: Rect,\n screenSize: Size,\n modelPreferences: IModelPreferences,\n) {\n const minEdgeSize =\n vlLocateMode(modelPreferences) === 'doubao-vision' ? 500 : 300;\n const defaultPadding = 160;\n\n const paddingSizeHorizontal =\n rect.width < minEdgeSize\n ? Math.ceil((minEdgeSize - rect.width) / 2)\n : defaultPadding;\n const paddingSizeVertical =\n rect.height < minEdgeSize\n ? Math.ceil((minEdgeSize - rect.height) / 2)\n : defaultPadding;\n rect.left = Math.max(0, rect.left - paddingSizeHorizontal);\n rect.width = Math.min(\n rect.width + paddingSizeHorizontal * 2,\n screenSize.width - rect.left,\n );\n rect.top = Math.max(0, rect.top - paddingSizeVertical);\n rect.height = Math.min(\n rect.height + paddingSizeVertical * 2,\n screenSize.height - rect.top,\n );\n return rect;\n}\n\nexport async function markupImageForLLM(\n screenshotBase64: string,\n tree: ElementTreeNode<BaseElement>,\n size: Size,\n) {\n const elementsInfo = treeToList(tree);\n const elementsPositionInfoWithoutText = elementsInfo!.filter(\n (elementInfo) => {\n if (elementInfo.attributes.nodeType === NodeType.TEXT) {\n return false;\n }\n return true;\n },\n );\n\n const imagePayload = await compositeElementInfoImg({\n inputImgBase64: screenshotBase64,\n elementsPositionInfo: elementsPositionInfoWithoutText,\n size,\n });\n return imagePayload;\n}\n\nexport function buildYamlFlowFromPlans(\n plans: PlanningAction[],\n actionSpace: DeviceAction<any>[],\n sleep?: number,\n): MidsceneYamlFlowItem[] {\n const flow: MidsceneYamlFlowItem[] = [];\n\n for (const plan of plans) {\n const verb = plan.type;\n\n const action = actionSpace.find((action) => action.name === verb);\n if (!action) {\n console.warn(\n `Cannot convert action ${verb} to yaml flow. Will ignore it.`,\n );\n continue;\n }\n\n const flowKey = action.interfaceAlias || verb;\n const flowParam = action.paramSchema\n ? dumpActionParam(plan.param || {}, action.paramSchema)\n : {};\n\n const flowItem: MidsceneYamlFlowItem = {\n [flowKey]: '',\n ...flowParam,\n };\n\n flow.push(flowItem);\n }\n\n if (sleep) {\n flow.push({\n sleep,\n });\n }\n\n return flow;\n}\n\n// Zod schemas for shared types\nexport const PointSchema = z.object({\n left: z.number(),\n top: z.number(),\n});\n\nexport const SizeSchema = z.object({\n width: z.number(),\n height: z.number(),\n dpr: z.number().optional(),\n});\n\nexport const RectSchema = PointSchema.and(SizeSchema).and(\n z.object({\n zoom: z.number().optional(),\n }),\n);\n\n// Zod schema for TMultimodalPrompt\nexport const TMultimodalPromptSchema = z.object({\n images: z\n .array(\n z.object({\n name: z.string(),\n url: z.string(),\n }),\n )\n .optional(),\n convertHttpImage2Base64: z.boolean().optional(),\n});\n\n// Zod schema for TUserPrompt\nexport const TUserPromptSchema = z.union([\n z.string(),\n z\n .object({\n prompt: z.string(),\n })\n .and(TMultimodalPromptSchema.partial()),\n]);\n\n// Generate TypeScript types from Zod schemas\nexport type TMultimodalPrompt = z.infer<typeof TMultimodalPromptSchema>;\nexport type TUserPrompt = z.infer<typeof TUserPromptSchema>;\n\nconst locateFieldFlagName = 'midscene_location_field_flag';\n\nconst MidsceneLocationResult = z\n .object({\n [locateFieldFlagName]: z.literal(true),\n prompt: TUserPromptSchema,\n\n // optional fields\n deepThink: z.boolean().optional(), // only available in vl model\n cacheable: z.boolean().optional(),\n xpath: z.boolean().optional(), // preset result for xpath\n\n // these two fields will only appear in the result\n center: z.tuple([z.number(), z.number()]),\n rect: RectSchema,\n })\n .passthrough();\n\nexport type MidsceneLocationResultType = z.infer<typeof MidsceneLocationResult>;\nexport const getMidsceneLocationSchema = () => {\n return MidsceneLocationResult;\n};\n\nexport const ifMidsceneLocatorField = (field: any): boolean => {\n // Handle optional fields by getting the inner type\n let actualField = field;\n if (actualField._def?.typeName === 'ZodOptional') {\n actualField = actualField._def.innerType;\n }\n\n // Check if this is a ZodUnion (the new MidsceneLocation structure)\n if (actualField._def?.typeName === 'ZodObject') {\n const shape = actualField._def.shape();\n return locateFieldFlagName in shape;\n }\n\n return false;\n};\n\nexport const dumpMidsceneLocatorField = (field: any): string => {\n assert(\n ifMidsceneLocatorField(field),\n 'field is not a midscene locator field',\n );\n\n // If field is a string, return it directly\n if (typeof field === 'string') {\n return field;\n }\n\n // If field is an object with prompt property\n if (field && typeof field === 'object' && field.prompt) {\n // If prompt is a string, return it directly\n if (typeof field.prompt === 'string') {\n return field.prompt;\n }\n // If prompt is a TUserPrompt object, extract the prompt string\n if (typeof field.prompt === 'object' && field.prompt.prompt) {\n return field.prompt.prompt; // TODO: dump images if necessary\n }\n }\n\n // Fallback: try to convert to string\n return String(field);\n};\n\nexport const findAllMidsceneLocatorField = (\n zodType?: z.ZodType<any>,\n requiredOnly?: boolean,\n): string[] => {\n if (!zodType) {\n return [];\n }\n\n // Check if this is a ZodObject by checking if it has a shape property\n const zodObject = zodType as any;\n if (zodObject._def?.typeName === 'ZodObject' && zodObject.shape) {\n const keys = Object.keys(zodObject.shape);\n return keys.filter((key) => {\n const field = zodObject.shape[key];\n if (!ifMidsceneLocatorField(field)) {\n return false;\n }\n\n // If requiredOnly is true, filter out optional fields\n if (requiredOnly) {\n return field._def?.typeName !== 'ZodOptional';\n }\n\n return true;\n });\n }\n\n // For other ZodType instances, we can't extract field names\n return [];\n};\n\nexport const dumpActionParam = (\n jsonObject: Record<string, any>,\n zodSchema: z.ZodType<any>,\n): Record<string, any> => {\n const locatorFields = findAllMidsceneLocatorField(zodSchema);\n const result = { ...jsonObject };\n\n for (const fieldName of locatorFields) {\n const fieldValue = result[fieldName];\n if (fieldValue) {\n // If it's already a string, keep it as is\n if (typeof fieldValue === 'string') {\n result[fieldName] = fieldValue;\n } else if (typeof fieldValue === 'object') {\n // Check if this field is actually a MidsceneLocationType object\n if (fieldValue.prompt) {\n // If prompt is a string, use it directly\n if (typeof fieldValue.prompt === 'string') {\n result[fieldName] = fieldValue.prompt;\n } else if (\n typeof fieldValue.prompt === 'object' &&\n fieldValue.prompt.prompt\n ) {\n // If prompt is a TUserPrompt object, extract the prompt string\n result[fieldName] = fieldValue.prompt.prompt;\n }\n }\n }\n }\n }\n\n return result;\n};\n\nexport const loadActionParam = (\n jsonObject: Record<string, any>,\n zodSchema: z.ZodType<any>,\n): Record<string, any> => {\n const locatorFields = findAllMidsceneLocatorField(zodSchema);\n const result = { ...jsonObject };\n\n for (const fieldName of locatorFields) {\n const fieldValue = result[fieldName];\n if (fieldValue && typeof fieldValue === 'string') {\n result[fieldName] = {\n [locateFieldFlagName]: true,\n prompt: fieldValue,\n };\n }\n }\n\n return result;\n};\n"],"names":["AIActionType","callAiFn","msgs","AIActionTypeValue","modelPreferences","jsonObject","callToGetJSONObject","defaultBboxSize","debugInspectUtils","getDebug","fillBboxParam","locate","width","height","adaptBbox","adaptQwenBbox","bbox","msg","JSON","Error","result","Math","adaptDoubaoBbox","assert","splitted","Number","Array","bboxList","item","x","y","vlLocateMode","adaptGeminiBbox","left","top","right","bottom","adaptBboxToRect","offsetX","offsetY","rect","warned","warnGPT4oSizeLimit","size","_getModelName","warningMsg","console","mergeRects","rects","minLeft","r","minTop","maxRight","maxBottom","expandSearchArea","screenSize","minEdgeSize","defaultPadding","paddingSizeHorizontal","paddingSizeVertical","markupImageForLLM","screenshotBase64","tree","elementsInfo","treeToList","elementsPositionInfoWithoutText","elementInfo","NodeType","imagePayload","compositeElementInfoImg","buildYamlFlowFromPlans","plans","actionSpace","sleep","flow","plan","verb","action","flowKey","flowParam","dumpActionParam","flowItem","PointSchema","z","SizeSchema","RectSchema","TMultimodalPromptSchema","TUserPromptSchema","locateFieldFlagName","MidsceneLocationResult","getMidsceneLocationSchema","ifMidsceneLocatorField","field","_actualField__def","_actualField__def1","actualField","shape","dumpMidsceneLocatorField","String","findAllMidsceneLocatorField","zodType","requiredOnly","_zodObject__def","zodObject","keys","Object","key","_field__def","zodSchema","locatorFields","fieldName","fieldValue","loadActionParam"],"mappings":";;;;;;;;AAmCO,IAAKA,sBAAYA,WAAAA,GAAAA,SAAZA,YAAY;;;;;;WAAZA;;AAQL,eAAeC,SACpBC,IAAY,EACZC,iBAA+B,EAC/BC,gBAAmC;IAEnC,MAAMC,aAAa,MAAMC,oBACvBJ,MACAC,mBACAC;IAGF,OAAO;QACL,SAASC,WAAW,OAAO;QAC3B,OAAOA,WAAW,KAAK;IACzB;AACF;AAEA,MAAME,kBAAkB;AACxB,MAAMC,oBAAoBC,SAAS;AAG5B,SAASC,cACdC,MAA2B,EAC3BC,KAAa,EACbC,MAAc,EACdT,gBAAmC;IAGnC,IAAKO,OAAe,OAAO,IAAI,CAACA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,IAAI,AAAD,GAAG;QAC5CA,OAAO,IAAI,GAAIA,OAAe,OAAO;QAErC,OAAQA,OAAe,OAAO;IAChC;IAEA,IAAIA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,IAAI,EACdA,OAAO,IAAI,GAAGG,UAAUH,OAAO,IAAI,EAAEC,OAAOC,QAAQT;IAGtD,OAAOO;AACT;AAEO,SAASI,cACdC,IAAc;IAEd,IAAIA,KAAK,MAAM,GAAG,GAAG;QACnB,MAAMC,MAAM,CAAC,oCAAoC,EAAEC,KAAK,SAAS,CAACF,MAAM,CAAC,CAAC;QAC1E,MAAM,IAAIG,MAAMF;IAClB;IAEA,MAAMG,SAA2C;QAC/CC,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE;QAClBK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE;QACC,YAAnB,OAAOA,IAAI,CAAC,EAAE,GACVK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE,IAClBK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE,GAAGT;QACN,YAAnB,OAAOS,IAAI,CAAC,EAAE,GACVK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE,IAClBK,KAAK,KAAK,CAACL,IAAI,CAAC,EAAE,GAAGT;KAC1B;IACD,OAAOa;AACT;AAEO,SAASE,gBACdN,IAAkC,EAClCJ,KAAa,EACbC,MAAc;IAEdU,OACEX,QAAQ,KAAKC,SAAS,GACtB;IAGF,IAAI,AAAgB,YAAhB,OAAOG,MAAmB;QAC5BO,OACE,+BAA+B,IAAI,CAACP,KAAK,IAAI,KAC7C,CAAC,iDAAiD,EAAEA,MAAM;QAE5D,MAAMQ,WAAWR,KAAK,KAAK,CAAC;QAC5B,IAAIQ,AAAoB,MAApBA,SAAS,MAAM,EACjB,OAAO;YACLH,KAAK,KAAK,CAAEI,OAAOD,QAAQ,CAAC,EAAE,IAAIZ,QAAS;YAC3CS,KAAK,KAAK,CAAEI,OAAOD,QAAQ,CAAC,EAAE,IAAIX,SAAU;YAC5CQ,KAAK,KAAK,CAAEI,OAAOD,QAAQ,CAAC,EAAE,IAAIZ,QAAS;YAC3CS,KAAK,KAAK,CAAEI,OAAOD,QAAQ,CAAC,EAAE,IAAIX,SAAU;SAC7C;QAEH,MAAM,IAAIM,MAAM,CAAC,iDAAiD,EAAEH,MAAM;IAC5E;IAEA,IAAIU,MAAM,OAAO,CAACV,SAASU,MAAM,OAAO,CAACV,IAAI,CAAC,EAAE,GAC9CA,OAAOA,IAAI,CAAC,EAAE;IAGhB,IAAIW,WAAqB,EAAE;IAC3B,IAAID,MAAM,OAAO,CAACV,SAAS,AAAmB,YAAnB,OAAOA,IAAI,CAAC,EAAE,EACvCA,KAAK,OAAO,CAAC,CAACY;QACZ,IAAI,AAAgB,YAAhB,OAAOA,QAAqBA,KAAK,QAAQ,CAAC,MAAM;YAClD,MAAM,CAACC,GAAGC,EAAE,GAAGF,KAAK,KAAK,CAAC;YAC1BD,SAAS,IAAI,CAACF,OAAOI,EAAE,IAAI,KAAKJ,OAAOK,EAAE,IAAI;QAC/C,OAAO,IAAI,AAAgB,YAAhB,OAAOF,QAAqBA,KAAK,QAAQ,CAAC,MAAM;YACzD,MAAM,CAACC,GAAGC,EAAE,GAAGF,KAAK,KAAK,CAAC;YAC1BD,SAAS,IAAI,CAACF,OAAOI,EAAE,IAAI,KAAKJ,OAAOK,EAAE,IAAI;QAC/C,OACEH,SAAS,IAAI,CAACF,OAAOG;IAEzB;SAEAD,WAAWX;IAGb,IAAIW,AAAoB,MAApBA,SAAS,MAAM,IAAUA,AAAoB,MAApBA,SAAS,MAAM,EAC1C,OAAO;QACLN,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS;QACnCS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU;QACpCQ,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS;QACnCS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU;KACrC;IAIH,IACEc,AAAoB,MAApBA,SAAS,MAAM,IACfA,AAAoB,MAApBA,SAAS,MAAM,IACfA,AAAoB,MAApBA,SAAS,MAAM,IACfA,AAAoB,MAApBA,SAAS,MAAM,EAEf,OAAO;QACLN,KAAK,GAAG,CACN,GACAA,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS,QAAQL,kBAAkB;QAE/Dc,KAAK,GAAG,CACN,GACAA,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU,QAAQN,kBAAkB;QAEhEc,KAAK,GAAG,CACNT,OACAS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS,QAAQL,kBAAkB;QAE/Dc,KAAK,GAAG,CACNR,QACAQ,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU,QAAQN,kBAAkB;KAEjE;IAGH,IAAIS,AAAgB,MAAhBA,KAAK,MAAM,EACb,OAAO;QACLK,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS;QACnCS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU;QACpCQ,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGf,QAAS;QACnCS,KAAK,KAAK,CAAEM,QAAQ,CAAC,EAAE,GAAGd,SAAU;KACrC;IAGH,MAAMI,MAAM,CAAC,0CAA0C,EAAEC,KAAK,SAAS,CAACF,MAAM,CAAC,CAAC;IAChF,MAAM,IAAIG,MAAMF;AAClB;AAEO,SAASH,UACdE,IAAc,EACdJ,KAAa,EACbC,MAAc,EACdT,gBAAmC;IAEnC,IACE2B,AAAmC,oBAAnCA,aAAa3B,qBACb2B,AAAmC,kBAAnCA,aAAa3B,mBAEb,OAAOkB,gBAAgBN,MAAMJ,OAAOC;IAGtC,IAAIkB,AAAmC,aAAnCA,aAAa3B,mBACf,OAAO4B,gBAAgBhB,MAAMJ,OAAOC;IAGtC,OAAOE,cAAcC;AACvB;AAEO,SAASgB,gBACdhB,IAAc,EACdJ,KAAa,EACbC,MAAc;IAEd,MAAMoB,OAAOZ,KAAK,KAAK,CAAEL,IAAI,CAAC,EAAE,GAAGJ,QAAS;IAC5C,MAAMsB,MAAMb,KAAK,KAAK,CAAEL,IAAI,CAAC,EAAE,GAAGH,SAAU;IAC5C,MAAMsB,QAAQd,KAAK,KAAK,CAAEL,IAAI,CAAC,EAAE,GAAGJ,QAAS;IAC7C,MAAMwB,SAASf,KAAK,KAAK,CAAEL,IAAI,CAAC,EAAE,GAAGH,SAAU;IAC/C,OAAO;QAACoB;QAAMC;QAAKC;QAAOC;KAAO;AACnC;AAEO,SAASC,gBACdrB,IAAc,EACdJ,KAAa,EACbC,MAAc,EACdT,gBAAmC,EACnCkC,UAAU,CAAC,EACXC,UAAU,CAAC;IAEX/B,kBAAkB,mBAAmBQ,MAAMJ,OAAOC,QAAQyB,SAASC;IACnE,MAAM,CAACN,MAAMC,KAAKC,OAAOC,OAAO,GAAGtB,UACjCE,MACAJ,OACAC,QACAT;IAEF,MAAMoC,OAAO;QACX,MAAMP,OAAOK;QACb,KAAKJ,MAAMK;QACX,OAAOJ,QAAQF;QACf,QAAQG,SAASF;IACnB;IACA1B,kBAAkB,4BAA4BgC;IAC9C,OAAOA;AACT;AAEA,IAAIC,SAAS;AACN,SAASC,mBACdC,IAAU,EACVvC,gBAAmC;QAG/BwC;IADJ,IAAIH,QAAQ;IACZ,IAAI,QAAAG,CAAAA,gBAAAA,aAAaxC,iBAAgB,IAA7BwC,KAAAA,IAAAA,cAAgC,WAAW,GAAG,QAAQ,CAAC,WAAW;QACpE,MAAMC,aAAa,CAAC,uEAAuE,EAAEF,KAAK,KAAK,CAAC,CAAC,EAAEA,KAAK,MAAM,CAAC,wFAAwF,CAAC;QAEhN,IACEtB,KAAK,GAAG,CAACsB,KAAK,KAAK,EAAEA,KAAK,MAAM,IAAI,QACpCtB,KAAK,GAAG,CAACsB,KAAK,KAAK,EAAEA,KAAK,MAAM,IAAI,KACpC;YACAG,QAAQ,IAAI,CAACD;YACbJ,SAAS;QACX;IACF,OAAO,IAAIE,KAAK,KAAK,GAAG,QAAQA,KAAK,MAAM,GAAG,MAAM;QAClDG,QAAQ,IAAI,CACV,CAAC,gCAAgC,EAAEH,KAAK,KAAK,CAAC,CAAC,EAAEA,KAAK,MAAM,CAAC,2EAA2E,CAAC;QAE3IF,SAAS;IACX;AACF;AAEO,SAASM,WAAWC,KAAa;IACtC,MAAMC,UAAU5B,KAAK,GAAG,IAAI2B,MAAM,GAAG,CAAC,CAACE,IAAMA,EAAE,IAAI;IACnD,MAAMC,SAAS9B,KAAK,GAAG,IAAI2B,MAAM,GAAG,CAAC,CAACE,IAAMA,EAAE,GAAG;IACjD,MAAME,WAAW/B,KAAK,GAAG,IAAI2B,MAAM,GAAG,CAAC,CAACE,IAAMA,EAAE,IAAI,GAAGA,EAAE,KAAK;IAC9D,MAAMG,YAAYhC,KAAK,GAAG,IAAI2B,MAAM,GAAG,CAAC,CAACE,IAAMA,EAAE,GAAG,GAAGA,EAAE,MAAM;IAC/D,OAAO;QACL,MAAMD;QACN,KAAKE;QACL,OAAOC,WAAWH;QAClB,QAAQI,YAAYF;IACtB;AACF;AAGO,SAASG,iBACdd,IAAU,EACVe,UAAgB,EAChBnD,gBAAmC;IAEnC,MAAMoD,cACJzB,AAAmC,oBAAnCA,aAAa3B,oBAAwC,MAAM;IAC7D,MAAMqD,iBAAiB;IAEvB,MAAMC,wBACJlB,KAAK,KAAK,GAAGgB,cACTnC,KAAK,IAAI,CAAEmC,AAAAA,CAAAA,cAAchB,KAAK,KAAI,IAAK,KACvCiB;IACN,MAAME,sBACJnB,KAAK,MAAM,GAAGgB,cACVnC,KAAK,IAAI,CAAEmC,AAAAA,CAAAA,cAAchB,KAAK,MAAK,IAAK,KACxCiB;IACNjB,KAAK,IAAI,GAAGnB,KAAK,GAAG,CAAC,GAAGmB,KAAK,IAAI,GAAGkB;IACpClB,KAAK,KAAK,GAAGnB,KAAK,GAAG,CACnBmB,KAAK,KAAK,GAAGkB,AAAwB,IAAxBA,uBACbH,WAAW,KAAK,GAAGf,KAAK,IAAI;IAE9BA,KAAK,GAAG,GAAGnB,KAAK,GAAG,CAAC,GAAGmB,KAAK,GAAG,GAAGmB;IAClCnB,KAAK,MAAM,GAAGnB,KAAK,GAAG,CACpBmB,KAAK,MAAM,GAAGmB,AAAsB,IAAtBA,qBACdJ,WAAW,MAAM,GAAGf,KAAK,GAAG;IAE9B,OAAOA;AACT;AAEO,eAAeoB,kBACpBC,gBAAwB,EACxBC,IAAkC,EAClCnB,IAAU;IAEV,MAAMoB,eAAeC,WAAWF;IAChC,MAAMG,kCAAkCF,aAAc,MAAM,CAC1D,CAACG;QACC,IAAIA,YAAY,UAAU,CAAC,QAAQ,KAAKC,SAAS,IAAI,EACnD,OAAO;QAET,OAAO;IACT;IAGF,MAAMC,eAAe,MAAMC,wBAAwB;QACjD,gBAAgBR;QAChB,sBAAsBI;QACtBtB;IACF;IACA,OAAOyB;AACT;AAEO,SAASE,uBACdC,KAAuB,EACvBC,WAAgC,EAChCC,KAAc;IAEd,MAAMC,OAA+B,EAAE;IAEvC,KAAK,MAAMC,QAAQJ,MAAO;QACxB,MAAMK,OAAOD,KAAK,IAAI;QAEtB,MAAME,SAASL,YAAY,IAAI,CAAC,CAACK,SAAWA,OAAO,IAAI,KAAKD;QAC5D,IAAI,CAACC,QAAQ;YACX/B,QAAQ,IAAI,CACV,CAAC,sBAAsB,EAAE8B,KAAK,8BAA8B,CAAC;YAE/D;QACF;QAEA,MAAME,UAAUD,OAAO,cAAc,IAAID;QACzC,MAAMG,YAAYF,OAAO,WAAW,GAChCG,gBAAgBL,KAAK,KAAK,IAAI,CAAC,GAAGE,OAAO,WAAW,IACpD,CAAC;QAEL,MAAMI,WAAiC;YACrC,CAACH,QAAQ,EAAE;YACX,GAAGC,SAAS;QACd;QAEAL,KAAK,IAAI,CAACO;IACZ;IAEA,IAAIR,OACFC,KAAK,IAAI,CAAC;QACRD;IACF;IAGF,OAAOC;AACT;AAGO,MAAMQ,cAAcC,EAAE,MAAM,CAAC;IAClC,MAAMA,EAAE,MAAM;IACd,KAAKA,EAAE,MAAM;AACf;AAEO,MAAMC,aAAaD,EAAE,MAAM,CAAC;IACjC,OAAOA,EAAE,MAAM;IACf,QAAQA,EAAE,MAAM;IAChB,KAAKA,EAAE,MAAM,GAAG,QAAQ;AAC1B;AAEO,MAAME,aAAaH,YAAY,GAAG,CAACE,YAAY,GAAG,CACvDD,EAAE,MAAM,CAAC;IACP,MAAMA,EAAE,MAAM,GAAG,QAAQ;AAC3B;AAIK,MAAMG,0BAA0BH,EAAE,MAAM,CAAC;IAC9C,QAAQA,EAAAA,KACA,CACJA,EAAE,MAAM,CAAC;QACP,MAAMA,EAAE,MAAM;QACd,KAAKA,EAAE,MAAM;IACf,IAED,QAAQ;IACX,yBAAyBA,EAAE,OAAO,GAAG,QAAQ;AAC/C;AAGO,MAAMI,oBAAoBJ,EAAE,KAAK,CAAC;IACvCA,EAAE,MAAM;IACRA,EAAAA,MACS,CAAC;QACN,QAAQA,EAAE,MAAM;IAClB,GACC,GAAG,CAACG,wBAAwB,OAAO;CACvC;AAMD,MAAME,sBAAsB;AAE5B,MAAMC,yBAAyBN,EAAAA,MACtB,CAAC;IACN,CAACK,oBAAoB,EAAEL,EAAE,OAAO,CAAC;IACjC,QAAQI;IAGR,WAAWJ,EAAE,OAAO,GAAG,QAAQ;IAC/B,WAAWA,EAAE,OAAO,GAAG,QAAQ;IAC/B,OAAOA,EAAE,OAAO,GAAG,QAAQ;IAG3B,QAAQA,EAAE,KAAK,CAAC;QAACA,EAAE,MAAM;QAAIA,EAAE,MAAM;KAAG;IACxC,MAAME;AACR,GACC,WAAW;AAGP,MAAMK,4BAA4B,IAChCD;AAGF,MAAME,yBAAyB,CAACC;QAGjCC,mBAKAC;IANJ,IAAIC,cAAcH;IAClB,IAAIC,AAAAA,SAAAA,CAAAA,oBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,kBAAkB,QAAQ,AAAD,MAAM,eACjCE,cAAcA,YAAY,IAAI,CAAC,SAAS;IAI1C,IAAID,AAAAA,SAAAA,CAAAA,qBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,mBAAkB,QAAQ,AAAD,MAAM,aAAa;QAC9C,MAAME,QAAQD,YAAY,IAAI,CAAC,KAAK;QACpC,OAAOP,uBAAuBQ;IAChC;IAEA,OAAO;AACT;AAEO,MAAMC,2BAA2B,CAACL;IACvCrE,OACEoE,uBAAuBC,QACvB;IAIF,IAAI,AAAiB,YAAjB,OAAOA,OACT,OAAOA;IAIT,IAAIA,SAAS,AAAiB,YAAjB,OAAOA,SAAsBA,MAAM,MAAM,EAAE;QAEtD,IAAI,AAAwB,YAAxB,OAAOA,MAAM,MAAM,EACrB,OAAOA,MAAM,MAAM;QAGrB,IAAI,AAAwB,YAAxB,OAAOA,MAAM,MAAM,IAAiBA,MAAM,MAAM,CAAC,MAAM,EACzD,OAAOA,MAAM,MAAM,CAAC,MAAM;IAE9B;IAGA,OAAOM,OAAON;AAChB;AAEO,MAAMO,8BAA8B,CACzCC,SACAC;QAQIC;IANJ,IAAI,CAACF,SACH,OAAO,EAAE;IAIX,MAAMG,YAAYH;IAClB,IAAIE,AAAAA,SAAAA,CAAAA,kBAAAA,UAAU,IAAI,AAAD,IAAbA,KAAAA,IAAAA,gBAAgB,QAAQ,AAAD,MAAM,eAAeC,UAAU,KAAK,EAAE;QAC/D,MAAMC,OAAOC,OAAO,IAAI,CAACF,UAAU,KAAK;QACxC,OAAOC,KAAK,MAAM,CAAC,CAACE;YAClB,MAAMd,QAAQW,UAAU,KAAK,CAACG,IAAI;YAClC,IAAI,CAACf,uBAAuBC,QAC1B,OAAO;YAIT,IAAIS,cAAc;oBACTM;gBAAP,OAAOA,AAAAA,SAAAA,CAAAA,cAAAA,MAAM,IAAI,AAAD,IAATA,KAAAA,IAAAA,YAAY,QAAQ,AAAD,MAAM;YAClC;YAEA,OAAO;QACT;IACF;IAGA,OAAO,EAAE;AACX;AAEO,MAAM3B,kBAAkB,CAC7B3E,YACAuG;IAEA,MAAMC,gBAAgBV,4BAA4BS;IAClD,MAAMxF,SAAS;QAAE,GAAGf,UAAU;IAAC;IAE/B,KAAK,MAAMyG,aAAaD,cAAe;QACrC,MAAME,aAAa3F,MAAM,CAAC0F,UAAU;QACpC,IAAIC,YAEF;YAAA,IAAI,AAAsB,YAAtB,OAAOA,YACT3F,MAAM,CAAC0F,UAAU,GAAGC;iBACf,IAAI,AAAsB,YAAtB,OAAOA,YAEhB;gBAAA,IAAIA,WAAW,MAAM,EAEnB;oBAAA,IAAI,AAA6B,YAA7B,OAAOA,WAAW,MAAM,EAC1B3F,MAAM,CAAC0F,UAAU,GAAGC,WAAW,MAAM;yBAChC,IACL,AAA6B,YAA7B,OAAOA,WAAW,MAAM,IACxBA,WAAW,MAAM,CAAC,MAAM,EAGxB3F,MAAM,CAAC0F,UAAU,GAAGC,WAAW,MAAM,CAAC,MAAM;gBAC9C;YACF;QACF;IAEJ;IAEA,OAAO3F;AACT;AAEO,MAAM4F,kBAAkB,CAC7B3G,YACAuG;IAEA,MAAMC,gBAAgBV,4BAA4BS;IAClD,MAAMxF,SAAS;QAAE,GAAGf,UAAU;IAAC;IAE/B,KAAK,MAAMyG,aAAaD,cAAe;QACrC,MAAME,aAAa3F,MAAM,CAAC0F,UAAU;QACpC,IAAIC,cAAc,AAAsB,YAAtB,OAAOA,YACvB3F,MAAM,CAAC0F,UAAU,GAAG;YAClB,CAACtB,oBAAoB,EAAE;YACvB,QAAQuB;QACV;IAEJ;IAEA,OAAO3F;AACT"}
|
|
@@ -5,6 +5,6 @@ import { generatePlaywrightTest, generatePlaywrightTestStream } from "./prompt/p
|
|
|
5
5
|
import { generateYamlTest, generateYamlTestStream } from "./prompt/yaml-generator.mjs";
|
|
6
6
|
import { AiExtractElementInfo, AiLocateElement, AiLocateSection } from "./inspect.mjs";
|
|
7
7
|
import { plan } from "./llm-planning.mjs";
|
|
8
|
-
import { AIActionType,
|
|
8
|
+
import { AIActionType, PointSchema, RectSchema, SizeSchema, TMultimodalPromptSchema, TUserPromptSchema, adaptBboxToRect, callAiFn, dumpActionParam, findAllMidsceneLocatorField, getMidsceneLocationSchema, loadActionParam } from "./common.mjs";
|
|
9
9
|
import { resizeImageForUiTars, vlmPlanning } from "./ui-tars-planning.mjs";
|
|
10
|
-
export { AIActionType, AiExtractElementInfo, AiLocateElement, AiLocateSection,
|
|
10
|
+
export { AIActionType, AiExtractElementInfo, AiLocateElement, AiLocateSection, PointSchema, RectSchema, SizeSchema, TMultimodalPromptSchema, TUserPromptSchema, adaptBboxToRect, call as callAi, callAiFn, callAiFnWithStringResponse, callToGetJSONObject, describeUserPage, dumpActionParam, elementByPositionWithElementInfo, findAllMidsceneLocatorField, generatePlaywrightTest, generatePlaywrightTestStream, generateYamlTest, generateYamlTestStream, getMidsceneLocationSchema, loadActionParam, plan, resizeImageForUiTars, systemPromptToLocateElement, vlmPlanning };
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"ai-model/inspect.mjs","sources":["webpack://@midscene/core/./src/ai-model/inspect.ts"],"sourcesContent":["import type {\n AIDataExtractionResponse,\n AIElementLocatorResponse,\n AIElementResponse,\n AISectionLocatorResponse,\n AIUsageInfo,\n BaseElement,\n ElementById,\n InsightExtractOption,\n Rect,\n ReferenceImage,\n TMultimodalPrompt,\n TUserPrompt,\n UIContext,\n} from '@/types';\nimport {\n type IModelPreferences,\n getIsUseQwenVl,\n vlLocateMode,\n} from '@midscene/shared/env';\nimport {\n cropByRect,\n paddingToMatchBlockByBase64,\n preProcessImageUrl,\n} from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport type {\n ChatCompletionSystemMessageParam,\n ChatCompletionUserMessageParam,\n} from 'openai/resources/index';\nimport {\n AIActionType,\n adaptBboxToRect,\n callAiFn,\n expandSearchArea,\n markupImageForLLM,\n mergeRects,\n} from './common';\nimport {\n extractDataQueryPrompt,\n systemPromptToExtract,\n} from './prompt/extraction';\nimport {\n findElementPrompt,\n systemPromptToLocateElement,\n} from './prompt/llm-locator';\nimport {\n sectionLocatorInstruction,\n systemPromptToLocateSection,\n} from './prompt/llm-section-locator';\nimport {\n describeUserPage,\n distance,\n distanceThreshold,\n elementByPositionWithElementInfo,\n} from './prompt/util';\nimport { callToGetJSONObject } from './service-caller/index';\n\nexport type AIArgs = [\n ChatCompletionSystemMessageParam,\n ...ChatCompletionUserMessageParam[],\n];\n\nconst debugInspect = getDebug('ai:inspect');\nconst debugSection = getDebug('ai:section');\n\nconst extraTextFromUserPrompt = (prompt: TUserPrompt): string => {\n if (typeof prompt === 'string') {\n return prompt;\n } else {\n return prompt.prompt;\n }\n};\n\nconst promptsToChatParam = async (\n multimodalPrompt: TMultimodalPrompt,\n): Promise<ChatCompletionUserMessageParam[]> => {\n const msgs: ChatCompletionUserMessageParam[] = [];\n if (multimodalPrompt?.images?.length) {\n msgs.push({\n role: 'user',\n content: [\n {\n type: 'text',\n text: 'Next, I will provide all the reference images.',\n },\n ],\n });\n\n for (const item of multimodalPrompt.images) {\n const base64 = await preProcessImageUrl(\n item.url,\n !!multimodalPrompt.convertHttpImage2Base64,\n );\n\n msgs.push({\n role: 'user',\n content: [\n {\n type: 'text',\n text: `reference image ${item.name}:`,\n },\n ],\n });\n\n msgs.push({\n role: 'user',\n content: [\n {\n type: 'image_url',\n image_url: {\n url: base64,\n detail: 'high',\n },\n },\n ],\n });\n }\n }\n return msgs;\n};\n\nexport async function AiLocateElement<\n ElementType extends BaseElement = BaseElement,\n>(options: {\n context: UIContext<ElementType>;\n targetElementDescription: TUserPrompt;\n referenceImage?: ReferenceImage;\n callAI?: typeof callAiFn<AIElementResponse | [number, number]>;\n searchConfig?: Awaited<ReturnType<typeof AiLocateSection>>;\n}): Promise<{\n parseResult: AIElementLocatorResponse;\n rect?: Rect;\n rawResponse: string;\n elementById: ElementById;\n usage?: AIUsageInfo;\n isOrderSensitive?: boolean;\n}> {\n const { context, targetElementDescription, callAI } = options;\n const { screenshotBase64 } = context;\n\n const modelPreferences: IModelPreferences = {\n intent: 'grounding',\n };\n\n const { description, elementById, insertElementByPosition } =\n await describeUserPage(context, modelPreferences);\n\n assert(\n targetElementDescription,\n 'cannot find the target element description',\n );\n\n const userInstructionPrompt = await findElementPrompt.format({\n pageDescription: description,\n targetElementDescription: extraTextFromUserPrompt(targetElementDescription),\n });\n const systemPrompt = systemPromptToLocateElement(\n vlLocateMode(modelPreferences),\n );\n\n let imagePayload = screenshotBase64;\n\n if (options.searchConfig) {\n assert(\n options.searchConfig.rect,\n 'searchArea is provided but its rect cannot be found. Failed to locate element',\n );\n assert(\n options.searchConfig.imageBase64,\n 'searchArea is provided but its imageBase64 cannot be found. Failed to locate element',\n );\n\n imagePayload = options.searchConfig.imageBase64;\n } else if (vlLocateMode(modelPreferences) === 'qwen-vl') {\n imagePayload = await paddingToMatchBlockByBase64(imagePayload);\n } else if (!vlLocateMode(modelPreferences)) {\n imagePayload = await markupImageForLLM(\n screenshotBase64,\n context.tree,\n context.size,\n );\n }\n\n const msgs: AIArgs = [\n { role: 'system', content: systemPrompt },\n {\n role: 'user',\n content: [\n {\n type: 'image_url',\n image_url: {\n url: imagePayload,\n detail: 'high',\n },\n },\n {\n type: 'text',\n text: userInstructionPrompt,\n },\n ],\n },\n ];\n\n if (typeof targetElementDescription !== 'string') {\n const addOns = await promptsToChatParam({\n images: targetElementDescription.images,\n convertHttpImage2Base64: targetElementDescription.convertHttpImage2Base64,\n });\n msgs.push(...addOns);\n }\n\n const callAIFn =\n callAI || callToGetJSONObject<AIElementResponse | [number, number]>;\n\n const res = await callAIFn(msgs, AIActionType.INSPECT_ELEMENT, {\n intent: 'grounding',\n });\n\n const rawResponse = JSON.stringify(res.content);\n\n let resRect: Rect | undefined;\n let matchedElements: AIElementLocatorResponse['elements'] =\n 'elements' in res.content ? res.content.elements : [];\n let errors: AIElementLocatorResponse['errors'] | undefined =\n 'errors' in res.content ? res.content.errors : [];\n try {\n if ('bbox' in res.content && Array.isArray(res.content.bbox)) {\n resRect = adaptBboxToRect(\n res.content.bbox,\n options.searchConfig?.rect?.width || context.size.width,\n options.searchConfig?.rect?.height || context.size.height,\n modelPreferences,\n options.searchConfig?.rect?.left,\n options.searchConfig?.rect?.top,\n );\n debugInspect('resRect', resRect);\n\n const rectCenter = {\n x: resRect.left + resRect.width / 2,\n y: resRect.top + resRect.height / 2,\n };\n let element = elementByPositionWithElementInfo(context.tree, rectCenter);\n\n const distanceToCenter = element\n ? distance({ x: element.center[0], y: element.center[1] }, rectCenter)\n : 0;\n\n if (!element || distanceToCenter > distanceThreshold) {\n element = insertElementByPosition(rectCenter);\n }\n\n if (element) {\n matchedElements = [element];\n errors = [];\n }\n }\n } catch (e) {\n const msg =\n e instanceof Error\n ? `Failed to parse bbox: ${e.message}`\n : 'unknown error in locate';\n if (!errors || errors?.length === 0) {\n errors = [msg];\n } else {\n errors.push(`(${msg})`);\n }\n }\n\n return {\n rect: resRect,\n parseResult: {\n elements: matchedElements,\n errors,\n },\n rawResponse,\n elementById,\n usage: res.usage,\n isOrderSensitive:\n typeof res.content === 'object' &&\n res.content !== null &&\n 'isOrderSensitive' in res.content\n ? (res.content as any).isOrderSensitive\n : undefined,\n };\n}\n\nexport async function AiLocateSection(options: {\n context: UIContext<BaseElement>;\n sectionDescription: TUserPrompt;\n callAI?: typeof callAiFn<AISectionLocatorResponse>;\n}): Promise<{\n rect?: Rect;\n imageBase64?: string;\n error?: string;\n rawResponse: string;\n usage?: AIUsageInfo;\n}> {\n const { context, sectionDescription } = options;\n const { screenshotBase64 } = context;\n\n const modelPreferences: IModelPreferences = {\n intent: 'grounding',\n };\n\n const systemPrompt = systemPromptToLocateSection(\n vlLocateMode(modelPreferences),\n );\n const sectionLocatorInstructionText = await sectionLocatorInstruction.format({\n sectionDescription: extraTextFromUserPrompt(sectionDescription),\n });\n const msgs: AIArgs = [\n { role: 'system', content: systemPrompt },\n {\n role: 'user',\n content: [\n {\n type: 'image_url',\n image_url: {\n url: screenshotBase64,\n detail: 'high',\n },\n },\n {\n type: 'text',\n text: sectionLocatorInstructionText,\n },\n ],\n },\n ];\n\n if (typeof sectionDescription !== 'string') {\n const addOns = await promptsToChatParam({\n images: sectionDescription.images,\n convertHttpImage2Base64: sectionDescription.convertHttpImage2Base64,\n });\n msgs.push(...addOns);\n }\n\n const result = await callAiFn<AISectionLocatorResponse>(\n msgs,\n AIActionType.EXTRACT_DATA,\n {\n intent: 'grounding',\n },\n );\n\n let sectionRect: Rect | undefined;\n const sectionBbox = result.content.bbox;\n if (sectionBbox) {\n const targetRect = adaptBboxToRect(\n sectionBbox,\n context.size.width,\n context.size.height,\n modelPreferences,\n );\n debugSection('original targetRect %j', targetRect);\n\n const referenceBboxList = result.content.references_bbox || [];\n debugSection('referenceBboxList %j', referenceBboxList);\n\n const referenceRects = referenceBboxList\n .filter((bbox) => Array.isArray(bbox))\n .map((bbox) => {\n return adaptBboxToRect(\n bbox,\n context.size.width,\n context.size.height,\n modelPreferences,\n );\n });\n debugSection('referenceRects %j', referenceRects);\n\n // merge the sectionRect and referenceRects\n const mergedRect = mergeRects([targetRect, ...referenceRects]);\n debugSection('mergedRect %j', mergedRect);\n\n // expand search area to at least 200 x 200\n sectionRect = expandSearchArea(mergedRect, context.size, modelPreferences);\n debugSection('expanded sectionRect %j', sectionRect);\n }\n\n let imageBase64 = screenshotBase64;\n if (sectionRect) {\n imageBase64 = await cropByRect(\n screenshotBase64,\n sectionRect,\n getIsUseQwenVl({\n intent: 'grounding',\n }),\n );\n }\n\n return {\n rect: sectionRect,\n imageBase64,\n error: result.content.error,\n rawResponse: JSON.stringify(result.content),\n usage: result.usage,\n };\n}\n\nexport async function AiExtractElementInfo<\n T,\n ElementType extends BaseElement = BaseElement,\n>(options: {\n dataQuery: string | Record<string, string>;\n multimodalPrompt?: TMultimodalPrompt;\n context: UIContext<ElementType>;\n extractOption?: InsightExtractOption;\n modelPreferences: IModelPreferences;\n}) {\n const {\n dataQuery,\n context,\n extractOption,\n multimodalPrompt,\n modelPreferences,\n } = options;\n const systemPrompt = systemPromptToExtract();\n\n const { screenshotBase64 } = context;\n const { description, elementById } = await describeUserPage(\n context,\n modelPreferences,\n {\n truncateTextLength: 200,\n filterNonTextContent: false,\n visibleOnly: false,\n domIncluded: extractOption?.domIncluded,\n },\n );\n\n const extractDataPromptText = await extractDataQueryPrompt(\n description,\n dataQuery,\n );\n\n const userContent: ChatCompletionUserMessageParam['content'] = [];\n\n if (extractOption?.screenshotIncluded !== false) {\n userContent.push({\n type: 'image_url',\n image_url: {\n url: screenshotBase64,\n detail: 'high',\n },\n });\n }\n\n userContent.push({\n type: 'text',\n text: extractDataPromptText,\n });\n\n const msgs: AIArgs = [\n { role: 'system', content: systemPrompt },\n {\n role: 'user',\n content: userContent,\n },\n ];\n\n if (options.extractOption?.returnThought) {\n msgs.push({\n role: 'user',\n content: 'Please provide reasons.',\n });\n }\n\n if (multimodalPrompt) {\n const addOns = await promptsToChatParam({\n images: multimodalPrompt.images,\n convertHttpImage2Base64: multimodalPrompt.convertHttpImage2Base64,\n });\n msgs.push(...addOns);\n }\n\n const result = await callAiFn<AIDataExtractionResponse<T>>(\n msgs,\n AIActionType.EXTRACT_DATA,\n modelPreferences,\n );\n return {\n parseResult: result.content,\n elementById,\n usage: result.usage,\n };\n}\n"],"names":["debugInspect","getDebug","debugSection","extraTextFromUserPrompt","prompt","promptsToChatParam","multimodalPrompt","_multimodalPrompt_images","msgs","item","base64","preProcessImageUrl","AiLocateElement","options","context","targetElementDescription","callAI","screenshotBase64","modelPreferences","description","elementById","insertElementByPosition","describeUserPage","assert","userInstructionPrompt","findElementPrompt","systemPrompt","systemPromptToLocateElement","vlLocateMode","imagePayload","paddingToMatchBlockByBase64","markupImageForLLM","addOns","callAIFn","callToGetJSONObject","res","AIActionType","rawResponse","JSON","resRect","matchedElements","errors","Array","_options_searchConfig_rect","_options_searchConfig_rect1","_options_searchConfig_rect2","_options_searchConfig_rect3","adaptBboxToRect","rectCenter","element","elementByPositionWithElementInfo","distanceToCenter","distance","distanceThreshold","e","msg","Error","undefined","AiLocateSection","sectionDescription","systemPromptToLocateSection","sectionLocatorInstructionText","sectionLocatorInstruction","result","callAiFn","sectionRect","sectionBbox","targetRect","referenceBboxList","referenceRects","bbox","mergedRect","mergeRects","expandSearchArea","imageBase64","cropByRect","getIsUseQwenVl","AiExtractElementInfo","_options_extractOption","dataQuery","extractOption","systemPromptToExtract","extractDataPromptText","extractDataQueryPrompt","userContent"],"mappings":";;;;;;;;;;AAgEA,MAAMA,eAAeC,SAAS;AAC9B,MAAMC,eAAeD,SAAS;AAE9B,MAAME,0BAA0B,CAACC;IAC/B,IAAI,AAAkB,YAAlB,OAAOA,QACT,OAAOA;IAEP,OAAOA,OAAO,MAAM;AAExB;AAEA,MAAMC,qBAAqB,OACzBC;QAGIC;IADJ,MAAMC,OAAyC,EAAE;IACjD,IAAID,QAAAA,mBAAAA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,iBAAkB,MAAM,AAAD,IAAvBA,KAAAA,IAAAA,yBAA0B,MAAM,EAAE;QACpCC,KAAK,IAAI,CAAC;YACR,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM;gBACR;aACD;QACH;QAEA,KAAK,MAAMC,QAAQH,iBAAiB,MAAM,CAAE;YAC1C,MAAMI,SAAS,MAAMC,mBACnBF,KAAK,GAAG,EACR,CAAC,CAACH,iBAAiB,uBAAuB;YAG5CE,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,MAAM,CAAC,gBAAgB,EAAEC,KAAK,IAAI,CAAC,CAAC,CAAC;oBACvC;iBACD;YACH;YAEAD,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,WAAW;4BACT,KAAKE;4BACL,QAAQ;wBACV;oBACF;iBACD;YACH;QACF;IACF;IACA,OAAOF;AACT;AAEO,eAAeI,gBAEpBC,OAMD;IAQC,MAAM,EAAEC,OAAO,EAAEC,wBAAwB,EAAEC,MAAM,EAAE,GAAGH;IACtD,MAAM,EAAEI,gBAAgB,EAAE,GAAGH;IAE7B,MAAMI,mBAAsC;QAC1C,QAAQ;IACV;IAEA,MAAM,EAAEC,WAAW,EAAEC,WAAW,EAAEC,uBAAuB,EAAE,GACzD,MAAMC,iBAAiBR,SAASI;IAElCK,OACER,0BACA;IAGF,MAAMS,wBAAwB,MAAMC,kBAAkB,MAAM,CAAC;QAC3D,iBAAiBN;QACjB,0BAA0BhB,wBAAwBY;IACpD;IACA,MAAMW,eAAeC,4BACnBC,aAAaV;IAGf,IAAIW,eAAeZ;IAEnB,IAAIJ,QAAQ,YAAY,EAAE;QACxBU,OACEV,QAAQ,YAAY,CAAC,IAAI,EACzB;QAEFU,OACEV,QAAQ,YAAY,CAAC,WAAW,EAChC;QAGFgB,eAAehB,QAAQ,YAAY,CAAC,WAAW;IACjD,OAAO,IAAIe,AAAmC,cAAnCA,aAAaV,mBACtBW,eAAe,MAAMC,4BAA4BD;SAC5C,IAAI,CAACD,aAAaV,mBACvBW,eAAe,MAAME,kBACnBd,kBACAH,QAAQ,IAAI,EACZA,QAAQ,IAAI;IAIhB,MAAMN,OAAe;QACnB;YAAE,MAAM;YAAU,SAASkB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKG;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAML;gBACR;aACD;QACH;KACD;IAED,IAAI,AAAoC,YAApC,OAAOT,0BAAuC;QAChD,MAAMiB,SAAS,MAAM3B,mBAAmB;YACtC,QAAQU,yBAAyB,MAAM;YACvC,yBAAyBA,yBAAyB,uBAAuB;QAC3E;QACAP,KAAK,IAAI,IAAIwB;IACf;IAEA,MAAMC,WACJjB,UAAUkB;IAEZ,MAAMC,MAAM,MAAMF,SAASzB,MAAM4B,aAAa,eAAe,EAAE;QAC7D,QAAQ;IACV;IAEA,MAAMC,cAAcC,KAAK,SAAS,CAACH,IAAI,OAAO;IAE9C,IAAII;IACJ,IAAIC,kBACF,cAAcL,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,QAAQ,GAAG,EAAE;IACvD,IAAIM,SACF,YAAYN,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,MAAM,GAAG,EAAE;IACnD,IAAI;QACF,IAAI,UAAUA,IAAI,OAAO,IAAIO,MAAM,OAAO,CAACP,IAAI,OAAO,CAAC,IAAI,GAAG;gBAG1DQ,4BAAAA,uBACAC,6BAAAA,wBAEAC,6BAAAA,wBACAC,6BAAAA;YANFP,UAAUQ,gBACRZ,IAAI,OAAO,CAAC,IAAI,EAChBQ,AAAAA,SAAAA,CAAAA,wBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,6BAAAA,sBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,2BAA4B,KAAK,AAAD,KAAK7B,QAAQ,IAAI,CAAC,KAAK,EACvD8B,AAAAA,SAAAA,CAAAA,yBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,uBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,MAAM,AAAD,KAAK9B,QAAQ,IAAI,CAAC,MAAM,EACzDI,kBAAAA,QACA2B,CAAAA,yBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,uBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,IAAI,UAChCC,CAAAA,yBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,uBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,GAAG;YAEjC9C,aAAa,WAAWuC;YAExB,MAAMS,aAAa;gBACjB,GAAGT,QAAQ,IAAI,GAAGA,QAAQ,KAAK,GAAG;gBAClC,GAAGA,QAAQ,GAAG,GAAGA,QAAQ,MAAM,GAAG;YACpC;YACA,IAAIU,UAAUC,iCAAiCpC,QAAQ,IAAI,EAAEkC;YAE7D,MAAMG,mBAAmBF,UACrBG,SAAS;gBAAE,GAAGH,QAAQ,MAAM,CAAC,EAAE;gBAAE,GAAGA,QAAQ,MAAM,CAAC,EAAE;YAAC,GAAGD,cACzD;YAEJ,IAAI,CAACC,WAAWE,mBAAmBE,mBACjCJ,UAAU5B,wBAAwB2B;YAGpC,IAAIC,SAAS;gBACXT,kBAAkB;oBAACS;iBAAQ;gBAC3BR,SAAS,EAAE;YACb;QACF;IACF,EAAE,OAAOa,GAAG;QACV,MAAMC,MACJD,aAAaE,QACT,CAAC,sBAAsB,EAAEF,EAAE,OAAO,EAAE,GACpC;QACN,IAAI,AAACb,UAAUA,AAAAA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM,AAAD,MAAM,GAGhCA,OAAO,IAAI,CAAC,CAAC,CAAC,EAAEc,IAAI,CAAC,CAAC;aAFtBd,SAAS;YAACc;SAAI;IAIlB;IAEA,OAAO;QACL,MAAMhB;QACN,aAAa;YACX,UAAUC;YACVC;QACF;QACAJ;QACAjB;QACA,OAAOe,IAAI,KAAK;QAChB,kBACE,AAAuB,YAAvB,OAAOA,IAAI,OAAO,IAClBA,AAAgB,SAAhBA,IAAI,OAAO,IACX,sBAAsBA,IAAI,OAAO,GAC5BA,IAAI,OAAO,CAAS,gBAAgB,GACrCsB;IACR;AACF;AAEO,eAAeC,gBAAgB7C,OAIrC;IAOC,MAAM,EAAEC,OAAO,EAAE6C,kBAAkB,EAAE,GAAG9C;IACxC,MAAM,EAAEI,gBAAgB,EAAE,GAAGH;IAE7B,MAAMI,mBAAsC;QAC1C,QAAQ;IACV;IAEA,MAAMQ,eAAekC,4BACnBhC,aAAaV;IAEf,MAAM2C,gCAAgC,MAAMC,0BAA0B,MAAM,CAAC;QAC3E,oBAAoB3D,wBAAwBwD;IAC9C;IACA,MAAMnD,OAAe;QACnB;YAAE,MAAM;YAAU,SAASkB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKT;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAM4C;gBACR;aACD;QACH;KACD;IAED,IAAI,AAA8B,YAA9B,OAAOF,oBAAiC;QAC1C,MAAM3B,SAAS,MAAM3B,mBAAmB;YACtC,QAAQsD,mBAAmB,MAAM;YACjC,yBAAyBA,mBAAmB,uBAAuB;QACrE;QACAnD,KAAK,IAAI,IAAIwB;IACf;IAEA,MAAM+B,SAAS,MAAMC,SACnBxD,MACA4B,aAAa,YAAY,EACzB;QACE,QAAQ;IACV;IAGF,IAAI6B;IACJ,MAAMC,cAAcH,OAAO,OAAO,CAAC,IAAI;IACvC,IAAIG,aAAa;QACf,MAAMC,aAAapB,gBACjBmB,aACApD,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAEFhB,aAAa,0BAA0BiE;QAEvC,MAAMC,oBAAoBL,OAAO,OAAO,CAAC,eAAe,IAAI,EAAE;QAC9D7D,aAAa,wBAAwBkE;QAErC,MAAMC,iBAAiBD,kBACpB,MAAM,CAAC,CAACE,OAAS5B,MAAM,OAAO,CAAC4B,OAC/B,GAAG,CAAC,CAACA,OACGvB,gBACLuB,MACAxD,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAGNhB,aAAa,qBAAqBmE;QAGlC,MAAME,aAAaC,WAAW;YAACL;eAAeE;SAAe;QAC7DnE,aAAa,iBAAiBqE;QAG9BN,cAAcQ,iBAAiBF,YAAYzD,QAAQ,IAAI,EAAEI;QACzDhB,aAAa,2BAA2B+D;IAC1C;IAEA,IAAIS,cAAczD;IAClB,IAAIgD,aACFS,cAAc,MAAMC,WAClB1D,kBACAgD,aACAW,eAAe;QACb,QAAQ;IACV;IAIJ,OAAO;QACL,MAAMX;QACNS;QACA,OAAOX,OAAO,OAAO,CAAC,KAAK;QAC3B,aAAazB,KAAK,SAAS,CAACyB,OAAO,OAAO;QAC1C,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAec,qBAGpBhE,OAMD;QAoDKiE;IAnDJ,MAAM,EACJC,SAAS,EACTjE,OAAO,EACPkE,aAAa,EACb1E,gBAAgB,EAChBY,gBAAgB,EACjB,GAAGL;IACJ,MAAMa,eAAeuD;IAErB,MAAM,EAAEhE,gBAAgB,EAAE,GAAGH;IAC7B,MAAM,EAAEK,WAAW,EAAEC,WAAW,EAAE,GAAG,MAAME,iBACzCR,SACAI,kBACA;QACE,oBAAoB;QACpB,sBAAsB;QACtB,aAAa;QACb,aAAa8D,QAAAA,gBAAAA,KAAAA,IAAAA,cAAe,WAAW;IACzC;IAGF,MAAME,wBAAwB,MAAMC,uBAClChE,aACA4D;IAGF,MAAMK,cAAyD,EAAE;IAEjE,IAAIJ,AAAAA,CAAAA,QAAAA,gBAAAA,KAAAA,IAAAA,cAAe,kBAAkB,AAAD,MAAM,OACxCI,YAAY,IAAI,CAAC;QACf,MAAM;QACN,WAAW;YACT,KAAKnE;YACL,QAAQ;QACV;IACF;IAGFmE,YAAY,IAAI,CAAC;QACf,MAAM;QACN,MAAMF;IACR;IAEA,MAAM1E,OAAe;QACnB;YAAE,MAAM;YAAU,SAASkB;QAAa;QACxC;YACE,MAAM;YACN,SAAS0D;QACX;KACD;IAED,IAAI,QAAAN,CAAAA,yBAAAA,QAAQ,aAAa,AAAD,IAApBA,KAAAA,IAAAA,uBAAuB,aAAa,EACtCtE,KAAK,IAAI,CAAC;QACR,MAAM;QACN,SAAS;IACX;IAGF,IAAIF,kBAAkB;QACpB,MAAM0B,SAAS,MAAM3B,mBAAmB;YACtC,QAAQC,iBAAiB,MAAM;YAC/B,yBAAyBA,iBAAiB,uBAAuB;QACnE;QACAE,KAAK,IAAI,IAAIwB;IACf;IAEA,MAAM+B,SAAS,MAAMC,SACnBxD,MACA4B,aAAa,YAAY,EACzBlB;IAEF,OAAO;QACL,aAAa6C,OAAO,OAAO;QAC3B3C;QACA,OAAO2C,OAAO,KAAK;IACrB;AACF"}
|
|
1
|
+
{"version":3,"file":"ai-model/inspect.mjs","sources":["webpack://@midscene/core/./src/ai-model/inspect.ts"],"sourcesContent":["import type {\n AIDataExtractionResponse,\n AIElementLocatorResponse,\n AIElementResponse,\n AISectionLocatorResponse,\n AIUsageInfo,\n BaseElement,\n ElementById,\n InsightExtractOption,\n Rect,\n ReferenceImage,\n UIContext,\n} from '@/types';\nimport {\n type IModelPreferences,\n getIsUseQwenVl,\n vlLocateMode,\n} from '@midscene/shared/env';\nimport {\n cropByRect,\n paddingToMatchBlockByBase64,\n preProcessImageUrl,\n} from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport type {\n ChatCompletionSystemMessageParam,\n ChatCompletionUserMessageParam,\n} from 'openai/resources/index';\nimport type { TMultimodalPrompt, TUserPrompt } from './common';\nimport {\n AIActionType,\n adaptBboxToRect,\n callAiFn,\n expandSearchArea,\n markupImageForLLM,\n mergeRects,\n} from './common';\nimport {\n extractDataQueryPrompt,\n systemPromptToExtract,\n} from './prompt/extraction';\nimport {\n findElementPrompt,\n systemPromptToLocateElement,\n} from './prompt/llm-locator';\nimport {\n sectionLocatorInstruction,\n systemPromptToLocateSection,\n} from './prompt/llm-section-locator';\nimport {\n describeUserPage,\n distance,\n distanceThreshold,\n elementByPositionWithElementInfo,\n} from './prompt/util';\nimport { callToGetJSONObject } from './service-caller/index';\n\nexport type AIArgs = [\n ChatCompletionSystemMessageParam,\n ...ChatCompletionUserMessageParam[],\n];\n\nconst debugInspect = getDebug('ai:inspect');\nconst debugSection = getDebug('ai:section');\n\nconst extraTextFromUserPrompt = (prompt: TUserPrompt): string => {\n if (typeof prompt === 'string') {\n return prompt;\n } else {\n return prompt.prompt;\n }\n};\n\nconst promptsToChatParam = async (\n multimodalPrompt: TMultimodalPrompt,\n): Promise<ChatCompletionUserMessageParam[]> => {\n const msgs: ChatCompletionUserMessageParam[] = [];\n if (multimodalPrompt?.images?.length) {\n msgs.push({\n role: 'user',\n content: [\n {\n type: 'text',\n text: 'Next, I will provide all the reference images.',\n },\n ],\n });\n\n for (const item of multimodalPrompt.images) {\n const base64 = await preProcessImageUrl(\n item.url,\n !!multimodalPrompt.convertHttpImage2Base64,\n );\n\n msgs.push({\n role: 'user',\n content: [\n {\n type: 'text',\n text: `reference image ${item.name}:`,\n },\n ],\n });\n\n msgs.push({\n role: 'user',\n content: [\n {\n type: 'image_url',\n image_url: {\n url: base64,\n detail: 'high',\n },\n },\n ],\n });\n }\n }\n return msgs;\n};\n\nexport async function AiLocateElement<\n ElementType extends BaseElement = BaseElement,\n>(options: {\n context: UIContext<ElementType>;\n targetElementDescription: TUserPrompt;\n referenceImage?: ReferenceImage;\n callAI?: typeof callAiFn<AIElementResponse | [number, number]>;\n searchConfig?: Awaited<ReturnType<typeof AiLocateSection>>;\n}): Promise<{\n parseResult: AIElementLocatorResponse;\n rect?: Rect;\n rawResponse: string;\n elementById: ElementById;\n usage?: AIUsageInfo;\n isOrderSensitive?: boolean;\n}> {\n const { context, targetElementDescription, callAI } = options;\n const { screenshotBase64 } = context;\n\n const modelPreferences: IModelPreferences = {\n intent: 'grounding',\n };\n\n const { description, elementById, insertElementByPosition } =\n await describeUserPage(context, modelPreferences);\n\n assert(\n targetElementDescription,\n 'cannot find the target element description',\n );\n\n const userInstructionPrompt = await findElementPrompt.format({\n pageDescription: description,\n targetElementDescription: extraTextFromUserPrompt(targetElementDescription),\n });\n const systemPrompt = systemPromptToLocateElement(\n vlLocateMode(modelPreferences),\n );\n\n let imagePayload = screenshotBase64;\n\n if (options.searchConfig) {\n assert(\n options.searchConfig.rect,\n 'searchArea is provided but its rect cannot be found. Failed to locate element',\n );\n assert(\n options.searchConfig.imageBase64,\n 'searchArea is provided but its imageBase64 cannot be found. Failed to locate element',\n );\n\n imagePayload = options.searchConfig.imageBase64;\n } else if (vlLocateMode(modelPreferences) === 'qwen-vl') {\n imagePayload = await paddingToMatchBlockByBase64(imagePayload);\n } else if (!vlLocateMode(modelPreferences)) {\n imagePayload = await markupImageForLLM(\n screenshotBase64,\n context.tree,\n context.size,\n );\n }\n\n const msgs: AIArgs = [\n { role: 'system', content: systemPrompt },\n {\n role: 'user',\n content: [\n {\n type: 'image_url',\n image_url: {\n url: imagePayload,\n detail: 'high',\n },\n },\n {\n type: 'text',\n text: userInstructionPrompt,\n },\n ],\n },\n ];\n\n if (typeof targetElementDescription !== 'string') {\n const addOns = await promptsToChatParam({\n images: targetElementDescription.images,\n convertHttpImage2Base64: targetElementDescription.convertHttpImage2Base64,\n });\n msgs.push(...addOns);\n }\n\n const callAIFn =\n callAI || callToGetJSONObject<AIElementResponse | [number, number]>;\n\n const res = await callAIFn(msgs, AIActionType.INSPECT_ELEMENT, {\n intent: 'grounding',\n });\n\n const rawResponse = JSON.stringify(res.content);\n\n let resRect: Rect | undefined;\n let matchedElements: AIElementLocatorResponse['elements'] =\n 'elements' in res.content ? res.content.elements : [];\n let errors: AIElementLocatorResponse['errors'] | undefined =\n 'errors' in res.content ? res.content.errors : [];\n try {\n if ('bbox' in res.content && Array.isArray(res.content.bbox)) {\n resRect = adaptBboxToRect(\n res.content.bbox,\n options.searchConfig?.rect?.width || context.size.width,\n options.searchConfig?.rect?.height || context.size.height,\n modelPreferences,\n options.searchConfig?.rect?.left,\n options.searchConfig?.rect?.top,\n );\n debugInspect('resRect', resRect);\n\n const rectCenter = {\n x: resRect.left + resRect.width / 2,\n y: resRect.top + resRect.height / 2,\n };\n let element = elementByPositionWithElementInfo(context.tree, rectCenter);\n\n const distanceToCenter = element\n ? distance({ x: element.center[0], y: element.center[1] }, rectCenter)\n : 0;\n\n if (!element || distanceToCenter > distanceThreshold) {\n element = insertElementByPosition(rectCenter);\n }\n\n if (element) {\n matchedElements = [element];\n errors = [];\n }\n }\n } catch (e) {\n const msg =\n e instanceof Error\n ? `Failed to parse bbox: ${e.message}`\n : 'unknown error in locate';\n if (!errors || errors?.length === 0) {\n errors = [msg];\n } else {\n errors.push(`(${msg})`);\n }\n }\n\n return {\n rect: resRect,\n parseResult: {\n elements: matchedElements,\n errors,\n },\n rawResponse,\n elementById,\n usage: res.usage,\n isOrderSensitive:\n typeof res.content === 'object' &&\n res.content !== null &&\n 'isOrderSensitive' in res.content\n ? (res.content as any).isOrderSensitive\n : undefined,\n };\n}\n\nexport async function AiLocateSection(options: {\n context: UIContext<BaseElement>;\n sectionDescription: TUserPrompt;\n callAI?: typeof callAiFn<AISectionLocatorResponse>;\n}): Promise<{\n rect?: Rect;\n imageBase64?: string;\n error?: string;\n rawResponse: string;\n usage?: AIUsageInfo;\n}> {\n const { context, sectionDescription } = options;\n const { screenshotBase64 } = context;\n\n const modelPreferences: IModelPreferences = {\n intent: 'grounding',\n };\n\n const systemPrompt = systemPromptToLocateSection(\n vlLocateMode(modelPreferences),\n );\n const sectionLocatorInstructionText = await sectionLocatorInstruction.format({\n sectionDescription: extraTextFromUserPrompt(sectionDescription),\n });\n const msgs: AIArgs = [\n { role: 'system', content: systemPrompt },\n {\n role: 'user',\n content: [\n {\n type: 'image_url',\n image_url: {\n url: screenshotBase64,\n detail: 'high',\n },\n },\n {\n type: 'text',\n text: sectionLocatorInstructionText,\n },\n ],\n },\n ];\n\n if (typeof sectionDescription !== 'string') {\n const addOns = await promptsToChatParam({\n images: sectionDescription.images,\n convertHttpImage2Base64: sectionDescription.convertHttpImage2Base64,\n });\n msgs.push(...addOns);\n }\n\n const result = await callAiFn<AISectionLocatorResponse>(\n msgs,\n AIActionType.EXTRACT_DATA,\n {\n intent: 'grounding',\n },\n );\n\n let sectionRect: Rect | undefined;\n const sectionBbox = result.content.bbox;\n if (sectionBbox) {\n const targetRect = adaptBboxToRect(\n sectionBbox,\n context.size.width,\n context.size.height,\n modelPreferences,\n );\n debugSection('original targetRect %j', targetRect);\n\n const referenceBboxList = result.content.references_bbox || [];\n debugSection('referenceBboxList %j', referenceBboxList);\n\n const referenceRects = referenceBboxList\n .filter((bbox) => Array.isArray(bbox))\n .map((bbox) => {\n return adaptBboxToRect(\n bbox,\n context.size.width,\n context.size.height,\n modelPreferences,\n );\n });\n debugSection('referenceRects %j', referenceRects);\n\n // merge the sectionRect and referenceRects\n const mergedRect = mergeRects([targetRect, ...referenceRects]);\n debugSection('mergedRect %j', mergedRect);\n\n // expand search area to at least 200 x 200\n sectionRect = expandSearchArea(mergedRect, context.size, modelPreferences);\n debugSection('expanded sectionRect %j', sectionRect);\n }\n\n let imageBase64 = screenshotBase64;\n if (sectionRect) {\n imageBase64 = await cropByRect(\n screenshotBase64,\n sectionRect,\n getIsUseQwenVl({\n intent: 'grounding',\n }),\n );\n }\n\n return {\n rect: sectionRect,\n imageBase64,\n error: result.content.error,\n rawResponse: JSON.stringify(result.content),\n usage: result.usage,\n };\n}\n\nexport async function AiExtractElementInfo<\n T,\n ElementType extends BaseElement = BaseElement,\n>(options: {\n dataQuery: string | Record<string, string>;\n multimodalPrompt?: TMultimodalPrompt;\n context: UIContext<ElementType>;\n extractOption?: InsightExtractOption;\n modelPreferences: IModelPreferences;\n}) {\n const {\n dataQuery,\n context,\n extractOption,\n multimodalPrompt,\n modelPreferences,\n } = options;\n const systemPrompt = systemPromptToExtract();\n\n const { screenshotBase64 } = context;\n const { description, elementById } = await describeUserPage(\n context,\n modelPreferences,\n {\n truncateTextLength: 200,\n filterNonTextContent: false,\n visibleOnly: false,\n domIncluded: extractOption?.domIncluded,\n },\n );\n\n const extractDataPromptText = await extractDataQueryPrompt(\n description,\n dataQuery,\n );\n\n const userContent: ChatCompletionUserMessageParam['content'] = [];\n\n if (extractOption?.screenshotIncluded !== false) {\n userContent.push({\n type: 'image_url',\n image_url: {\n url: screenshotBase64,\n detail: 'high',\n },\n });\n }\n\n userContent.push({\n type: 'text',\n text: extractDataPromptText,\n });\n\n const msgs: AIArgs = [\n { role: 'system', content: systemPrompt },\n {\n role: 'user',\n content: userContent,\n },\n ];\n\n if (options.extractOption?.returnThought) {\n msgs.push({\n role: 'user',\n content: 'Please provide reasons.',\n });\n }\n\n if (multimodalPrompt) {\n const addOns = await promptsToChatParam({\n images: multimodalPrompt.images,\n convertHttpImage2Base64: multimodalPrompt.convertHttpImage2Base64,\n });\n msgs.push(...addOns);\n }\n\n const result = await callAiFn<AIDataExtractionResponse<T>>(\n msgs,\n AIActionType.EXTRACT_DATA,\n modelPreferences,\n );\n return {\n parseResult: result.content,\n elementById,\n usage: result.usage,\n };\n}\n"],"names":["debugInspect","getDebug","debugSection","extraTextFromUserPrompt","prompt","promptsToChatParam","multimodalPrompt","_multimodalPrompt_images","msgs","item","base64","preProcessImageUrl","AiLocateElement","options","context","targetElementDescription","callAI","screenshotBase64","modelPreferences","description","elementById","insertElementByPosition","describeUserPage","assert","userInstructionPrompt","findElementPrompt","systemPrompt","systemPromptToLocateElement","vlLocateMode","imagePayload","paddingToMatchBlockByBase64","markupImageForLLM","addOns","callAIFn","callToGetJSONObject","res","AIActionType","rawResponse","JSON","resRect","matchedElements","errors","Array","_options_searchConfig_rect","_options_searchConfig_rect1","_options_searchConfig_rect2","_options_searchConfig_rect3","adaptBboxToRect","rectCenter","element","elementByPositionWithElementInfo","distanceToCenter","distance","distanceThreshold","e","msg","Error","undefined","AiLocateSection","sectionDescription","systemPromptToLocateSection","sectionLocatorInstructionText","sectionLocatorInstruction","result","callAiFn","sectionRect","sectionBbox","targetRect","referenceBboxList","referenceRects","bbox","mergedRect","mergeRects","expandSearchArea","imageBase64","cropByRect","getIsUseQwenVl","AiExtractElementInfo","_options_extractOption","dataQuery","extractOption","systemPromptToExtract","extractDataPromptText","extractDataQueryPrompt","userContent"],"mappings":";;;;;;;;;;AA+DA,MAAMA,eAAeC,SAAS;AAC9B,MAAMC,eAAeD,SAAS;AAE9B,MAAME,0BAA0B,CAACC;IAC/B,IAAI,AAAkB,YAAlB,OAAOA,QACT,OAAOA;IAEP,OAAOA,OAAO,MAAM;AAExB;AAEA,MAAMC,qBAAqB,OACzBC;QAGIC;IADJ,MAAMC,OAAyC,EAAE;IACjD,IAAID,QAAAA,mBAAAA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,iBAAkB,MAAM,AAAD,IAAvBA,KAAAA,IAAAA,yBAA0B,MAAM,EAAE;QACpCC,KAAK,IAAI,CAAC;YACR,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM;gBACR;aACD;QACH;QAEA,KAAK,MAAMC,QAAQH,iBAAiB,MAAM,CAAE;YAC1C,MAAMI,SAAS,MAAMC,mBACnBF,KAAK,GAAG,EACR,CAAC,CAACH,iBAAiB,uBAAuB;YAG5CE,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,MAAM,CAAC,gBAAgB,EAAEC,KAAK,IAAI,CAAC,CAAC,CAAC;oBACvC;iBACD;YACH;YAEAD,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,WAAW;4BACT,KAAKE;4BACL,QAAQ;wBACV;oBACF;iBACD;YACH;QACF;IACF;IACA,OAAOF;AACT;AAEO,eAAeI,gBAEpBC,OAMD;IAQC,MAAM,EAAEC,OAAO,EAAEC,wBAAwB,EAAEC,MAAM,EAAE,GAAGH;IACtD,MAAM,EAAEI,gBAAgB,EAAE,GAAGH;IAE7B,MAAMI,mBAAsC;QAC1C,QAAQ;IACV;IAEA,MAAM,EAAEC,WAAW,EAAEC,WAAW,EAAEC,uBAAuB,EAAE,GACzD,MAAMC,iBAAiBR,SAASI;IAElCK,OACER,0BACA;IAGF,MAAMS,wBAAwB,MAAMC,kBAAkB,MAAM,CAAC;QAC3D,iBAAiBN;QACjB,0BAA0BhB,wBAAwBY;IACpD;IACA,MAAMW,eAAeC,4BACnBC,aAAaV;IAGf,IAAIW,eAAeZ;IAEnB,IAAIJ,QAAQ,YAAY,EAAE;QACxBU,OACEV,QAAQ,YAAY,CAAC,IAAI,EACzB;QAEFU,OACEV,QAAQ,YAAY,CAAC,WAAW,EAChC;QAGFgB,eAAehB,QAAQ,YAAY,CAAC,WAAW;IACjD,OAAO,IAAIe,AAAmC,cAAnCA,aAAaV,mBACtBW,eAAe,MAAMC,4BAA4BD;SAC5C,IAAI,CAACD,aAAaV,mBACvBW,eAAe,MAAME,kBACnBd,kBACAH,QAAQ,IAAI,EACZA,QAAQ,IAAI;IAIhB,MAAMN,OAAe;QACnB;YAAE,MAAM;YAAU,SAASkB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKG;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAML;gBACR;aACD;QACH;KACD;IAED,IAAI,AAAoC,YAApC,OAAOT,0BAAuC;QAChD,MAAMiB,SAAS,MAAM3B,mBAAmB;YACtC,QAAQU,yBAAyB,MAAM;YACvC,yBAAyBA,yBAAyB,uBAAuB;QAC3E;QACAP,KAAK,IAAI,IAAIwB;IACf;IAEA,MAAMC,WACJjB,UAAUkB;IAEZ,MAAMC,MAAM,MAAMF,SAASzB,MAAM4B,aAAa,eAAe,EAAE;QAC7D,QAAQ;IACV;IAEA,MAAMC,cAAcC,KAAK,SAAS,CAACH,IAAI,OAAO;IAE9C,IAAII;IACJ,IAAIC,kBACF,cAAcL,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,QAAQ,GAAG,EAAE;IACvD,IAAIM,SACF,YAAYN,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,MAAM,GAAG,EAAE;IACnD,IAAI;QACF,IAAI,UAAUA,IAAI,OAAO,IAAIO,MAAM,OAAO,CAACP,IAAI,OAAO,CAAC,IAAI,GAAG;gBAG1DQ,4BAAAA,uBACAC,6BAAAA,wBAEAC,6BAAAA,wBACAC,6BAAAA;YANFP,UAAUQ,gBACRZ,IAAI,OAAO,CAAC,IAAI,EAChBQ,AAAAA,SAAAA,CAAAA,wBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,6BAAAA,sBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,2BAA4B,KAAK,AAAD,KAAK7B,QAAQ,IAAI,CAAC,KAAK,EACvD8B,AAAAA,SAAAA,CAAAA,yBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,uBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,MAAM,AAAD,KAAK9B,QAAQ,IAAI,CAAC,MAAM,EACzDI,kBAAAA,QACA2B,CAAAA,yBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,uBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,IAAI,UAChCC,CAAAA,yBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,uBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,GAAG;YAEjC9C,aAAa,WAAWuC;YAExB,MAAMS,aAAa;gBACjB,GAAGT,QAAQ,IAAI,GAAGA,QAAQ,KAAK,GAAG;gBAClC,GAAGA,QAAQ,GAAG,GAAGA,QAAQ,MAAM,GAAG;YACpC;YACA,IAAIU,UAAUC,iCAAiCpC,QAAQ,IAAI,EAAEkC;YAE7D,MAAMG,mBAAmBF,UACrBG,SAAS;gBAAE,GAAGH,QAAQ,MAAM,CAAC,EAAE;gBAAE,GAAGA,QAAQ,MAAM,CAAC,EAAE;YAAC,GAAGD,cACzD;YAEJ,IAAI,CAACC,WAAWE,mBAAmBE,mBACjCJ,UAAU5B,wBAAwB2B;YAGpC,IAAIC,SAAS;gBACXT,kBAAkB;oBAACS;iBAAQ;gBAC3BR,SAAS,EAAE;YACb;QACF;IACF,EAAE,OAAOa,GAAG;QACV,MAAMC,MACJD,aAAaE,QACT,CAAC,sBAAsB,EAAEF,EAAE,OAAO,EAAE,GACpC;QACN,IAAI,AAACb,UAAUA,AAAAA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM,AAAD,MAAM,GAGhCA,OAAO,IAAI,CAAC,CAAC,CAAC,EAAEc,IAAI,CAAC,CAAC;aAFtBd,SAAS;YAACc;SAAI;IAIlB;IAEA,OAAO;QACL,MAAMhB;QACN,aAAa;YACX,UAAUC;YACVC;QACF;QACAJ;QACAjB;QACA,OAAOe,IAAI,KAAK;QAChB,kBACE,AAAuB,YAAvB,OAAOA,IAAI,OAAO,IAClBA,AAAgB,SAAhBA,IAAI,OAAO,IACX,sBAAsBA,IAAI,OAAO,GAC5BA,IAAI,OAAO,CAAS,gBAAgB,GACrCsB;IACR;AACF;AAEO,eAAeC,gBAAgB7C,OAIrC;IAOC,MAAM,EAAEC,OAAO,EAAE6C,kBAAkB,EAAE,GAAG9C;IACxC,MAAM,EAAEI,gBAAgB,EAAE,GAAGH;IAE7B,MAAMI,mBAAsC;QAC1C,QAAQ;IACV;IAEA,MAAMQ,eAAekC,4BACnBhC,aAAaV;IAEf,MAAM2C,gCAAgC,MAAMC,0BAA0B,MAAM,CAAC;QAC3E,oBAAoB3D,wBAAwBwD;IAC9C;IACA,MAAMnD,OAAe;QACnB;YAAE,MAAM;YAAU,SAASkB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKT;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAM4C;gBACR;aACD;QACH;KACD;IAED,IAAI,AAA8B,YAA9B,OAAOF,oBAAiC;QAC1C,MAAM3B,SAAS,MAAM3B,mBAAmB;YACtC,QAAQsD,mBAAmB,MAAM;YACjC,yBAAyBA,mBAAmB,uBAAuB;QACrE;QACAnD,KAAK,IAAI,IAAIwB;IACf;IAEA,MAAM+B,SAAS,MAAMC,SACnBxD,MACA4B,aAAa,YAAY,EACzB;QACE,QAAQ;IACV;IAGF,IAAI6B;IACJ,MAAMC,cAAcH,OAAO,OAAO,CAAC,IAAI;IACvC,IAAIG,aAAa;QACf,MAAMC,aAAapB,gBACjBmB,aACApD,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAEFhB,aAAa,0BAA0BiE;QAEvC,MAAMC,oBAAoBL,OAAO,OAAO,CAAC,eAAe,IAAI,EAAE;QAC9D7D,aAAa,wBAAwBkE;QAErC,MAAMC,iBAAiBD,kBACpB,MAAM,CAAC,CAACE,OAAS5B,MAAM,OAAO,CAAC4B,OAC/B,GAAG,CAAC,CAACA,OACGvB,gBACLuB,MACAxD,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAGNhB,aAAa,qBAAqBmE;QAGlC,MAAME,aAAaC,WAAW;YAACL;eAAeE;SAAe;QAC7DnE,aAAa,iBAAiBqE;QAG9BN,cAAcQ,iBAAiBF,YAAYzD,QAAQ,IAAI,EAAEI;QACzDhB,aAAa,2BAA2B+D;IAC1C;IAEA,IAAIS,cAAczD;IAClB,IAAIgD,aACFS,cAAc,MAAMC,WAClB1D,kBACAgD,aACAW,eAAe;QACb,QAAQ;IACV;IAIJ,OAAO;QACL,MAAMX;QACNS;QACA,OAAOX,OAAO,OAAO,CAAC,KAAK;QAC3B,aAAazB,KAAK,SAAS,CAACyB,OAAO,OAAO;QAC1C,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAec,qBAGpBhE,OAMD;QAoDKiE;IAnDJ,MAAM,EACJC,SAAS,EACTjE,OAAO,EACPkE,aAAa,EACb1E,gBAAgB,EAChBY,gBAAgB,EACjB,GAAGL;IACJ,MAAMa,eAAeuD;IAErB,MAAM,EAAEhE,gBAAgB,EAAE,GAAGH;IAC7B,MAAM,EAAEK,WAAW,EAAEC,WAAW,EAAE,GAAG,MAAME,iBACzCR,SACAI,kBACA;QACE,oBAAoB;QACpB,sBAAsB;QACtB,aAAa;QACb,aAAa8D,QAAAA,gBAAAA,KAAAA,IAAAA,cAAe,WAAW;IACzC;IAGF,MAAME,wBAAwB,MAAMC,uBAClChE,aACA4D;IAGF,MAAMK,cAAyD,EAAE;IAEjE,IAAIJ,AAAAA,CAAAA,QAAAA,gBAAAA,KAAAA,IAAAA,cAAe,kBAAkB,AAAD,MAAM,OACxCI,YAAY,IAAI,CAAC;QACf,MAAM;QACN,WAAW;YACT,KAAKnE;YACL,QAAQ;QACV;IACF;IAGFmE,YAAY,IAAI,CAAC;QACf,MAAM;QACN,MAAMF;IACR;IAEA,MAAM1E,OAAe;QACnB;YAAE,MAAM;YAAU,SAASkB;QAAa;QACxC;YACE,MAAM;YACN,SAAS0D;QACX;KACD;IAED,IAAI,QAAAN,CAAAA,yBAAAA,QAAQ,aAAa,AAAD,IAApBA,KAAAA,IAAAA,uBAAuB,aAAa,EACtCtE,KAAK,IAAI,CAAC;QACR,MAAM;QACN,SAAS;IACX;IAGF,IAAIF,kBAAkB;QACpB,MAAM0B,SAAS,MAAM3B,mBAAmB;YACtC,QAAQC,iBAAiB,MAAM;YAC/B,yBAAyBA,iBAAiB,uBAAuB;QACnE;QACAE,KAAK,IAAI,IAAIwB;IACf;IAEA,MAAM+B,SAAS,MAAMC,SACnBxD,MACA4B,aAAa,YAAY,EACzBlB;IAEF,OAAO;QACL,aAAa6C,OAAO,OAAO;QAC3B3C;QACA,OAAO2C,OAAO,KAAK;IACrB;AACF"}
|
|
@@ -76,7 +76,6 @@ async function plan(userInstruction, opts) {
|
|
|
76
76
|
const element = elementById(locateResult);
|
|
77
77
|
if (element) action.param[field].id = element.id;
|
|
78
78
|
}
|
|
79
|
-
action.locate = action.param[field];
|
|
80
79
|
});
|
|
81
80
|
});
|
|
82
81
|
assert(!planFromAI.error, `Failed to plan actions: ${planFromAI.error}`);
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"ai-model/llm-planning.mjs","sources":["webpack://@midscene/core/./src/ai-model/llm-planning.ts"],"sourcesContent":["import type {\n DeviceAction,\n PageType,\n PlanningAIResponse,\n UIContext,\n} from '@/types';\nimport { type IModelPreferences, vlLocateMode } from '@midscene/shared/env';\nimport { paddingToMatchBlockByBase64 } from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport {\n AIActionType,\n type AIArgs,\n buildYamlFlowFromPlans,\n callAiFn,\n fillBboxParam,\n findAllMidsceneLocatorField,\n markupImageForLLM,\n warnGPT4oSizeLimit,\n} from './common';\nimport {\n automationUserPrompt,\n generateTaskBackgroundContext,\n systemPromptToTaskPlanning,\n} from './prompt/llm-planning';\nimport { describeUserPage } from './prompt/util';\n\nconst debug = getDebug('planning');\n\nexport async function plan(\n userInstruction: string,\n opts: {\n context: UIContext;\n pageType: PageType;\n actionSpace: DeviceAction<any>[];\n callAI?: typeof callAiFn<PlanningAIResponse>;\n log?: string;\n actionContext?: string;\n },\n): Promise<PlanningAIResponse> {\n const { callAI, context } = opts || {};\n const { screenshotBase64, size } = context;\n\n const modelPreferences: IModelPreferences = {\n intent: 'planning',\n };\n const { description: pageDescription, elementById } = await describeUserPage(\n context,\n modelPreferences,\n );\n\n const systemPrompt = await systemPromptToTaskPlanning({\n actionSpace: opts.actionSpace,\n vlMode: vlLocateMode(modelPreferences),\n });\n const taskBackgroundContextText = generateTaskBackgroundContext(\n userInstruction,\n opts.log,\n opts.actionContext,\n );\n const userInstructionPrompt = await automationUserPrompt(\n vlLocateMode(modelPreferences),\n ).format({\n pageDescription,\n taskBackgroundContext: taskBackgroundContextText,\n });\n\n let imagePayload = screenshotBase64;\n if (vlLocateMode(modelPreferences) === 'qwen-vl') {\n imagePayload = await paddingToMatchBlockByBase64(imagePayload);\n } else if (!vlLocateMode(modelPreferences)) {\n imagePayload = await markupImageForLLM(\n screenshotBase64,\n context.tree,\n context.size,\n );\n }\n\n warnGPT4oSizeLimit(size, modelPreferences);\n\n const msgs: AIArgs = [\n { role: 'system', content: systemPrompt },\n {\n role: 'user',\n content: [\n {\n type: 'image_url',\n image_url: {\n url: imagePayload,\n detail: 'high',\n },\n },\n {\n type: 'text',\n text: userInstructionPrompt,\n },\n ],\n },\n ];\n\n const call = callAI || callAiFn;\n const { content, usage } = await call(\n msgs,\n AIActionType.PLAN,\n modelPreferences,\n );\n const rawResponse = JSON.stringify(content, undefined, 2);\n const planFromAI = content;\n\n const actions =\n (planFromAI.action?.type ? [planFromAI.action] : planFromAI.actions) || [];\n const returnValue: PlanningAIResponse = {\n ...planFromAI,\n actions,\n rawResponse,\n usage,\n yamlFlow: buildYamlFlowFromPlans(\n actions,\n opts.actionSpace,\n planFromAI.sleep,\n ),\n };\n\n assert(planFromAI, \"can't get plans from AI\");\n\n // TODO: use zod.parse to parse the action.param, and then fill the bbox param.\n actions.forEach((action) => {\n const type = action.type;\n const actionInActionSpace = opts.actionSpace.find(\n (action) => action.name === type,\n );\n const locateFields = actionInActionSpace\n ? findAllMidsceneLocatorField(actionInActionSpace.paramSchema)\n : [];\n\n debug('locateFields', locateFields);\n\n locateFields.forEach((field) => {\n const locateResult = action.param[field];\n if (locateResult) {\n if (vlLocateMode(modelPreferences)) {\n action.param[field] = fillBboxParam(\n locateResult,\n size.width,\n size.height,\n modelPreferences,\n );\n } else {\n const element = elementById(locateResult);\n if (element) {\n action.param[field].id = element.id;\n }\n }\n }\n
|
|
1
|
+
{"version":3,"file":"ai-model/llm-planning.mjs","sources":["webpack://@midscene/core/./src/ai-model/llm-planning.ts"],"sourcesContent":["import type {\n DeviceAction,\n PageType,\n PlanningAIResponse,\n UIContext,\n} from '@/types';\nimport { type IModelPreferences, vlLocateMode } from '@midscene/shared/env';\nimport { paddingToMatchBlockByBase64 } from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport {\n AIActionType,\n type AIArgs,\n buildYamlFlowFromPlans,\n callAiFn,\n fillBboxParam,\n findAllMidsceneLocatorField,\n markupImageForLLM,\n warnGPT4oSizeLimit,\n} from './common';\nimport {\n automationUserPrompt,\n generateTaskBackgroundContext,\n systemPromptToTaskPlanning,\n} from './prompt/llm-planning';\nimport { describeUserPage } from './prompt/util';\n\nconst debug = getDebug('planning');\n\nexport async function plan(\n userInstruction: string,\n opts: {\n context: UIContext;\n pageType: PageType;\n actionSpace: DeviceAction<any>[];\n callAI?: typeof callAiFn<PlanningAIResponse>;\n log?: string;\n actionContext?: string;\n },\n): Promise<PlanningAIResponse> {\n const { callAI, context } = opts || {};\n const { screenshotBase64, size } = context;\n\n const modelPreferences: IModelPreferences = {\n intent: 'planning',\n };\n const { description: pageDescription, elementById } = await describeUserPage(\n context,\n modelPreferences,\n );\n\n const systemPrompt = await systemPromptToTaskPlanning({\n actionSpace: opts.actionSpace,\n vlMode: vlLocateMode(modelPreferences),\n });\n const taskBackgroundContextText = generateTaskBackgroundContext(\n userInstruction,\n opts.log,\n opts.actionContext,\n );\n const userInstructionPrompt = await automationUserPrompt(\n vlLocateMode(modelPreferences),\n ).format({\n pageDescription,\n taskBackgroundContext: taskBackgroundContextText,\n });\n\n let imagePayload = screenshotBase64;\n if (vlLocateMode(modelPreferences) === 'qwen-vl') {\n imagePayload = await paddingToMatchBlockByBase64(imagePayload);\n } else if (!vlLocateMode(modelPreferences)) {\n imagePayload = await markupImageForLLM(\n screenshotBase64,\n context.tree,\n context.size,\n );\n }\n\n warnGPT4oSizeLimit(size, modelPreferences);\n\n const msgs: AIArgs = [\n { role: 'system', content: systemPrompt },\n {\n role: 'user',\n content: [\n {\n type: 'image_url',\n image_url: {\n url: imagePayload,\n detail: 'high',\n },\n },\n {\n type: 'text',\n text: userInstructionPrompt,\n },\n ],\n },\n ];\n\n const call = callAI || callAiFn;\n const { content, usage } = await call(\n msgs,\n AIActionType.PLAN,\n modelPreferences,\n );\n const rawResponse = JSON.stringify(content, undefined, 2);\n const planFromAI = content;\n\n const actions =\n (planFromAI.action?.type ? [planFromAI.action] : planFromAI.actions) || [];\n const returnValue: PlanningAIResponse = {\n ...planFromAI,\n actions,\n rawResponse,\n usage,\n yamlFlow: buildYamlFlowFromPlans(\n actions,\n opts.actionSpace,\n planFromAI.sleep,\n ),\n };\n\n assert(planFromAI, \"can't get plans from AI\");\n\n // TODO: use zod.parse to parse the action.param, and then fill the bbox param.\n actions.forEach((action) => {\n const type = action.type;\n const actionInActionSpace = opts.actionSpace.find(\n (action) => action.name === type,\n );\n const locateFields = actionInActionSpace\n ? findAllMidsceneLocatorField(actionInActionSpace.paramSchema)\n : [];\n\n debug('locateFields', locateFields);\n\n locateFields.forEach((field) => {\n const locateResult = action.param[field];\n if (locateResult) {\n if (vlLocateMode(modelPreferences)) {\n action.param[field] = fillBboxParam(\n locateResult,\n size.width,\n size.height,\n modelPreferences,\n );\n } else {\n const element = elementById(locateResult);\n if (element) {\n action.param[field].id = element.id;\n }\n }\n }\n });\n });\n // in Qwen-VL, error means error. In GPT-4o, error may mean more actions are needed.\n assert(!planFromAI.error, `Failed to plan actions: ${planFromAI.error}`);\n\n if (\n actions.length === 0 &&\n returnValue.more_actions_needed_by_instruction &&\n !returnValue.sleep\n ) {\n console.warn(\n 'No actions planned for the prompt, but model said more actions are needed:',\n userInstruction,\n );\n }\n\n return returnValue;\n}\n"],"names":["debug","getDebug","plan","userInstruction","opts","_planFromAI_action","callAI","context","screenshotBase64","size","modelPreferences","pageDescription","elementById","describeUserPage","systemPrompt","systemPromptToTaskPlanning","vlLocateMode","taskBackgroundContextText","generateTaskBackgroundContext","userInstructionPrompt","automationUserPrompt","imagePayload","paddingToMatchBlockByBase64","markupImageForLLM","warnGPT4oSizeLimit","msgs","call","callAiFn","content","usage","AIActionType","rawResponse","JSON","undefined","planFromAI","actions","returnValue","buildYamlFlowFromPlans","assert","action","type","actionInActionSpace","locateFields","findAllMidsceneLocatorField","field","locateResult","fillBboxParam","element","console"],"mappings":";;;;;;;AA2BA,MAAMA,QAAQC,SAAS;AAEhB,eAAeC,KACpBC,eAAuB,EACvBC,IAOC;QAwEEC;IAtEH,MAAM,EAAEC,MAAM,EAAEC,OAAO,EAAE,GAAGH,QAAQ,CAAC;IACrC,MAAM,EAAEI,gBAAgB,EAAEC,IAAI,EAAE,GAAGF;IAEnC,MAAMG,mBAAsC;QAC1C,QAAQ;IACV;IACA,MAAM,EAAE,aAAaC,eAAe,EAAEC,WAAW,EAAE,GAAG,MAAMC,iBAC1DN,SACAG;IAGF,MAAMI,eAAe,MAAMC,2BAA2B;QACpD,aAAaX,KAAK,WAAW;QAC7B,QAAQY,aAAaN;IACvB;IACA,MAAMO,4BAA4BC,8BAChCf,iBACAC,KAAK,GAAG,EACRA,KAAK,aAAa;IAEpB,MAAMe,wBAAwB,MAAMC,qBAClCJ,aAAaN,mBACb,MAAM,CAAC;QACPC;QACA,uBAAuBM;IACzB;IAEA,IAAII,eAAeb;IACnB,IAAIQ,AAAmC,cAAnCA,aAAaN,mBACfW,eAAe,MAAMC,4BAA4BD;SAC5C,IAAI,CAACL,aAAaN,mBACvBW,eAAe,MAAME,kBACnBf,kBACAD,QAAQ,IAAI,EACZA,QAAQ,IAAI;IAIhBiB,mBAAmBf,MAAMC;IAEzB,MAAMe,OAAe;QACnB;YAAE,MAAM;YAAU,SAASX;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKO;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAMF;gBACR;aACD;QACH;KACD;IAED,MAAMO,OAAOpB,UAAUqB;IACvB,MAAM,EAAEC,OAAO,EAAEC,KAAK,EAAE,GAAG,MAAMH,KAC/BD,MACAK,aAAa,IAAI,EACjBpB;IAEF,MAAMqB,cAAcC,KAAK,SAAS,CAACJ,SAASK,QAAW;IACvD,MAAMC,aAAaN;IAEnB,MAAMO,UACH9B,AAAAA,CAAAA,SAAAA,CAAAA,qBAAAA,WAAW,MAAM,AAAD,IAAhBA,KAAAA,IAAAA,mBAAmB,IAAI,AAAD,IAAI;QAAC6B,WAAW,MAAM;KAAC,GAAGA,WAAW,OAAM,KAAM,EAAE;IAC5E,MAAME,cAAkC;QACtC,GAAGF,UAAU;QACbC;QACAJ;QACAF;QACA,UAAUQ,uBACRF,SACA/B,KAAK,WAAW,EAChB8B,WAAW,KAAK;IAEpB;IAEAI,OAAOJ,YAAY;IAGnBC,QAAQ,OAAO,CAAC,CAACI;QACf,MAAMC,OAAOD,OAAO,IAAI;QACxB,MAAME,sBAAsBrC,KAAK,WAAW,CAAC,IAAI,CAC/C,CAACmC,SAAWA,OAAO,IAAI,KAAKC;QAE9B,MAAME,eAAeD,sBACjBE,4BAA4BF,oBAAoB,WAAW,IAC3D,EAAE;QAENzC,MAAM,gBAAgB0C;QAEtBA,aAAa,OAAO,CAAC,CAACE;YACpB,MAAMC,eAAeN,OAAO,KAAK,CAACK,MAAM;YACxC,IAAIC,cACF,IAAI7B,aAAaN,mBACf6B,OAAO,KAAK,CAACK,MAAM,GAAGE,cACpBD,cACApC,KAAK,KAAK,EACVA,KAAK,MAAM,EACXC;iBAEG;gBACL,MAAMqC,UAAUnC,YAAYiC;gBAC5B,IAAIE,SACFR,OAAO,KAAK,CAACK,MAAM,CAAC,EAAE,GAAGG,QAAQ,EAAE;YAEvC;QAEJ;IACF;IAEAT,OAAO,CAACJ,WAAW,KAAK,EAAE,CAAC,wBAAwB,EAAEA,WAAW,KAAK,EAAE;IAEvE,IACEC,AAAmB,MAAnBA,QAAQ,MAAM,IACdC,YAAY,kCAAkC,IAC9C,CAACA,YAAY,KAAK,EAElBY,QAAQ,IAAI,CACV,8EACA7C;IAIJ,OAAOiC;AACT"}
|
|
@@ -19,23 +19,42 @@ const descriptionForAction = (action, locatorSchemaTypeDescription)=>{
|
|
|
19
19
|
const shape = action.paramSchema.shape;
|
|
20
20
|
const paramLines = [];
|
|
21
21
|
const getTypeName = (field)=>{
|
|
22
|
-
var
|
|
23
|
-
const
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
22
|
+
var _actualField__def;
|
|
23
|
+
const unwrapField = (f)=>{
|
|
24
|
+
if (!f._def) return f;
|
|
25
|
+
const typeName = f._def.typeName;
|
|
26
|
+
if ('ZodOptional' === typeName || 'ZodNullable' === typeName || 'ZodDefault' === typeName) return unwrapField(f._def.innerType);
|
|
27
|
+
return f;
|
|
28
|
+
};
|
|
29
|
+
const actualField = unwrapField(field);
|
|
30
|
+
const fieldTypeName = null == (_actualField__def = actualField._def) ? void 0 : _actualField__def.typeName;
|
|
31
|
+
if ('ZodString' === fieldTypeName) return 'string';
|
|
32
|
+
if ('ZodNumber' === fieldTypeName) return 'number';
|
|
33
|
+
if ('ZodBoolean' === fieldTypeName) return 'boolean';
|
|
34
|
+
if ('ZodArray' === fieldTypeName) return 'array';
|
|
35
|
+
if ('ZodObject' === fieldTypeName) {
|
|
29
36
|
if (ifMidsceneLocatorField(actualField)) return locatorSchemaTypeDescription;
|
|
30
37
|
return 'object';
|
|
31
38
|
}
|
|
32
|
-
|
|
33
|
-
|
|
39
|
+
if ('ZodEnum' === fieldTypeName) {
|
|
40
|
+
var _actualField__def_values, _actualField__def1;
|
|
41
|
+
const values = (null == (_actualField__def1 = actualField._def) ? void 0 : null == (_actualField__def_values = _actualField__def1.values) ? void 0 : _actualField__def_values.map((option)=>String(`'${option}'`)).join(', ')) ?? 'enum';
|
|
42
|
+
return `enum(${values})`;
|
|
43
|
+
}
|
|
44
|
+
console.warn('failed to parse Zod type. This may lead to wrong params from the LLM.\n', actualField._def);
|
|
45
|
+
return actualField.toString();
|
|
34
46
|
};
|
|
35
47
|
const getDescription = (field)=>{
|
|
36
|
-
var
|
|
37
|
-
const
|
|
48
|
+
var _actualField__def;
|
|
49
|
+
const unwrapField = (f)=>{
|
|
50
|
+
if (!f._def) return f;
|
|
51
|
+
const typeName = f._def.typeName;
|
|
52
|
+
if ('ZodOptional' === typeName || 'ZodNullable' === typeName || 'ZodDefault' === typeName) return unwrapField(f._def.innerType);
|
|
53
|
+
return f;
|
|
54
|
+
};
|
|
38
55
|
if ("description" in field) return field.description || null;
|
|
56
|
+
const actualField = unwrapField(field);
|
|
57
|
+
if ("description" in actualField) return actualField.description || null;
|
|
39
58
|
if ((null == (_actualField__def = actualField._def) ? void 0 : _actualField__def.typeName) === 'ZodObject') {
|
|
40
59
|
if ('midscene_location_field_flag' in actualField._def.shape()) return 'Location information for the target element';
|
|
41
60
|
}
|
|
@@ -101,9 +120,11 @@ this and output the JSON:
|
|
|
101
120
|
"more_actions_needed_by_instruction": false,
|
|
102
121
|
"action": {
|
|
103
122
|
"type": "Tap",
|
|
104
|
-
"
|
|
105
|
-
"
|
|
106
|
-
|
|
123
|
+
"param": {
|
|
124
|
+
"locate": {
|
|
125
|
+
"bbox": [100, 100, 200, 200],
|
|
126
|
+
"prompt": "The 'Yes' button in popup"
|
|
127
|
+
}
|
|
107
128
|
}
|
|
108
129
|
}
|
|
109
130
|
}
|
|
@@ -193,8 +214,9 @@ By viewing the page screenshot and description, you should consider this and out
|
|
|
193
214
|
{
|
|
194
215
|
"thought": "Click the language switch button to open the language options.",
|
|
195
216
|
"type": "Tap",
|
|
196
|
-
"param":
|
|
197
|
-
|
|
217
|
+
"param": {
|
|
218
|
+
"locate": { id: "c81c4e9a33", prompt: "The language switch button" }
|
|
219
|
+
}
|
|
198
220
|
},
|
|
199
221
|
{
|
|
200
222
|
"thought": "Wait for 1 second to ensure the language options are displayed.",
|
|
@@ -214,16 +236,16 @@ Wrong output:
|
|
|
214
236
|
{
|
|
215
237
|
"thought": "Click the language switch button to open the language options.",
|
|
216
238
|
"type": "Tap",
|
|
217
|
-
"param":
|
|
218
|
-
|
|
219
|
-
{ "id": "c81c4e9a33" }, // WRONG: prompt is missing, this is not a valid LocateParam
|
|
239
|
+
"param": {
|
|
240
|
+
"locate": { "id": "c81c4e9a33" } // WRONG: prompt is missing, this is not a valid LocateParam
|
|
220
241
|
}
|
|
221
242
|
},
|
|
222
243
|
{
|
|
223
244
|
"thought": "Click the English option",
|
|
224
245
|
"type": "Tap",
|
|
225
|
-
"param":
|
|
226
|
-
|
|
246
|
+
"param": {
|
|
247
|
+
"locate": null // WRONG: if the element is not on the page, you should not plan this action
|
|
248
|
+
}
|
|
227
249
|
}
|
|
228
250
|
],
|
|
229
251
|
"more_actions_needed_by_instruction": false, // WRONG: should be true
|