@rpascene/core 0.30.8

package/dist/es/ai-model/inspect.mjs

@@ -0,0 +1,286 @@
+import { cropByRect, paddingToMatchBlockByBase64, preProcessImageUrl } from "@rpascene/shared/img";
+import { getDebug } from "@rpascene/shared/logger";
+import { assert } from "@rpascene/shared/utils";
+import { AIActionType, adaptBboxToRect, expandSearchArea, markupImageForLLM, mergeRects } from "./common.mjs";
+import { extractDataQueryPrompt, systemPromptToExtract } from "./prompt/extraction.mjs";
+import { findElementPrompt, systemPromptToLocateElement } from "./prompt/llm-locator.mjs";
+import { sectionLocatorInstruction, systemPromptToLocateSection } from "./prompt/llm-section-locator.mjs";
+import { describeUserPage, distance, distanceThreshold, elementByPositionWithElementInfo } from "./prompt/util.mjs";
+import { callAIWithObjectResponse } from "./service-caller/index.mjs";
+const debugInspect = getDebug('ai:inspect');
+const debugSection = getDebug('ai:section');
+const extraTextFromUserPrompt = (prompt)=>{
+    if ('string' == typeof prompt) return prompt;
+    return prompt.prompt;
+};
+const promptsToChatParam = async (multimodalPrompt)=>{
+    var _multimodalPrompt_images;
+    const msgs = [];
+    if (null == multimodalPrompt ? void 0 : null == (_multimodalPrompt_images = multimodalPrompt.images) ? void 0 : _multimodalPrompt_images.length) {
+        msgs.push({
+            role: 'user',
+            content: [
+                {
+                    type: 'text',
+                    text: 'Next, I will provide all the reference images.'
+                }
+            ]
+        });
+        for (const item of multimodalPrompt.images){
+            const base64 = await preProcessImageUrl(item.url, !!multimodalPrompt.convertHttpImage2Base64);
+            msgs.push({
+                role: 'user',
+                content: [
+                    {
+                        type: 'text',
+                        text: `reference image ${item.name}:`
+                    }
+                ]
+            });
+            msgs.push({
+                role: 'user',
+                content: [
+                    {
+                        type: 'image_url',
+                        image_url: {
+                            url: base64,
+                            detail: 'high'
+                        }
+                    }
+                ]
+            });
+        }
+    }
+    return msgs;
+};
+async function AiLocateElement(options) {
+    const { context, targetElementDescription, callAIFn, modelConfig } = options;
+    const { vlMode } = modelConfig;
+    const { screenshotBase64 } = context;
+    const { description, elementById, insertElementByPosition } = await describeUserPage(context, {
+        vlMode
+    });
+    assert(targetElementDescription, "cannot find the target element description");
+    const userInstructionPrompt = await findElementPrompt.format({
+        pageDescription: description,
+        targetElementDescription: extraTextFromUserPrompt(targetElementDescription)
+    });
+    const systemPrompt = systemPromptToLocateElement(vlMode);
+    let imagePayload = screenshotBase64;
+    let imageWidth = context.size.width;
+    let imageHeight = context.size.height;
+    let originalImageWidth = imageWidth;
+    let originalImageHeight = imageHeight;
+    if (options.searchConfig) {
+        var _options_searchConfig_rect, _options_searchConfig_rect1;
+        assert(options.searchConfig.rect, 'searchArea is provided but its rect cannot be found. Failed to locate element');
+        assert(options.searchConfig.imageBase64, 'searchArea is provided but its imageBase64 cannot be found. Failed to locate element');
+        imagePayload = options.searchConfig.imageBase64;
+        imageWidth = null == (_options_searchConfig_rect = options.searchConfig.rect) ? void 0 : _options_searchConfig_rect.width;
+        imageHeight = null == (_options_searchConfig_rect1 = options.searchConfig.rect) ? void 0 : _options_searchConfig_rect1.height;
+        originalImageWidth = imageWidth;
+        originalImageHeight = imageHeight;
+    } else if ('qwen-vl' === vlMode) {
+        const paddedResult = await paddingToMatchBlockByBase64(imagePayload);
+        imageWidth = paddedResult.width;
+        imageHeight = paddedResult.height;
+        imagePayload = paddedResult.imageBase64;
+    } else if ('qwen3-vl' === vlMode) ;
+    else if (!vlMode) imagePayload = await markupImageForLLM(screenshotBase64, context.tree, context.size);
+    const msgs = [
+        {
+            role: 'system',
+            content: systemPrompt
+        },
+        {
+            role: 'user',
+            content: [
+                {
+                    type: 'image_url',
+                    image_url: {
+                        url: imagePayload,
+                        detail: 'high'
+                    }
+                },
+                {
+                    type: 'text',
+                    text: userInstructionPrompt
+                }
+            ]
+        }
+    ];
+    if ('string' != typeof targetElementDescription) {
+        const addOns = await promptsToChatParam({
+            images: targetElementDescription.images,
+            convertHttpImage2Base64: targetElementDescription.convertHttpImage2Base64
+        });
+        msgs.push(...addOns);
+    }
+    const res = await callAIFn(msgs, AIActionType.INSPECT_ELEMENT, modelConfig);
+    const rawResponse = JSON.stringify(res.content);
+    let resRect;
+    let matchedElements = 'elements' in res.content ? res.content.elements : [];
+    let errors = 'errors' in res.content ? res.content.errors : [];
+    try {
+        if ('bbox' in res.content && Array.isArray(res.content.bbox)) {
+            var _options_searchConfig_rect2, _options_searchConfig, _options_searchConfig_rect3, _options_searchConfig1;
+            resRect = adaptBboxToRect(res.content.bbox, imageWidth, imageHeight, null == (_options_searchConfig = options.searchConfig) ? void 0 : null == (_options_searchConfig_rect2 = _options_searchConfig.rect) ? void 0 : _options_searchConfig_rect2.left, null == (_options_searchConfig1 = options.searchConfig) ? void 0 : null == (_options_searchConfig_rect3 = _options_searchConfig1.rect) ? void 0 : _options_searchConfig_rect3.top, originalImageWidth, originalImageHeight, vlMode);
+            debugInspect('resRect', resRect);
+            const rectCenter = {
+                x: resRect.left + resRect.width / 2,
+                y: resRect.top + resRect.height / 2
+            };
+            let element = elementByPositionWithElementInfo(context.tree, rectCenter);
+            const distanceToCenter = element ? distance({
+                x: element.center[0],
+                y: element.center[1]
+            }, rectCenter) : 0;
+            if (!element || distanceToCenter > distanceThreshold) element = insertElementByPosition(rectCenter);
+            if (element) {
+                matchedElements = [
+                    element
+                ];
+                errors = [];
+            }
+        }
+    } catch (e) {
+        const msg = e instanceof Error ? `Failed to parse bbox: ${e.message}` : 'unknown error in locate';
+        if (errors && (null == errors ? void 0 : errors.length) !== 0) errors.push(`(${msg})`);
+        else errors = [
+            msg
+        ];
+    }
+    return {
+        rect: resRect,
+        parseResult: {
+            elements: matchedElements,
+            errors
+        },
+        rawResponse,
+        elementById,
+        usage: res.usage,
+        isOrderSensitive: 'object' == typeof res.content && null !== res.content && 'isOrderSensitive' in res.content ? res.content.isOrderSensitive : void 0
+    };
+}
+async function AiLocateSection(options) {
+    const { context, sectionDescription, modelConfig } = options;
+    const { vlMode } = modelConfig;
+    const { screenshotBase64 } = context;
+    const systemPrompt = systemPromptToLocateSection(vlMode);
+    const sectionLocatorInstructionText = await sectionLocatorInstruction.format({
+        sectionDescription: extraTextFromUserPrompt(sectionDescription)
+    });
+    const msgs = [
+        {
+            role: 'system',
+            content: systemPrompt
+        },
+        {
+            role: 'user',
+            content: [
+                {
+                    type: 'image_url',
+                    image_url: {
+                        url: screenshotBase64,
+                        detail: 'high'
+                    }
+                },
+                {
+                    type: 'text',
+                    text: sectionLocatorInstructionText
+                }
+            ]
+        }
+    ];
+    if ('string' != typeof sectionDescription) {
+        const addOns = await promptsToChatParam({
+            images: sectionDescription.images,
+            convertHttpImage2Base64: sectionDescription.convertHttpImage2Base64
+        });
+        msgs.push(...addOns);
+    }
+    const result = await callAIWithObjectResponse(msgs, AIActionType.EXTRACT_DATA, modelConfig);
+    let sectionRect;
+    const sectionBbox = result.content.bbox;
+    if (sectionBbox) {
+        const targetRect = adaptBboxToRect(sectionBbox, context.size.width, context.size.height, 0, 0, context.size.width, context.size.height, vlMode);
+        debugSection('original targetRect %j', targetRect);
+        const referenceBboxList = result.content.references_bbox || [];
+        debugSection('referenceBboxList %j', referenceBboxList);
+        const referenceRects = referenceBboxList.filter((bbox)=>Array.isArray(bbox)).map((bbox)=>adaptBboxToRect(bbox, context.size.width, context.size.height, 0, 0, context.size.width, context.size.height, vlMode));
+        debugSection('referenceRects %j', referenceRects);
+        const mergedRect = mergeRects([
+            targetRect,
+            ...referenceRects
+        ]);
+        debugSection('mergedRect %j', mergedRect);
+        sectionRect = expandSearchArea(mergedRect, context.size, vlMode);
+        debugSection('expanded sectionRect %j', sectionRect);
+    }
+    let imageBase64 = screenshotBase64;
+    if (sectionRect) {
+        const croppedResult = await cropByRect(screenshotBase64, sectionRect, 'qwen-vl' === vlMode);
+        imageBase64 = croppedResult.imageBase64;
+        sectionRect.width = croppedResult.width;
+        sectionRect.height = croppedResult.height;
+    }
+    return {
+        rect: sectionRect,
+        imageBase64,
+        error: result.content.error,
+        rawResponse: JSON.stringify(result.content),
+        usage: result.usage
+    };
+}
+async function AiExtractElementInfo(options) {
+    const { dataQuery, context, extractOption, multimodalPrompt, modelConfig } = options;
+    const { vlMode } = modelConfig;
+    const systemPrompt = systemPromptToExtract();
+    const { screenshotBase64 } = context;
+    const { description, elementById } = await describeUserPage(context, {
+        truncateTextLength: 200,
+        filterNonTextContent: false,
+        visibleOnly: false,
+        domIncluded: null == extractOption ? void 0 : extractOption.domIncluded,
+        vlMode
+    });
+    const extractDataPromptText = await extractDataQueryPrompt(description, dataQuery);
+    const userContent = [];
+    if ((null == extractOption ? void 0 : extractOption.screenshotIncluded) !== false) userContent.push({
+        type: 'image_url',
+        image_url: {
+            url: screenshotBase64,
+            detail: 'high'
+        }
+    });
+    userContent.push({
+        type: 'text',
+        text: extractDataPromptText
+    });
+    const msgs = [
+        {
+            role: 'system',
+            content: systemPrompt
+        },
+        {
+            role: 'user',
+            content: userContent
+        }
+    ];
+    if (multimodalPrompt) {
+        const addOns = await promptsToChatParam({
+            images: multimodalPrompt.images,
+            convertHttpImage2Base64: multimodalPrompt.convertHttpImage2Base64
+        });
+        msgs.push(...addOns);
+    }
+    const result = await callAIWithObjectResponse(msgs, AIActionType.EXTRACT_DATA, modelConfig);
+    return {
+        parseResult: result.content,
+        elementById,
+        usage: result.usage
+    };
+}
+export { AiExtractElementInfo, AiLocateElement, AiLocateSection };
+
+//# sourceMappingURL=inspect.mjs.map

package/dist/es/ai-model/inspect.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"ai-model\\inspect.mjs","sources":["webpack://@rpascene/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 type { IModelConfig } from '@rpascene/shared/env';\nimport {\n  cropByRect,\n  paddingToMatchBlockByBase64,\n  preProcessImageUrl,\n} from '@rpascene/shared/img';\nimport { getDebug } from '@rpascene/shared/logger';\nimport { assert } from '@rpascene/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  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 { callAIWithObjectResponse } 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  callAIFn: typeof callAIWithObjectResponse<\n    AIElementResponse | [number, number]\n  >;\n  searchConfig?: Awaited<ReturnType<typeof AiLocateSection>>;\n  modelConfig: IModelConfig;\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, callAIFn, modelConfig } = options;\n  const { vlMode } = modelConfig;\n  const { screenshotBase64 } = context;\n\n  const { description, elementById, insertElementByPosition } =\n    await describeUserPage(context, { vlMode });\n\n  assert(\n    targetElementDescription,\n    'cannot find the target element description',\n  );\n  const userInstructionPrompt = await findElementPrompt.format({\n    pageDescription: description,\n    targetElementDescription: extraTextFromUserPrompt(targetElementDescription),\n  });\n  const systemPrompt = systemPromptToLocateElement(vlMode);\n\n  let imagePayload = screenshotBase64;\n  let imageWidth = context.size.width;\n  let imageHeight = context.size.height;\n  let originalImageWidth = imageWidth;\n  let originalImageHeight = imageHeight;\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    imageWidth = options.searchConfig.rect?.width;\n    imageHeight = options.searchConfig.rect?.height;\n    originalImageWidth = imageWidth;\n    originalImageHeight = imageHeight;\n  } else if (vlMode === 'qwen-vl') {\n    const paddedResult = await paddingToMatchBlockByBase64(imagePayload);\n    imageWidth = paddedResult.width;\n    imageHeight = paddedResult.height;\n    imagePayload = paddedResult.imageBase64;\n  } else if (vlMode === 'qwen3-vl') {\n    // const paddedResult = await paddingToMatchBlockByBase64(imagePayload, 32);\n    // imageWidth = paddedResult.width;\n    // imageHeight = paddedResult.height;\n    // imagePayload = paddedResult.imageBase64;\n  } else if (!vlMode) {\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  // console.log(msgs, 'msgs')\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 res = await callAIFn(msgs, AIActionType.INSPECT_ELEMENT, modelConfig);\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        imageWidth,\n        imageHeight,\n        options.searchConfig?.rect?.left,\n        options.searchConfig?.rect?.top,\n        originalImageWidth,\n        originalImageHeight,\n        vlMode,\n      );\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  modelConfig: IModelConfig;\n}): Promise<{\n  rect?: Rect;\n  imageBase64?: string;\n  error?: string;\n  rawResponse: string;\n  usage?: AIUsageInfo;\n}> {\n  const { context, sectionDescription, modelConfig } = options;\n  const { vlMode } = modelConfig;\n  const { screenshotBase64 } = context;\n\n  const systemPrompt = systemPromptToLocateSection(vlMode);\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 callAIWithObjectResponse<AISectionLocatorResponse>(\n    msgs,\n    AIActionType.EXTRACT_DATA,\n    modelConfig,\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      0,\n      0,\n      context.size.width,\n      context.size.height,\n      vlMode,\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          0,\n          0,\n          context.size.width,\n          context.size.height,\n          vlMode,\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, vlMode);\n    debugSection('expanded sectionRect %j', sectionRect);\n  }\n\n  let imageBase64 = screenshotBase64;\n  if (sectionRect) {\n    const croppedResult = await cropByRect(\n      screenshotBase64,\n      sectionRect,\n      vlMode === 'qwen-vl',\n    );\n    imageBase64 = croppedResult.imageBase64;\n    sectionRect.width = croppedResult.width;\n    sectionRect.height = croppedResult.height;\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  modelConfig: IModelConfig;\n}) {\n  const { dataQuery, context, extractOption, multimodalPrompt, modelConfig } =\n    options;\n  const { vlMode } = modelConfig;\n  const systemPrompt = systemPromptToExtract();\n\n  const { screenshotBase64 } = context;\n\n  const { description, elementById } = await describeUserPage(context, {\n    truncateTextLength: 200,\n    filterNonTextContent: false,\n    visibleOnly: false,\n    domIncluded: extractOption?.domIncluded,\n    vlMode,\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 (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 callAIWithObjectResponse<AIDataExtractionResponse<T>>(\n    msgs,\n    AIActionType.EXTRACT_DATA,\n    modelConfig,\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","callAIFn","modelConfig","vlMode","screenshotBase64","description","elementById","insertElementByPosition","describeUserPage","assert","userInstructionPrompt","findElementPrompt","systemPrompt","systemPromptToLocateElement","imagePayload","imageWidth","imageHeight","originalImageWidth","originalImageHeight","_options_searchConfig_rect","_options_searchConfig_rect1","paddedResult","paddingToMatchBlockByBase64","markupImageForLLM","addOns","res","AIActionType","rawResponse","JSON","resRect","matchedElements","errors","Array","_options_searchConfig_rect2","_options_searchConfig_rect3","adaptBboxToRect","rectCenter","element","elementByPositionWithElementInfo","distanceToCenter","distance","distanceThreshold","e","msg","Error","undefined","AiLocateSection","sectionDescription","systemPromptToLocateSection","sectionLocatorInstructionText","sectionLocatorInstruction","result","callAIWithObjectResponse","sectionRect","sectionBbox","targetRect","referenceBboxList","referenceRects","bbox","mergedRect","mergeRects","expandSearchArea","imageBase64","croppedResult","cropByRect","AiExtractElementInfo","dataQuery","extractOption","systemPromptToExtract","extractDataPromptText","extractDataQueryPrompt","userContent"],"mappings":";;;;;;;;;AA0DA,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,OASD;IAQC,MAAM,EAAEC,OAAO,EAAEC,wBAAwB,EAAEC,QAAQ,EAAEC,WAAW,EAAE,GAAGJ;IACrE,MAAM,EAAEK,MAAM,EAAE,GAAGD;IACnB,MAAM,EAAEE,gBAAgB,EAAE,GAAGL;IAE7B,MAAM,EAAEM,WAAW,EAAEC,WAAW,EAAEC,uBAAuB,EAAE,GACzD,MAAMC,iBAAiBT,SAAS;QAAEI;IAAO;IAE3CM,OACET,0BACA;IAEF,MAAMU,wBAAwB,MAAMC,kBAAkB,MAAM,CAAC;QAC3D,iBAAiBN;QACjB,0BAA0BjB,wBAAwBY;IACpD;IACA,MAAMY,eAAeC,4BAA4BV;IAEjD,IAAIW,eAAeV;IACnB,IAAIW,aAAahB,QAAQ,IAAI,CAAC,KAAK;IACnC,IAAIiB,cAAcjB,QAAQ,IAAI,CAAC,MAAM;IACrC,IAAIkB,qBAAqBF;IACzB,IAAIG,sBAAsBF;IAE1B,IAAIlB,QAAQ,YAAY,EAAE;YAWXqB,4BACCC;QAXdX,OACEX,QAAQ,YAAY,CAAC,IAAI,EACzB;QAEFW,OACEX,QAAQ,YAAY,CAAC,WAAW,EAChC;QAGFgB,eAAehB,QAAQ,YAAY,CAAC,WAAW;QAC/CiB,aAAa,QAAAI,CAAAA,6BAAAA,QAAQ,YAAY,CAAC,IAAI,AAAD,IAAxBA,KAAAA,IAAAA,2BAA2B,KAAK;QAC7CH,cAAc,QAAAI,CAAAA,8BAAAA,QAAQ,YAAY,CAAC,IAAI,AAAD,IAAxBA,KAAAA,IAAAA,4BAA2B,MAAM;QAC/CH,qBAAqBF;QACrBG,sBAAsBF;IACxB,OAAO,IAAIb,AAAW,cAAXA,QAAsB;QAC/B,MAAMkB,eAAe,MAAMC,4BAA4BR;QACvDC,aAAaM,aAAa,KAAK;QAC/BL,cAAcK,aAAa,MAAM;QACjCP,eAAeO,aAAa,WAAW;IACzC,OAAO,IAAIlB,AAAW,eAAXA;SAKJ,IAAI,CAACA,QACVW,eAAe,MAAMS,kBACnBnB,kBACAL,QAAQ,IAAI,EACZA,QAAQ,IAAI;IAIhB,MAAMN,OAAe;QACnB;YAAE,MAAM;YAAU,SAASmB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKE;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAMJ;gBACR;aACD;QACH;KACD;IAED,IAAI,AAAoC,YAApC,OAAOV,0BAAuC;QAChD,MAAMwB,SAAS,MAAMlC,mBAAmB;YACtC,QAAQU,yBAAyB,MAAM;YACvC,yBAAyBA,yBAAyB,uBAAuB;QAC3E;QACAP,KAAK,IAAI,IAAI+B;IACf;IAEA,MAAMC,MAAM,MAAMxB,SAASR,MAAMiC,aAAa,eAAe,EAAExB;IAE/D,MAAMyB,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;gBAK1DQ,6BAAAA,uBACAC,6BAAAA;YALFL,UAAUM,gBACRV,IAAI,OAAO,CAAC,IAAI,EAChBV,YACAC,aAAAA,QACAiB,CAAAA,wBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,sBAAsB,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,EAC/BjB,oBACAC,qBACAf;YAGFlB,aAAa,WAAW4C;YAExB,MAAMO,aAAa;gBACjB,GAAGP,QAAQ,IAAI,GAAGA,QAAQ,KAAK,GAAG;gBAClC,GAAGA,QAAQ,GAAG,GAAGA,QAAQ,MAAM,GAAG;YACpC;YACA,IAAIQ,UAAUC,iCAAiCvC,QAAQ,IAAI,EAAEqC;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,UAAU9B,wBAAwB6B;YAGpC,IAAIC,SAAS;gBACXP,kBAAkB;oBAACO;iBAAQ;gBAC3BN,SAAS,EAAE;YACb;QACF;IACF,EAAE,OAAOW,GAAG;QACV,MAAMC,MACJD,aAAaE,QACT,CAAC,sBAAsB,EAAEF,EAAE,OAAO,EAAE,GACpC;QACN,IAAI,AAACX,UAAUA,AAAAA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM,AAAD,MAAM,GAGhCA,OAAO,IAAI,CAAC,CAAC,CAAC,EAAEY,IAAI,CAAC,CAAC;aAFtBZ,SAAS;YAACY;SAAI;IAIlB;IAEA,OAAO;QACL,MAAMd;QACN,aAAa;YACX,UAAUC;YACVC;QACF;QACAJ;QACArB;QACA,OAAOmB,IAAI,KAAK;QAChB,kBACE,AAAuB,YAAvB,OAAOA,IAAI,OAAO,IAChBA,AAAgB,SAAhBA,IAAI,OAAO,IACX,sBAAsBA,IAAI,OAAO,GAC9BA,IAAI,OAAO,CAAS,gBAAgB,GACrCoB;IACR;AACF;AAEO,eAAeC,gBAAgBhD,OAIrC;IAOC,MAAM,EAAEC,OAAO,EAAEgD,kBAAkB,EAAE7C,WAAW,EAAE,GAAGJ;IACrD,MAAM,EAAEK,MAAM,EAAE,GAAGD;IACnB,MAAM,EAAEE,gBAAgB,EAAE,GAAGL;IAE7B,MAAMa,eAAeoC,4BAA4B7C;IACjD,MAAM8C,gCAAgC,MAAMC,0BAA0B,MAAM,CAAC;QAC3E,oBAAoB9D,wBAAwB2D;IAC9C;IACA,MAAMtD,OAAe;QACnB;YAAE,MAAM;YAAU,SAASmB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKR;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAM6C;gBACR;aACD;QACH;KACD;IAED,IAAI,AAA8B,YAA9B,OAAOF,oBAAiC;QAC1C,MAAMvB,SAAS,MAAMlC,mBAAmB;YACtC,QAAQyD,mBAAmB,MAAM;YACjC,yBAAyBA,mBAAmB,uBAAuB;QACrE;QACAtD,KAAK,IAAI,IAAI+B;IACf;IAEA,MAAM2B,SAAS,MAAMC,yBACnB3D,MACAiC,aAAa,YAAY,EACzBxB;IAGF,IAAImD;IACJ,MAAMC,cAAcH,OAAO,OAAO,CAAC,IAAI;IACvC,IAAIG,aAAa;QACf,MAAMC,aAAapB,gBACjBmB,aACAvD,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAEFhB,aAAa,0BAA0BoE;QAEvC,MAAMC,oBAAoBL,OAAO,OAAO,CAAC,eAAe,IAAI,EAAE;QAC9DhE,aAAa,wBAAwBqE;QAErC,MAAMC,iBAAiBD,kBACpB,MAAM,CAAC,CAACE,OAAS1B,MAAM,OAAO,CAAC0B,OAC/B,GAAG,CAAC,CAACA,OACGvB,gBACLuB,MACA3D,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAGNhB,aAAa,qBAAqBsE;QAGlC,MAAME,aAAaC,WAAW;YAACL;eAAeE;SAAe;QAC7DtE,aAAa,iBAAiBwE;QAG9BN,cAAcQ,iBAAiBF,YAAY5D,QAAQ,IAAI,EAAEI;QACzDhB,aAAa,2BAA2BkE;IAC1C;IAEA,IAAIS,cAAc1D;IAClB,IAAIiD,aAAa;QACf,MAAMU,gBAAgB,MAAMC,WAC1B5D,kBACAiD,aACAlD,AAAW,cAAXA;QAEF2D,cAAcC,cAAc,WAAW;QACvCV,YAAY,KAAK,GAAGU,cAAc,KAAK;QACvCV,YAAY,MAAM,GAAGU,cAAc,MAAM;IAC3C;IAEA,OAAO;QACL,MAAMV;QACNS;QACA,OAAOX,OAAO,OAAO,CAAC,KAAK;QAC3B,aAAavB,KAAK,SAAS,CAACuB,OAAO,OAAO;QAC1C,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAec,qBAGpBnE,OAMD;IACC,MAAM,EAAEoE,SAAS,EAAEnE,OAAO,EAAEoE,aAAa,EAAE5E,gBAAgB,EAAEW,WAAW,EAAE,GACxEJ;IACF,MAAM,EAAEK,MAAM,EAAE,GAAGD;IACnB,MAAMU,eAAewD;IAErB,MAAM,EAAEhE,gBAAgB,EAAE,GAAGL;IAE7B,MAAM,EAAEM,WAAW,EAAEC,WAAW,EAAE,GAAG,MAAME,iBAAiBT,SAAS;QACnE,oBAAoB;QACpB,sBAAsB;QACtB,aAAa;QACb,aAAaoE,QAAAA,gBAAAA,KAAAA,IAAAA,cAAe,WAAW;QACvChE;IACF;IAEA,MAAMkE,wBAAwB,MAAMC,uBAClCjE,aACA6D;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,MAAM5E,OAAe;QACnB;YAAE,MAAM;YAAU,SAASmB;QAAa;QACxC;YACE,MAAM;YACN,SAAS2D;QACX;KACD;IAED,IAAIhF,kBAAkB;QACpB,MAAMiC,SAAS,MAAMlC,mBAAmB;YACtC,QAAQC,iBAAiB,MAAM;YAC/B,yBAAyBA,iBAAiB,uBAAuB;QACnE;QACAE,KAAK,IAAI,IAAI+B;IACf;IAEA,MAAM2B,SAAS,MAAMC,yBACnB3D,MACAiC,aAAa,YAAY,EACzBxB;IAEF,OAAO;QACL,aAAaiD,OAAO,OAAO;QAC3B7C;QACA,OAAO6C,OAAO,KAAK;IACrB;AACF"}

package/dist/es/ai-model/llm-planning.mjs

@@ -0,0 +1,140 @@
+import { paddingToMatchBlockByBase64 } from "@rpascene/shared/img";
+import { getDebug } from "@rpascene/shared/logger";
+import { assert } from "@rpascene/shared/utils";
+import { AIActionType, buildYamlFlowFromPlans, fillBboxParam, findAllRpasceneLocatorField, markupImageForLLM, warnGPT4oSizeLimit } from "./common.mjs";
+import { systemPromptToTaskPlanning } from "./prompt/llm-planning.mjs";
+import { describeUserPage } from "./prompt/util.mjs";
+import { callAIWithObjectResponse } from "./service-caller/index.mjs";
+const debug = getDebug('planning');
+async function plan(userInstruction, opts) {
+    var _opts_conversationHistory, _planFromAI_action;
+    const { context, modelConfig, conversationHistory } = opts;
+    const { screenshotBase64, size } = context;
+    const { modelName, vlMode } = modelConfig;
+    const { description: pageDescription, elementById } = await describeUserPage(context, {
+        vlMode
+    });
+    const systemPrompt = await systemPromptToTaskPlanning({
+        actionSpace: opts.actionSpace,
+        vlMode: vlMode
+    });
+    let imagePayload = screenshotBase64;
+    let imageWidth = size.width;
+    let imageHeight = size.height;
+    const rightLimit = imageWidth;
+    const bottomLimit = imageHeight;
+    if ('qwen-vl' === vlMode) {
+        const paddedResult = await paddingToMatchBlockByBase64(imagePayload);
+        imageWidth = paddedResult.width;
+        imageHeight = paddedResult.height;
+        imagePayload = paddedResult.imageBase64;
+    } else if ('qwen3-vl' === vlMode) ;
+    else if (!vlMode) imagePayload = await markupImageForLLM(screenshotBase64, context.tree, {
+        width: imageWidth,
+        height: imageHeight
+    });
+    warnGPT4oSizeLimit(size, modelName);
+    const historyLog = (null == (_opts_conversationHistory = opts.conversationHistory) ? void 0 : _opts_conversationHistory.snapshot()) || [];
+    const knowledgeContext = opts.actionContext ? [
+        {
+            role: 'user',
+            content: [
+                {
+                    type: 'text',
+                    text: `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>`
+                }
+            ]
+        }
+    ] : [];
+    const instruction = [
+        {
+            role: 'user',
+            content: [
+                {
+                    type: 'text',
+                    text: `<user_instruction>${userInstruction}</user_instruction>`
+                }
+            ]
+        }
+    ];
+    const msgs = [
+        {
+            role: 'system',
+            content: systemPrompt
+        },
+        ...knowledgeContext,
+        ...instruction,
+        ...historyLog,
+        {
+            role: 'user',
+            content: [
+                {
+                    type: 'image_url',
+                    image_url: {
+                        url: imagePayload,
+                        detail: 'high'
+                    }
+                },
+                ...vlMode ? [] : [
+                    {
+                        type: 'text',
+                        text: pageDescription
+                    }
+                ]
+            ]
+        }
+    ];
+    const { content, usage } = await callAIWithObjectResponse(msgs, AIActionType.PLAN, modelConfig);
+    const rawResponse = JSON.stringify(content, void 0, 2);
+    const planFromAI = content;
+    const actions = ((null == (_planFromAI_action = planFromAI.action) ? void 0 : _planFromAI_action.type) ? [
+        planFromAI.action
+    ] : planFromAI.actions) || [];
+    const returnValue = {
+        ...planFromAI,
+        actions,
+        rawResponse,
+        usage,
+        yamlFlow: buildYamlFlowFromPlans(actions, opts.actionSpace, planFromAI.sleep)
+    };
+    assert(planFromAI, "can't get plans from AI");
+    actions.forEach((action)=>{
+        const type = action.type;
+        const actionInActionSpace = opts.actionSpace.find((action)=>action.name === type);
+        debug('actionInActionSpace matched', actionInActionSpace);
+        const locateFields = actionInActionSpace ? findAllRpasceneLocatorField(actionInActionSpace.paramSchema) : [];
+        debug('locateFields', locateFields);
+        locateFields.forEach((field)=>{
+            const locateResult = action.param[field];
+            if (locateResult) if (vlMode) action.param[field] = fillBboxParam(locateResult, imageWidth, imageHeight, rightLimit, bottomLimit, vlMode);
+            else {
+                const element = elementById(locateResult);
+                if (element) action.param[field].id = element.id;
+            }
+        });
+    });
+    assert(!planFromAI.error, `Failed to plan actions: ${planFromAI.error}`);
+    if (0 === actions.length && returnValue.more_actions_needed_by_instruction && !returnValue.sleep) console.warn('No actions planned for the prompt, but model said more actions are needed:', userInstruction);
+    null == conversationHistory || conversationHistory.append({
+        role: 'assistant',
+        content: [
+            {
+                type: 'text',
+                text: rawResponse
+            }
+        ]
+    });
+    null == conversationHistory || conversationHistory.append({
+        role: 'user',
+        content: [
+            {
+                type: 'text',
+                text: 'I have finished the action previously planned'
+            }
+        ]
+    });
+    return returnValue;
+}
+export { plan };
+
+//# sourceMappingURL=llm-planning.mjs.map

package/dist/es/ai-model/llm-planning.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"ai-model\\llm-planning.mjs","sources":["webpack://@rpascene/core/./src/ai-model/llm-planning.ts"],"sourcesContent":["import type {\n  DeviceAction,\n  InterfaceType,\n  PlanningAIResponse,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig } from '@rpascene/shared/env';\nimport { paddingToMatchBlockByBase64 } from '@rpascene/shared/img';\nimport { getDebug } from '@rpascene/shared/logger';\nimport { assert } from '@rpascene/shared/utils';\nimport type {\n  ChatCompletionContentPart,\n  ChatCompletionMessageParam,\n} from 'openai/resources/index';\nimport {\n  AIActionType,\n  buildYamlFlowFromPlans,\n  fillBboxParam,\n  findAllRpasceneLocatorField,\n  markupImageForLLM,\n  warnGPT4oSizeLimit,\n} from './common';\nimport type { ConversationHistory } from './conversation-history';\nimport { systemPromptToTaskPlanning } from './prompt/llm-planning';\nimport { describeUserPage } from './prompt/util';\nimport { callAIWithObjectResponse } from './service-caller/index';\n\nconst debug = getDebug('planning');\n\nexport async function plan(\n  userInstruction: string,\n  opts: {\n    context: UIContext;\n    interfaceType: InterfaceType;\n    actionSpace: DeviceAction<any>[];\n    actionContext?: string;\n    modelConfig: IModelConfig;\n    conversationHistory?: ConversationHistory;\n  },\n): Promise<PlanningAIResponse> {\n  const { context, modelConfig, conversationHistory } = opts;\n  const { screenshotBase64, size } = context;\n\n  const { modelName, vlMode } = modelConfig;\n\n  const { description: pageDescription, elementById } = await describeUserPage(\n    context,\n    { vlMode },\n  );\n  const systemPrompt = await systemPromptToTaskPlanning({\n    actionSpace: opts.actionSpace,\n    vlMode: vlMode,\n  });\n\n  let imagePayload = screenshotBase64;\n  let imageWidth = size.width;\n  let imageHeight = size.height;\n  const rightLimit = imageWidth;\n  const bottomLimit = imageHeight;\n  if (vlMode === 'qwen-vl') {\n    const paddedResult = await paddingToMatchBlockByBase64(imagePayload);\n    imageWidth = paddedResult.width;\n    imageHeight = paddedResult.height;\n    imagePayload = paddedResult.imageBase64;\n  } else if (vlMode === 'qwen3-vl') {\n    // const paddedResult = await paddingToMatchBlockByBase64(imagePayload, 32);\n    // imageWidth = paddedResult.width;\n    // imageHeight = paddedResult.height;\n    // imagePayload = paddedResult.imageBase64;\n  } else if (!vlMode) {\n    imagePayload = await markupImageForLLM(screenshotBase64, context.tree, {\n      width: imageWidth,\n      height: imageHeight,\n    });\n  }\n\n  warnGPT4oSizeLimit(size, modelName);\n\n  const historyLog = opts.conversationHistory?.snapshot() || [];\n  // .filter((item) => item.role === 'assistant') || [];\n\n  const knowledgeContext: ChatCompletionMessageParam[] = opts.actionContext\n    ? [\n      {\n        role: 'user',\n        content: [\n          {\n            type: 'text',\n            text: `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>`,\n          },\n        ],\n      },\n    ]\n    : [];\n\n  const instruction: ChatCompletionMessageParam[] = [\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `<user_instruction>${userInstruction}</user_instruction>`,\n        },\n      ],\n    },\n  ];\n\n  const msgs: ChatCompletionMessageParam[] = [\n    { role: 'system', content: systemPrompt },\n    ...knowledgeContext,\n    ...instruction,\n    ...historyLog,\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n        ...(vlMode\n          ? []\n          : ([\n            {\n              type: 'text',\n              text: pageDescription,\n            },\n          ] as ChatCompletionContentPart[])),\n      ],\n    },\n  ];\n\n  const { content, usage } = await callAIWithObjectResponse<PlanningAIResponse>(\n    msgs,\n    AIActionType.PLAN,\n    modelConfig,\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\n    debug('actionInActionSpace matched', actionInActionSpace);\n    const locateFields = actionInActionSpace\n      ? findAllRpasceneLocatorField(actionInActionSpace.paramSchema)\n      : [];\n\n    debug('locateFields', locateFields);\n\n    locateFields.forEach((field) => {\n      const locateResult = action.param[field];\n      if (locateResult) {\n        if (vlMode) {\n          action.param[field] = fillBboxParam(\n            locateResult,\n            imageWidth,\n            imageHeight,\n            rightLimit,\n            bottomLimit,\n            vlMode,\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  conversationHistory?.append({\n    role: 'assistant',\n    content: [\n      {\n        type: 'text',\n        text: rawResponse,\n      },\n    ],\n  });\n  conversationHistory?.append({\n    role: 'user',\n    content: [\n      {\n        type: 'text',\n        text: 'I have finished the action previously planned',\n      },\n    ],\n  });\n\n  return returnValue;\n}\n"],"names":["debug","getDebug","plan","userInstruction","opts","_opts_conversationHistory","_planFromAI_action","context","modelConfig","conversationHistory","screenshotBase64","size","modelName","vlMode","pageDescription","elementById","describeUserPage","systemPrompt","systemPromptToTaskPlanning","imagePayload","imageWidth","imageHeight","rightLimit","bottomLimit","paddedResult","paddingToMatchBlockByBase64","markupImageForLLM","warnGPT4oSizeLimit","historyLog","knowledgeContext","instruction","msgs","content","usage","callAIWithObjectResponse","AIActionType","rawResponse","JSON","undefined","planFromAI","actions","returnValue","buildYamlFlowFromPlans","assert","action","type","actionInActionSpace","locateFields","findAllRpasceneLocatorField","field","locateResult","fillBboxParam","element","console"],"mappings":";;;;;;;AA2BA,MAAMA,QAAQC,SAAS;AAEhB,eAAeC,KACpBC,eAAuB,EACvBC,IAOC;QAwCkBC,2BAiEhBC;IAvGH,MAAM,EAAEC,OAAO,EAAEC,WAAW,EAAEC,mBAAmB,EAAE,GAAGL;IACtD,MAAM,EAAEM,gBAAgB,EAAEC,IAAI,EAAE,GAAGJ;IAEnC,MAAM,EAAEK,SAAS,EAAEC,MAAM,EAAE,GAAGL;IAE9B,MAAM,EAAE,aAAaM,eAAe,EAAEC,WAAW,EAAE,GAAG,MAAMC,iBAC1DT,SACA;QAAEM;IAAO;IAEX,MAAMI,eAAe,MAAMC,2BAA2B;QACpD,aAAad,KAAK,WAAW;QAC7B,QAAQS;IACV;IAEA,IAAIM,eAAeT;IACnB,IAAIU,aAAaT,KAAK,KAAK;IAC3B,IAAIU,cAAcV,KAAK,MAAM;IAC7B,MAAMW,aAAaF;IACnB,MAAMG,cAAcF;IACpB,IAAIR,AAAW,cAAXA,QAAsB;QACxB,MAAMW,eAAe,MAAMC,4BAA4BN;QACvDC,aAAaI,aAAa,KAAK;QAC/BH,cAAcG,aAAa,MAAM;QACjCL,eAAeK,aAAa,WAAW;IACzC,OAAO,IAAIX,AAAW,eAAXA;SAKJ,IAAI,CAACA,QACVM,eAAe,MAAMO,kBAAkBhB,kBAAkBH,QAAQ,IAAI,EAAE;QACrE,OAAOa;QACP,QAAQC;IACV;IAGFM,mBAAmBhB,MAAMC;IAEzB,MAAMgB,aAAavB,AAAAA,SAAAA,CAAAA,4BAAAA,KAAK,mBAAmB,AAAD,IAAvBA,KAAAA,IAAAA,0BAA0B,QAAQ,EAAC,KAAK,EAAE;IAG7D,MAAMwB,mBAAiDzB,KAAK,aAAa,GACrE;QACA;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,CAAC,yBAAyB,EAAEA,KAAK,aAAa,CAAC,0BAA0B,CAAC;gBAClF;aACD;QACH;KACD,GACC,EAAE;IAEN,MAAM0B,cAA4C;QAChD;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,CAAC,kBAAkB,EAAE3B,gBAAgB,mBAAmB,CAAC;gBACjE;aACD;QACH;KACD;IAED,MAAM4B,OAAqC;QACzC;YAAE,MAAM;YAAU,SAASd;QAAa;WACrCY;WACAC;WACAF;QACH;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKT;wBACL,QAAQ;oBACV;gBACF;mBACIN,SACA,EAAE,GACD;oBACD;wBACE,MAAM;wBACN,MAAMC;oBACR;iBACD;aACJ;QACH;KACD;IAED,MAAM,EAAEkB,OAAO,EAAEC,KAAK,EAAE,GAAG,MAAMC,yBAC/BH,MACAI,aAAa,IAAI,EACjB3B;IAEF,MAAM4B,cAAcC,KAAK,SAAS,CAACL,SAASM,QAAW;IACvD,MAAMC,aAAaP;IAEnB,MAAMQ,UACHlC,AAAAA,CAAAA,SAAAA,CAAAA,qBAAAA,WAAW,MAAM,AAAD,IAAhBA,KAAAA,IAAAA,mBAAmB,IAAI,AAAD,IAAI;QAACiC,WAAW,MAAM;KAAC,GAAGA,WAAW,OAAM,KAAM,EAAE;IAC5E,MAAME,cAAkC;QACtC,GAAGF,UAAU;QACbC;QACAJ;QACAH;QACA,UAAUS,uBACRF,SACApC,KAAK,WAAW,EAChBmC,WAAW,KAAK;IAEpB;IAEAI,OAAOJ,YAAY;IAGnBC,QAAQ,OAAO,CAAC,CAACI;QACf,MAAMC,OAAOD,OAAO,IAAI;QACxB,MAAME,sBAAsB1C,KAAK,WAAW,CAAC,IAAI,CAC/C,CAACwC,SAAWA,OAAO,IAAI,KAAKC;QAG9B7C,MAAM,+BAA+B8C;QACrC,MAAMC,eAAeD,sBACjBE,4BAA4BF,oBAAoB,WAAW,IAC3D,EAAE;QAEN9C,MAAM,gBAAgB+C;QAEtBA,aAAa,OAAO,CAAC,CAACE;YACpB,MAAMC,eAAeN,OAAO,KAAK,CAACK,MAAM;YACxC,IAAIC,cACF,IAAIrC,QACF+B,OAAO,KAAK,CAACK,MAAM,GAAGE,cACpBD,cACA9B,YACAC,aACAC,YACAC,aACAV;iBAEG;gBACL,MAAMuC,UAAUrC,YAAYmC;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,8EACAlD;IAIJM,QAAAA,uBAAAA,oBAAqB,MAAM,CAAC;QAC1B,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAM2B;YACR;SACD;IACH;IACA3B,QAAAA,uBAAAA,oBAAqB,MAAM,CAAC;QAC1B,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAM;YACR;SACD;IACH;IAEA,OAAOgC;AACT"}

package/dist/es/ai-model/prompt/assertion.mjs

@@ -0,0 +1,31 @@
+const assertSchema = {
+    type: 'json_schema',
+    json_schema: {
+        name: 'assert',
+        strict: true,
+        schema: {
+            type: 'object',
+            properties: {
+                pass: {
+                    type: 'boolean',
+                    description: 'Whether the assertion passed or failed'
+                },
+                thought: {
+                    type: [
+                        'string',
+                        'null'
+                    ],
+                    description: 'The thought process behind the assertion'
+                }
+            },
+            required: [
+                'pass',
+                'thought'
+            ],
+            additionalProperties: false
+        }
+    }
+};
+export { assertSchema };
+
+//# sourceMappingURL=assertion.mjs.map

package/dist/es/ai-model/prompt/assertion.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"ai-model\\prompt\\assertion.mjs","sources":["webpack://@rpascene/core/./src/ai-model/prompt/assertion.ts"],"sourcesContent":["import type { ResponseFormatJSONSchema } from 'openai/resources/index';\n\nexport const assertSchema: ResponseFormatJSONSchema = {\n  type: 'json_schema',\n  json_schema: {\n    name: 'assert',\n    strict: true,\n    schema: {\n      type: 'object',\n      properties: {\n        pass: {\n          type: 'boolean',\n          description: 'Whether the assertion passed or failed',\n        },\n        thought: {\n          type: ['string', 'null'],\n          description: 'The thought process behind the assertion',\n        },\n      },\n      required: ['pass', 'thought'],\n      additionalProperties: false,\n    },\n  },\n};\n"],"names":["assertSchema"],"mappings":"AAEO,MAAMA,eAAyC;IACpD,MAAM;IACN,aAAa;QACX,MAAM;QACN,QAAQ;QACR,QAAQ;YACN,MAAM;YACN,YAAY;gBACV,MAAM;oBACJ,MAAM;oBACN,aAAa;gBACf;gBACA,SAAS;oBACP,MAAM;wBAAC;wBAAU;qBAAO;oBACxB,aAAa;gBACf;YACF;YACA,UAAU;gBAAC;gBAAQ;aAAU;YAC7B,sBAAsB;QACxB;IACF;AACF"}

package/dist/es/ai-model/prompt/common.mjs

@@ -0,0 +1,7 @@
+function bboxDescription(vlMode) {
+    if ('gemini' === vlMode) return 'box_2d bounding box for the target element, should be [ymin, xmin, ymax, xmax] normalized to 0-1000.';
+    return '2d bounding box as [xmin, ymin, xmax, ymax]';
+}
+export { bboxDescription };
+
+//# sourceMappingURL=common.mjs.map

package/dist/es/ai-model/prompt/common.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"ai-model\\prompt\\common.mjs","sources":["webpack://@rpascene/core/./src/ai-model/prompt/common.ts"],"sourcesContent":["import type { TVlModeTypes } from '@rpascene/shared/env';\nexport function bboxDescription(vlMode: TVlModeTypes | undefined) {\n  if (vlMode === 'gemini') {\n    return 'box_2d bounding box for the target element, should be [ymin, xmin, ymax, xmax] normalized to 0-1000.';\n  }\n  return '2d bounding box as [xmin, ymin, xmax, ymax]';\n}\n"],"names":["bboxDescription","vlMode"],"mappings":"AACO,SAASA,gBAAgBC,MAAgC;IAC9D,IAAIA,AAAW,aAAXA,QACF,OAAO;IAET,OAAO;AACT"}

package/dist/es/ai-model/prompt/describe.mjs

@@ -0,0 +1,44 @@
+import { getPreferredLanguage } from "@rpascene/shared/env";
+const elementDescriberInstruction = ()=>`
+Describe the element in the red rectangle for precise identification. Use ${getPreferredLanguage()}.
+
+CRITICAL REQUIREMENTS:
+1. UNIQUENESS: The description must uniquely identify this element on the current page
+2. UNIVERSALITY: Use generic, reusable selectors that work across different contexts
+3. PRECISION: Be specific enough to distinguish from similar elements
+
+DESCRIPTION STRUCTURE:
+1. Element type (button, input, link, div, etc.)
+2. Primary identifier (in order of preference):
+   - Unique text content: "with text 'Login'"
+   - Unique attribute: "with aria-label 'Search'"
+   - Unique class/ID: "with class 'primary-button'"
+   - Unique position: "in header navigation"
+3. Secondary identifiers (if needed for uniqueness):
+   - Visual features: "blue background", "with icon"
+   - Relative position: "below search bar", "in sidebar"
+   - Parent context: "in login form", "in main menu"
+
+GUIDELINES:
+- Keep description under 25 words
+- Prioritize semantic identifiers over visual ones
+- Use consistent terminology across similar elements
+- Avoid page-specific or temporary content
+- Don't mention the red rectangle or selection box
+- Focus on stable, reusable characteristics
+
+EXAMPLES:
+- "Login button with text 'Sign In'"
+- "Search input with placeholder 'Enter keywords'"
+- "Navigation link with text 'Home' in header"
+- "Submit button in contact form"
+- "Menu icon with aria-label 'Open menu'"
+
+Return JSON:
+{
+  "description": "unique element identifier",
+  "error"?: "error message if any"
+}`;
+export { elementDescriberInstruction };
+
+//# sourceMappingURL=describe.mjs.map

package/dist/es/ai-model/prompt/describe.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"ai-model\\prompt\\describe.mjs","sources":["webpack://@rpascene/core/./src/ai-model/prompt/describe.ts"],"sourcesContent":["import { getPreferredLanguage } from '@rpascene/shared/env';\n\nexport const elementDescriberInstruction = () => {\n  return `\nDescribe the element in the red rectangle for precise identification. Use ${getPreferredLanguage()}.\n\nCRITICAL REQUIREMENTS:\n1. UNIQUENESS: The description must uniquely identify this element on the current page\n2. UNIVERSALITY: Use generic, reusable selectors that work across different contexts\n3. PRECISION: Be specific enough to distinguish from similar elements\n\nDESCRIPTION STRUCTURE:\n1. Element type (button, input, link, div, etc.)\n2. Primary identifier (in order of preference):\n   - Unique text content: \"with text 'Login'\"\n   - Unique attribute: \"with aria-label 'Search'\"\n   - Unique class/ID: \"with class 'primary-button'\"\n   - Unique position: \"in header navigation\"\n3. Secondary identifiers (if needed for uniqueness):\n   - Visual features: \"blue background\", \"with icon\"\n   - Relative position: \"below search bar\", \"in sidebar\"\n   - Parent context: \"in login form\", \"in main menu\"\n\nGUIDELINES:\n- Keep description under 25 words\n- Prioritize semantic identifiers over visual ones\n- Use consistent terminology across similar elements\n- Avoid page-specific or temporary content\n- Don't mention the red rectangle or selection box\n- Focus on stable, reusable characteristics\n\nEXAMPLES:\n- \"Login button with text 'Sign In'\"\n- \"Search input with placeholder 'Enter keywords'\"\n- \"Navigation link with text 'Home' in header\"\n- \"Submit button in contact form\"\n- \"Menu icon with aria-label 'Open menu'\"\n\nReturn JSON:\n{\n  \"description\": \"unique element identifier\",\n  \"error\"?: \"error message if any\"\n}`;\n};\n"],"names":["elementDescriberInstruction","getPreferredLanguage"],"mappings":";AAEO,MAAMA,8BAA8B,IAClC,CAAC;0EACgE,EAAEC,uBAAuB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAsClG,CAAC"}