{"id":53734,"date":"2025-02-16T15:06:30","date_gmt":"2025-02-16T07:06:30","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53734\/"},"modified":"2025-02-16T15:06:30","modified_gmt":"2025-02-16T07:06:30","slug":"extractthinkergemini-2-0","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53734\/","title":{"rendered":"ExtractThinker+Gemini 2.0"},"content":{"rendered":"<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8 Google \u7684 Gemini 2.0 \u6a21\u578b\u5982\u4f55\u4e0e ExtractThinker\uff08\u4e00\u79cd\u65e8\u5728\u534f\u8c03 OCR\u3001\u5206\u7c7b\u3001\u6587\u6863\u62c6\u5206\u548c\u6570\u636e\u63d0\u53d6\u7ba1\u9053\u7684\u5f00\u6e90\u6846\u67b6\uff09\u7ed3\u5408\u4f7f\u7528\uff0c\u4ece\u800c\u589e\u5f3a\u667a\u80fd\u6587\u6863\u5904\u7406 (IDP)\u3002\u6211\u4eec\u5c06\u4ecb\u7ecd Google Document AI \u5982\u4f55\u878d\u5165\u5176\u4e2d\uff0c\u4ee5\u53ca Gemini 2.0 Flash \u7684\u65b0\u529f\u80fd\uff0c\u5e76\u901a\u8fc7\u4ee3\u7801\u793a\u4f8b\u548c\u5b9a\u4ef7\u89c1\u89e3\u603b\u7ed3\u6240\u6709\u5185\u5bb9\u3002<\/p>\n<h2>1\u3001\u7b80\u4ecb<\/h2>\n<p>\u667a\u80fd\u6587\u6863\u5904\u7406 (IDP) \u662f\u5c06\u975e\u7ed3\u6784\u5316\u6570\u636e\uff08\u5982\u53d1\u7968\u3001\u9a7e\u9a76\u6267\u7167\u548c\u62a5\u544a\uff09\u8f6c\u6362\u4e3a\u7ed3\u6784\u5316\u3001\u53ef\u64cd\u4f5c\u4fe1\u606f\u7684\u5173\u952e\u5de5\u4f5c\u6d41\u7a0b\u3002\u867d\u7136\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u73b0\u5728\u53ef\u4ee5\u76f4\u63a5\u5904\u7406\u56fe\u50cf\u548c PDF\uff0c\u4f46\u4ec5\u4ec5\u5c06\u56fe\u50cf\u8f93\u5165 LLM \u5e76\u5e0c\u671b\u83b7\u5f97\u5b8c\u7f8e\u7684\u7ed3\u679c\u901a\u5e38\u662f\u4e0d\u591f\u7684\u3002\u76f8\u53cd\uff0c\u5f3a\u5927\u7684 IDP \u7ba1\u9053\u7ed3\u5408\u4e86\uff1a<\/p>\n<ul>\n<li>OCR \u6216\u5176\u4ed6\u5e03\u5c40\u63d0\u53d6\u5de5\u5177\uff08\u5982 Google Document AI\u3001Tesseract \u6216 PyPDF\uff09\u3002<\/li>\n<li>\u5206\u7c7b\u4ee5\u8bc6\u522b\u6587\u6863\u7c7b\u578b\uff08\u53d1\u7968\u3001\u5408\u540c\u3001\u8bb8\u53ef\u8bc1\u7b49\uff09\u3002<\/li>\n<li>\u62c6\u5206\u4ee5\u5904\u7406\u5927\u578b\u7ec4\u5408\u6587\u4ef6\u5e76\u5c06\u5176\u5206\u89e3\u4e3a\u903b\u8f91\u90e8\u5206\u3002<\/li>\n<li>\u63d0\u53d6\u4ee5\u5c06\u4fe1\u606f\u6620\u5c04\u5230\u7ed3\u6784\u5316\u7684 Pydantic \u6a21\u578b\u4e2d \u2014 \u4f8b\u5982\u63d0\u53d6\u53d1\u7968\u53f7\u3001\u65e5\u671f\u3001\u603b\u91d1\u989d\u6216\u89e3\u91ca\u56fe\u8868\u6570\u636e\u3002<\/li>\n<\/ul>\n<p>ExtractThinker \u662f\u4e00\u4e2a\u5f00\u7bb1\u5373\u7528\u7684\u5e93\uff0c\u53ef\u8ba9\u4f60\u5c06\u5b83\u4eec\u4e0e Google \u5168\u65b0\u7684 Gemini 2.0 \u6a21\u578b\u65e0\u7f1d\u96c6\u6210\u3002<\/p>\n<h2>2\u3001Google Document AI<\/h2>\n<p>\u5728\u6df1\u5165\u7814\u7a76\u57fa\u4e8e LLM \u7684\u63d0\u53d6\u4e4b\u524d\uff0c\u8ba9\u6211\u4eec\u5148\u8c08\u8c08 Google Document AI\u3002\u8fd9\u662f Google Cloud \u63d0\u4f9b\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u63d0\u4f9b OCR\u3001\u7ed3\u6784\u89e3\u6790\u3001\u5206\u7c7b\u548c\u4e13\u7528\u57df\u63d0\u53d6\u5668\uff08\u4f8b\u5982\u53d1\u7968\u89e3\u6790\u3001W2 \u8868\u683c\u3001\u94f6\u884c\u5bf9\u8d26\u5355\u7b49\uff09\u3002<\/p>\n<p>Document AI \u63d0\u4f9b\uff1a<\/p>\n<ul>\n<li>\u6587\u6863 OCR \u6bcf 1,000 \u9875 1.50 \u7f8e\u5143\uff08\u6bcf\u6708\u6700\u591a 500 \u4e07\u9875\uff09\uff0c\u6279\u91cf\u8d8a\u5927\u53ef\u4eab\u53d7\u66f4\u591a\u6298\u6263\u3002<\/li>\n<li>\u8868\u5355\u89e3\u6790\u5668\u548c\u81ea\u5b9a\u4e49\u63d0\u53d6\u5668\uff0c\u6bcf 1,000 \u9875 30 \u7f8e\u5143\uff08\u6bcf\u6708 100 \u4e07\u9875\u540e\u6709\u6298\u6263\uff09\u3002<\/li>\n<li>\u5e03\u5c40\u89e3\u6790\u5668\uff0c\u6bcf 1,000 \u9875 10 \u7f8e\u5143\u3002<\/li>\n<li>\u9884\u8bad\u7ec3\u7684\u4e13\u7528\u5904\u7406\u5668\uff08\u5982\u7f8e\u56fd\u9a7e\u7167\u89e3\u6790\u5668\u6216\u53d1\u7968\u89e3\u6790\u5668\uff09\uff0c\u6bcf\u6587\u6863\u6216\u6bcf\u9875\u6536\u8d39\uff08\u4f8b\u5982\uff0c\u53d1\u7968\u89e3\u6790\u6bcf 10 \u9875 0.10 \u7f8e\u5143\uff09\u3002<\/li>\n<\/ul>\n<p>\u4f7f\u7528 ExtractThinker \u65f6\uff0c\u4f60\u53ef\u4ee5\u9644\u52a0 DocumentLoaderDocumentAI\uff0c\u4ee5\u7edf\u4e00 Document AI OCR \u6216\u57fa\u4e8e LLM \u7684\u7ba1\u9053\u8fdb\u884c\u8868\u5355\u89e3\u6790\u3002\u534f\u540c\u4f5c\u7528\u975e\u5e38\u5f3a\u5927\uff1aDocument AI \u53ef\u9760\u5730\u63d0\u53d6\u6587\u672c\uff0c\u800c Gemini \u6216\u5176\u4ed6\u6a21\u578b\u89e3\u91ca\u8be5\u6587\u672c\uff08\u52a0\u4e0a\u56fe\u50cf\uff09\u4ee5\u751f\u6210\u9ad8\u7ea7\u7ed3\u6784\u5316\u8f93\u51fa\u3002<\/p>\n<p>\u4f60\u53ef\u4ee5\u4f7f\u7528\u4efb\u4f55\u5904\u7406\u5668\uff0c\u4f46\u53ea\u80fd\u4f7f\u7528\u6587\u6863 OCR \u6216\u5e03\u5c40\u89e3\u6790\u5668\u3002\u6587\u6863 OCR \u5e94\u4e0e\u89c6\u89c9\u914d\u5bf9\uff0c\u5f53\u89c6\u89c9\u4e0d\u53ef\u7528\u65f6\u5e94\u4f7f\u7528\u5e03\u5c40\u89e3\u6790\u5668\u3002\u5982\u679c\u53ef\u80fd\u7684\u8bdd\uff0c\u89c6\u89c9\u662f\u9996\u9009\uff0c\u56e0\u4e3a\u5b83\u4f1a\u4e3a LLM \u63d0\u4f9b\u5927\u91cf\u80cc\u666f\u4fe1\u606f\uff0c\u4f46\u4f60\u53ef\u4ee5\u4f7f\u7528 Layout \u6765\u5b8c\u6210\u989d\u5916\u7684\u5de5\u4f5c\u3002\u4f60\u4e5f\u53ef\u4ee5\u53ea\u4f7f\u7528 Gemini\uff0c\u5b83\u53ea\u4f1a\u4f7f\u7528\u89c6\u89c9\u8fdb\u884c\u8bfb\u53d6\u3002<\/p>\n<blockquote><p>\n  \u63d0\u793a\uff1a\u5982\u679c\u4f60\u53ea\u60f3\u4f7f\u7528\u201c\u7eaf LLM \u65b9\u6cd5\u201d\u5c06\u6587\u6863\u8bfb\u53d6\u4e3a\u56fe\u50cf\uff0c\u53ef\u80fd\u4f1a\u9047\u5230\u5e7b\u89c9\u6216\u626b\u63cf\u7cbe\u5ea6\u5dee\u7684\u95ee\u9898\u3002\u5c06 Document AI \u4e0e LLM \u76f8\u7ed3\u5408\u901a\u5e38\u4f1a\u4ea7\u751f\u66f4\u7cbe\u786e\u3001\u66f4\u5177\u6210\u672c\u6548\u76ca\u7684\u7ed3\u679c\u3002\n<\/p><\/blockquote>\n<h2>3\u3001Gemini 2.0<\/h2>\n<p>Google \u7684 Gemini 2.0 \u662f\u5176\u591a\u6a21\u6001\u6a21\u578b\u7cfb\u5217\u7684\u4e0b\u4e00\u4e2a\u53d1\u5c55\uff0c\u652f\u6301\u6587\u672c\u3001\u56fe\u50cf\u3001\u97f3\u9891\u4ee5\u53ca\u9ad8\u7ea7\u201c\u4ee3\u7406\u201d\u529f\u80fd\u3002\u5728 Gemini 2.0 \u4e2d\uff0c\u6709\u591a\u4e2a\u53d8\u4f53\uff1a<\/p>\n<ul>\n<li>Gemini 2.0 Flash\uff1a\u4e00\u79cd\u5b9e\u9a8c\u6027\u4f46\u9ad8\u901f\u7684\u6a21\u578b\uff0c\u5177\u6709\u5f3a\u5927\u7684 IDP \u5de5\u4f5c\u6d41\u6027\u80fd\u3002\u975e\u5e38\u9002\u5408\u5feb\u901f\u4ece\u6587\u6863\u4e2d\u63d0\u53d6\u6570\u636e\uff0c\u5e76\u4ee5\u4f4e\u5ef6\u8fdf\u5927\u89c4\u6a21\u5904\u7406\u6587\u672c\u6216\u56fe\u50cf\u3002<\/li>\n<li>Gemini 2.0 Thinking\uff1a\u4e00\u79cd\u66f4\u201c\u6ce8\u91cd\u63a8\u7406\u201d\u7684\u6a21\u578b\uff0c\u5b83\u80fd\u591f\u5904\u7406\u6781\u5176\u590d\u6742\u7684\u4efb\u52a1\uff0c\u5177\u6709\u66f4\u6df1\u7684\u601d\u8def\u548c\u5de5\u5177\u4f7f\u7528\u3002<\/li>\n<\/ul>\n<p>\u53ef\u4ee5\u5c06 Gemini 2.0 \u4e0e ExtractThinker \u4e00\u8d77\u4f7f\u7528\u3002Gemini 2.0 Flash \u7279\u522b\u9002\u5408 IDP\uff0c\u56e0\u4e3a\uff1a<\/p>\n<ul>\n<li>\u5b83\u901f\u5ea6\u66f4\u5feb\uff0c\u975e\u5e38\u9002\u5408\u5927\u89c4\u6a21\u63d0\u53d6\u4efb\u52a1\u3002<\/li>\n<li>\u5b83\u652f\u6301\u591a\u6a21\u5f0f\uff08\u56fe\u50cf + \u6587\u672c\uff09\u8f93\u5165\uff0c\u8fd9\u5bf9\u4e8e\u9605\u8bfb\u626b\u63cf\u7684\u9875\u9762\u6216\u56fe\u8868\u81f3\u5173\u91cd\u8981\u3002<\/li>\n<li>\u5b83\u53ef\u4ee5\u5f88\u597d\u5730\u5904\u7406\u7b80\u5355\u7684\u5206\u7c7b\u548c\u7ed3\u6784\u5316\u8f93\u51fa\uff0c\u6210\u672c\u4f4e\u4e8e\u201c\u9ad8\u9636\u201d\u6a21\u578b\u3002<\/li>\n<\/ul>\n<h2>4\u3001ExtractThinker\uff1aLLM \u7684 IDP<\/h2>\n<p>ExtractThinker \u662f\u4e00\u4e2a\u7075\u6d3b\u7684\u5e93\uff0c\u5b83\u62bd\u8c61\u4e86\u6784\u5efa\u667a\u80fd\u6587\u6863\u5904\u7406\u6d41\u7a0b\u7684\u590d\u6742\u6027\u3002\u5b83\u53ef\u4ee5\u5e2e\u52a9\u4f60\uff1a<\/p>\n<ul>\n<li>\u901a\u8fc7\u4e0d\u540c\u7684 DocumentLoader\uff08Tesseract\u3001PyPDF\u3001Google Document AI\u3001AWS Textract \u7b49\uff09\u52a0\u8f7d\u6587\u6863\u3002<\/li>\n<li>\u5c06\u7ec4\u5408\u6587\u6863\u62c6\u5206\u4e3a\u5355\u72ec\u7684\u6587\u6863\uff08\u4f8b\u5982\uff0c\u5c06\u4e24\u4e2a\u6587\u6863\u62c6\u5206\u4e3a\u4e00\u4e2a\u6587\u6863\uff09\u3002<\/li>\n<li>\u4f7f\u7528\u591a\u79cd\u7b56\u7565\u548c\u65b9\u6cd5\u5bf9\u6bcf\u4e2a\u6587\u6863\u8fdb\u884c\u5206\u7c7b<\/li>\n<li>\u4f7f\u7528\u57fa\u4e8e Pydantic \u7684\u201c\u5408\u7ea6\u201d\u63d0\u53d6\u7ed3\u6784\u5316\u6570\u636e\u3002<\/li>\n<\/ul>\n<p>\u4ee5\u4e0b\u662f ExtractThinker \u901a\u5e38\u5982\u4f55\u5904\u7406 IDP \u5de5\u4f5c\u6d41\u7a0b\u7684\u9ad8\u7ea7\u793a\u610f\u56fe\uff1a<\/p>\n<p>  ExtractThinker \u6d41\u7a0b\u200b\u200b <\/p>\n<p>\u4f46\u9996\u5148\u8bf7\u5148\u5b89\u88c5\u5b83\uff1a<\/p>\n<pre><code>pip install extract-thinker<\/code><\/pre>\n<h3>4.1 \u6587\u6863\u52a0\u8f7d\u5668<\/h3>\n<p>\u4e0b\u9762\u662f\u4e00\u4e2a\u6700\u5c0f\u4ee3\u7801\u7247\u6bb5\uff0c\u5c55\u793a\u4e86\u5982\u4f55\u4f7f\u7528 Google Document AI \u4f5c\u4e3a <code>DocumentLoader<\/code> \u52a0\u8f7d PDF\uff1a<\/p>\n<pre><code>from extract_thinker.document_loader.document_loader_google_document_ai import DocumentLoaderDocumentAI\n\n# Initialize the DocumentLoader for Google Document AI\ndoc_loader = DocumentLoaderDocumentAI(\n    project_id=\"YOUR_PROJECT_ID\",\n    location=\"us\", # or eu\n    processor_id=\"YOUR_PROCESSOR_ID\",\n    credentials=\"path\/to\/google_credentials.json\"\n)\n\n# Now load or extract:\npdf_path = \"path\/to\/your\/bulk_documents.pdf\"\n# You can directly call:\npages = doc_loader.load(pdf_path)<\/code><\/pre>\n<h3>4.2 \u63d0\u53d6<\/h3>\n<p>\u5b9a\u4e49 <code>DocumentLoader<\/code> \u540e\uff0c\u4e0b\u4e00\u6b65\u5c31\u662f\u63d0\u53d6\u7ed3\u6784\u5316\u6570\u636e\u3002ExtractThinker \u901a\u8fc7\u57fa\u4e8e Pydantic \u7684 Contract \u5b9e\u73b0\u6b64\u76ee\u7684 &#8211; \u4e00\u79cd\u63cf\u8ff0\u4ece\u6587\u6863\u4e2d\u63d0\u53d6\u54ea\u4e9b\u5b57\u6bb5\u7684\u67b6\u6784\u3002\u65e0\u8bba\u4f60\u662f\u4ece\u53d1\u7968\u4e2d\u89e3\u6790\u884c\u9879\u76ee\u8fd8\u662f\u4ece\u9a7e\u9a76\u6267\u7167\u4e2d\u89e3\u6790\u5b57\u6bb5\uff0c\u5de5\u4f5c\u6d41\u7a0b\u90fd\u4fdd\u6301\u4e00\u81f4\u3002<\/p>\n<pre><code>from extract_thinker import Contract\nfrom extract_thinker import Extractor\nfrom pydantic import Field\nfrom typing import List\n\nclass InvoiceLineItem(Contract):\n    description: str = Field(description=\"Description of the item\")\n    quantity: int = Field(description=\"Quantity of items purchased\")\n    unit_price: float = Field(description=\"Price per unit\")\n    amount: float = Field(description=\"Total amount for this line\")\n\nclass InvoiceContract(Contract):\n    invoice_number: str = Field(description=\"Unique invoice identifier\")\n    invoice_date: str = Field(description=\"Date of the invoice\")\n    total_amount: float = Field(description=\"Overall total amount\")\n    line_items: List[InvoiceLineItem] = Field(description=\"List of items in this invoice\")<\/code><\/pre>\n<p>\u7136\u540e\u6211\u4eec\u7ee7\u7eed\u6267\u884c Extractor\u3002\u4f7f\u7528 <code>DocumentLoader<\/code> \u5c06 Gemini 2.0 Flash \u6307\u5b9a\u4e3a LLM\u3002\u53ef\u4ee5\u968f\u610f\u66f4\u6362\u5176\u4ed6\u52a0\u8f7d\u5668\uff0c\u4f8b\u5982 <code>DocumentLoaderTesseract<\/code> \u6216 <code>DocumentLoaderAzureForm<\/code> \uff1a<\/p>\n<pre><code># Create Extractor &amp; attach the loader\nextractor = Extractor()\nextractor.load_document_loader(doc_loader)\n\n# Assign Gemini 2.0 Flash model for extraction\nextractor.load_llm(\"vertex_ai\/gemini-2.0-flash-exp\")<\/code><\/pre>\n<p>\u7136\u540e\uff0c\u5bf9\u5176\u8fdb\u884c\u5904\u7406\u4ee5\u8fdb\u884c\u63d0\u53d6\u3002\u60a8\u53ea\u9700\u4f20\u9012\u8def\u5f84\u6216\u6d41\u4ee5\u53ca\u5b9a\u4e49\u7684\u5408\u540c\u3002\u4f60\u8fd8\u6709\u5176\u4ed6\u53ef\u9009\u5b57\u6bb5\uff0c\u4f8b\u5982 vision\uff0c\u5982\u679c\u662f vision \u6a21\u578b\uff0c\u5b83\u4f1a\u5c06\u9875\u9762\u8f6c\u6362\u4e3a\u8981\u5728\u6a21\u578b\u5185\u90e8\u4f7f\u7528\u7684\u56fe\u50cf\u3002<\/p>\n<pre><code>extracted_invoice = extractor.extract(\n    source=test_file_path,\n    response_model=InvoiceContract,\n    vision=False                     \n)\n\n# Access the structured data\nprint(\"Invoice Number:\", extracted_invoice.invoice_number)\nprint(\"Invoice Date:\", extracted_invoice.invoice_date)\nprint(\"Total Amount:\", extracted_invoice.total_amount)\nfor item in extracted_invoice.line_items:\n    print(f\"Item {item.description}: x{item.quantity} at ${item.unit_price} each.\")<\/code><\/pre>\n<h3>4.3 \u5206\u7c7b<\/h3>\n<p>\u5047\u8bbe\u6211\u4eec\u60f3\u5c06\u6587\u6863\u5206\u7c7b\u4e3a\u201c\u8f66\u8f86\u767b\u8bb0\u201d\u6216\u201c\u9a7e\u9a76\u6267\u7167\u201d\u3002\u6211\u4eec\u53ef\u4ee5\u4e3a\u6bcf\u79cd\u7c7b\u578b\u5b9a\u4e49\u57fa\u4e8e Pydantic \u7684\u5408\u540c\uff0c\u5e76\u5c06\u5b83\u4eec\u6620\u5c04\u5230 <code>ExtractThinker<\/code> \u4e2d\u7684\u5206\u7c7b\u5bf9\u8c61\u3002\u50cf Gemini 2.0 Flash \u8fd9\u6837\u7684\u5feb\u901f\u6a21\u578b\u53ef\u4ee5\u5904\u7406\u5927\u591a\u6570\u6587\u6863\uff0c\u4f46\u5982\u679c\u5176\u7f6e\u4fe1\u5ea6\u4f4e\u4e8e\u9608\u503c\uff0c\u6211\u4eec\u53ef\u4ee5\u5347\u7ea7\u5230\u66f4\u5f3a\u5927\u4e14\u66f4\u6162\u7684 Gemini 2.0 Thinking\u3002ExtractThinker \u7684\u591a\u5c42\u65b9\u6cd5\u53ef\u4ee5\u81ea\u52a8\u5316\u8fd9\u79cd\u56de\u9000\u903b\u8f91\uff0c\u5728\u6210\u672c\u6548\u7387\u548c\u7a33\u5065\u51c6\u786e\u6027\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\u3002<\/p>\n<p>  \u5e26\u56de\u9000\u7684\u5206\u7c7b\u6d41\u7a0b <\/p>\n<p>\u4e0b\u9762\u662f\u4e00\u4e2a\u7b80\u5316\u7684\u56fe\u793a\uff0c\u5176\u4e2d\u5305\u542b\u4e24\u4e2a\u5206\u7c7b\u5bf9\u8c61\uff1a\u4e00\u4e2a\u7528\u4e8e\u201c\u9a7e\u9a76\u6267\u7167\u201d\uff0c\u53e6\u4e00\u4e2a\u7528\u4e8e\u201c\u8f66\u8f86\u767b\u8bb0\u201d\uff0c \u4f60\u53ef\u4ee5\u7c7b\u4f3c\u5730\u5b9a\u4e49\u201c\u53d1\u7968\u201d\u3001\u201c\u4fe1\u7528\u7968\u636e\u201d\u7b49\u3002<\/p>\n<pre><code>from extract_thinker.models.classification import Classification\nfrom tests.models.driver_license import DriverLicense\nfrom tests.models.vehicle_registration import VehicleRegistration\n\ndriver_license_class = Classification(\n    name=\"Driver License\",\n    description=\"A document representing a driver's license\",\n    contract=DriverLicense\n)\n\nvehicle_registration_class = Classification(\n    name=\"Vehicle Registration\",\n    description=\"Official document registering a vehicle ownership\",\n    contract=VehicleRegistration\n)\n\nmy_classifications = [driver_license_class, vehicle_registration_class]<\/code><\/pre>\n<p>\u4e3a\u4e86\u5b9e\u73b0\u540e\u5907\u65b9\u6cd5\uff0c\u6211\u4eec\u8bbe\u7f6e\u4e86\u4e24\u4e2a Extractor \u5b9e\u4f8b \u2014 \u4e00\u4e2a\u7531 Gemini 2.0 Flash\uff08\u5feb\u901f\uff09\u63d0\u4f9b\u652f\u6301\uff0c\u53e6\u4e00\u4e2a\u7531 Gemini 2.0 Thinking\uff08\u540e\u5907\uff09\u63d0\u4f9b\u652f\u6301\u3002\u6bcf\u4e2a\u63d0\u53d6\u5668\u90fd\u4e0e\u4e00\u4e2a\u53ef\u8bfb\u53d6 PDF \u6216\u56fe\u50cf\u7684 <code>DocumentLoader<\/code> \u76f8\u5173\u8054\uff0c\u4f8b\u5982 <code>DocumentLoaderPyPdf<\/code>\u3002<\/p>\n<pre><code>from extract_thinker import Extractor\nfrom extract_thinker.document_loader.document_loader_pypdf import DocumentLoaderPyPdf\nfrom extract_thinker.process import Process, ClassificationStrategy\nfrom extract_thinker.models.classification import Classification\n\n# 1. Define your DocumentLoader\npdf_loader = DocumentLoaderPyPdf()\n\n# 2. Create two Extractors: fast &amp; fallback\nflash_extractor = Extractor(pdf_loader)\nflash_extractor.load_llm(\"vertex_ai\/gemini-2.0-flash-exp\")\n\nthinking_extractor = Extractor(pdf_loader)\nthinking_extractor.load_llm(\"vertex_ai\/gemini-2.0-thinking-exp\")\n\n# 3. Define Classifications (e.g. \"Vehicle Registration\" &amp; \"Driver License\")\nvehicle_registration = Classification(name=\"Vehicle Registration\", description=\"...\")\ndriver_license = Classification(name=\"Driver License\", description=\"...\")\nmy_classifications = [vehicle_registration, driver_license]\n\n# 4. Build a Process and add both extractors in separate layers\nprocess = Process()\nprocess.add_classify_extractor([\n    [flash_extractor],     # Layer 1\n    [thinking_extractor]   # Layer 2 (fallback)\n])\n\n# 5. Perform classification with CONSENSUS_WITH_THRESHOLD at 0.9\npdf_path = \"path\/to\/document.pdf\"\nresult = process.classify(\n    pdf_path,\n    my_classifications,\n    strategy=ClassificationStrategy.CONSENSUS_WITH_THRESHOLD,\n    threshold=9,  # i.e., 0.9 confidence,\n    image=False # if vision is allowed, you can add it in the classification\n)\n\nprint(\"Classified as:\", result.name)\nprint(\"Confidence:\", result.confidence)<\/code><\/pre>\n<h3>4.4 \u4f7f\u7528\u62c6\u5206\u5668\u62c6\u5206\u6587\u6863<\/h3>\n<p>\u8bb8\u591a IDP \u5de5\u4f5c\u6d41\u7a0b\u6d89\u53ca\u591a\u9875 PDF \u6216\u6df7\u5408\u6587\u6863\u96c6\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a\u6587\u4ef6\u53ef\u80fd\u540c\u65f6\u5305\u542b\u53d1\u7968\u548c\u9a7e\u9a76\u6267\u7167\u3002ExtractThinker \u63d0\u4f9b\u62c6\u5206\u7b56\u7565\u6765\u81ea\u52a8\u5206\u5272\u6587\u6863\u3002\u4e24\u79cd\u4e3b\u8981\u7b56\u7565\u662f\uff1a<\/p>\n<ul>\n<li>EAGER\uff1a\u4e00\u6b21\u5904\u7406\u6574\u4e2a\u6587\u4ef6\uff0c\u9884\u5148\u786e\u5b9a\u6240\u6709\u62c6\u5206\u70b9\u3002<\/li>\n<li>LAZY\uff1a\u9010\u6b65\u6bd4\u8f83\u9875\u9762\uff0c\u51b3\u5b9a\u5728\u54ea\u91cc\u62c6\u5206\u3002<\/li>\n<\/ul>\n<p>\u4e0b\u9762\uff0c\u6211\u4eec\u7528\u4e00\u4e2a\u5047\u8bbe\u7684 PDF \u6f14\u793a\u4e86 EAGER \u62c6\u5206\uff0c\u8be5 PDF \u7ed3\u5408\u4e86\u4e0d\u540c\u7684\u5f62\u5f0f\uff1a<\/p>\n<pre><code>from extract_thinker.process import Process\nfrom extract_thinker.splitter import SplittingStrategy\nfrom extract_thinker.image_splitter import ImageSplitter\n\n# 1. Prepare a Process\nprocess = Process()\n\n# 2. Assign a DocumentLoader (e.g., Tesseract, PyPdf, etc.) or an Extractor later\n# Here we do it at the extractor level or process level:\n# process.load_document_loader(my_loader)\n\n# 3. Specify which Splitter to use\nimage_splitter = ImageSplitter(model=\"vertex_ai\/gemini-2.0-flash-exp\")\nprocess.load_splitter(image_splitter)\n\n# 4. Provide classifications\u2014like \"Invoice\" vs \"Driver License\"\n# (already defined as my_classifications or from a tree)<\/code><\/pre>\n<p>\u5728\u8fd0\u884c\u65f6\uff0c EAGER \u62c6\u5206\u5c06\u626b\u63cf\u6574\u4e2a\u6587\u6863\uff0c\u68c0\u6d4b\u903b\u8f91\u8fb9\u754c\uff08\u57fa\u4e8e\u5185\u5bb9\u5dee\u5f02\uff09\uff0c\u5e76\u521b\u5efa\u8f83\u5c0f\u7684\u201c\u5b50\u6587\u6863\u201d\uff0c\u7136\u540e\u53ef\u4ee5\u5bf9\u6bcf\u4e2a\u5b50\u6587\u6863\u8fdb\u884c\u5206\u7c7b\u548c\u63d0\u53d6\u3002<\/p>\n<p>\u8981\u5c06\u5b83\u4eec\u653e\u5728\u4e00\u8d77\uff0c\u4f60\u9700\u8981\uff1a<\/p>\n<ul>\n<li>\u4e00\u4e2a\u8fc7\u7a0b\uff08\u5e26\u6709\u5df2\u52a0\u8f7d\u7684\u62c6\u5206\u5668\uff09\u3002<\/li>\n<li>\u7528\u4e8e\u8bc6\u522b\u6bcf\u4e2a\u9875\u9762\u7684\u5206\u7c7b\u3002<\/li>\n<li>\u5df2\u5206\u914d\u63d0\u53d6\u5668\u6216 LLM\u3002<\/li>\n<\/ul>\n<p>\u4f60\u53ef\u4ee5\u5728\u4e00\u6b21\u94fe\u5f0f\u8c03\u7528\u4e2d\u52a0\u8f7d\u6587\u4ef6\u3001\u62c6\u5206\u6587\u4ef6\u5e76\u63d0\u53d6\u6240\u6709\u5185\u5bb9\uff1a<\/p>\n<pre><code>from extract_thinker.models.splitting_strategy import SplittingStrategy\n\nBULK_DOC_PATH = \"path\/to\/combined_documents.pdf\"\n\nresult = process.load_file(BULK_DOC_PATH)\n    .split(my_classifications, strategy=SplittingStrategy.EAGER)\n    .extract(vision=True)\n\n# 'result' is a list of extracted objects, each matching a classification's contract\nfor doc_content in result:\n    print(f\"Extracted document type: {type(doc_content).__name__}\")\n    print(doc_content.json(indent=2))<\/code><\/pre>\n<ul>\n<li><code>load_file(...)<\/code>: \u52a0\u8f7d\u7ec4\u5408 PDF\u3002<\/li>\n<li><code>split(...)<\/code>: \u4f7f\u7528 EAGER \u7b56\u7565\u5bf9\u5185\u5bb9\u8fdb\u884c\u5206\u6bb5\uff0c\u7531 Splitter \u6a21\u578b\u548c\u4f60\u7684\u5206\u7c7b\u6307\u5bfc\u3002<\/li>\n<li><code>extract(...)<\/code>\uff1a\u8c03\u7528\u4f60\u9009\u62e9\u7684 LLM \u5c06\u6bcf\u4e2a\u62c6\u5206\u5757\u89e3\u6790\u4e3a\u7ed3\u6784\u5316\u7684 Pydantic \u6a21\u578b\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u6709\u6548\u5730\u5904\u7406\u5927\u578b\u6216\u591a\u6587\u6863\u8f93\u5165\uff0c\u786e\u4fdd\u6bcf\u4e2a\u5b50\u6587\u6863\u90fd\u5f97\u5230\u6b63\u786e\u5206\u7c7b\uff0c\u7136\u540e\u7528\u6700\u5c11\u7684\u989d\u5916\u4ee3\u7801\u63d0\u53d6\u3002<\/p>\n<p>\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528 <code>ImageSplitter<\/code>\uff0c\u4f46\u5c31 Flash-Thinking \u800c\u8a00\uff0c\u76ee\u524d\u4e0d\u652f\u6301\u56fe\u50cf\u3002\u4f60\u53ef\u4ee5\u6539\u7528 <code>TextSplitter<\/code>\u3002<\/p>\n<blockquote><p>\n  Document AI &nbsp;Splitter &nbsp;vs. ExtractThinker\n<\/p><\/blockquote>\n<p>Google Document AI \u8fd8\u63d0\u4f9b\u4e86\u4e00\u4e2a <code>Splitter<\/code> \u5904\u7406\u5668\uff0c\u5b83\u53ef\u4ee5\u8bc6\u522b\u5b50\u6587\u6863\u8fb9\u754c\u5e76\u4e3a\u6bcf\u4e2a\u7247\u6bb5\u5206\u914d\u4e00\u4e2a\u7f6e\u4fe1\u5ea6\u5206\u6570\u3002\u5b83\u8f93\u51fa\u7ed3\u6784\u5316\u7684 JSON\uff08\u5217\u51fa\u9875\u9762\u8303\u56f4\u3001\u5206\u7c7b\u6807\u7b7e\u7b49\u7684\u5b9e\u4f53\uff09\u3002\u4f46\u662f\uff0c\u5b83\u6709\u660e\u663e\u7684\u9650\u5236\u2014\u2014\u4f8b\u5982\uff0c\u4e0d\u652f\u6301\u62c6\u5206\u5927\u578b\uff08\u8d85\u8fc7 30 \u9875\uff09\u903b\u8f91\u6587\u6863\uff0c\u5e76\u4e14\u62c6\u5206\u5668\u53ea\u4f1a\u5728\u9875\u9762\u8fb9\u754c\u5904\u62c6\u5206\u6587\u6863\uff0c\u800c\u4e0d\u4f1a\u771f\u6b63\u4e3a\u4f60\u62c6\u5206 PDF\u3002<\/p>\n<p>\u76f8\u6bd4\u4e4b\u4e0b\uff0cExtractThinker \u7684\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li>\u6ca1\u6709\u4e25\u683c\u7684\u9875\u9762\u9650\u5236\u2014\u2014\u5b83\u53ef\u4ee5\u901a\u8fc7\u5206\u5757\u6216\u589e\u91cf\u7b56\u7565\uff08\u4f8b\u5982\u60f0\u6027\uff09\u5206\u6790\u4efb\u610f\u957f\u7684\u6587\u4ef6\u3002<\/li>\n<li>\u96c6\u6210\u5206\u7c7b\u903b\u8f91\u2014\u2014\u62c6\u5206\u51b3\u7b56\u53ef\u4ee5\u7531 LLM \u6d1e\u5bdf\uff08\u4f8b\u5982 Gemini 2.0\uff09\u800c\u4e0d\u662f\u56fa\u5b9a\u7684\u9875\u9762\u7ea7\u542f\u53d1\u5f0f\u65b9\u6cd5\u9a71\u52a8\u3002<\/li>\n<li>\u6267\u884c\u6574\u4e2a\u7ba1\u9053\u2014\u2014\u5728\u5355\u4e2a\u5de5\u4f5c\u6d41\u4e2d\u8fdb\u884c\u63d0\u53d6\u3001\u5206\u7c7b\u548c\u62c6\u5206\uff0c\u4f7f\u7528\u56de\u9000\u903b\u8f91\u548c\u7528\u4e8e\u7ed3\u6784\u5316\u6570\u636e\u7684\u9ad8\u7ea7\u57fa\u4e8e Pydantic \u7684\u5408\u540c\u3002<\/li>\n<\/ul>\n<p>\u7b80\u800c\u8a00\u4e4b\uff0c\u867d\u7136 Document AI \u7684\u62c6\u5206\u5668\u9002\u7528\u4e8e\u8f83\u7b80\u5355\u7684\u60c5\u51b5\uff08\u7279\u522b\u662f\u5982\u679c\u4f60\u60f3\u8981\u4e00\u4e2a\u7528\u4e8e\u8f83\u77ed\u6587\u6863\u7684\u5f00\u7bb1\u5373\u7528\u7684\u5904\u7406\u5668\uff09\uff0c\u4f46 ExtractThinker \u7684\u62c6\u5206\u5668\u66f4\u52a0\u7075\u6d3b\uff0c\u53ef\u4ee5\u7edf\u4e00\u9ad8\u7ea7\u5206\u7c7b\u6216\u591a\u5c42 LLM \u903b\u8f91\u2014\u2014\u4e00\u79cd\u7528\u4e8e\u5927\u89c4\u6a21\u3001\u590d\u6742 IDP \u7ba1\u9053\u7684\u4e00\u4f53\u5316\u65b9\u6cd5\u3002<\/p>\n<h2>5\u3001\u4ef7\u683c\u6bd4\u8f83<\/h2>\n<p>\u867d\u7136\u8c37\u6b4c\u5c1a\u672a\u6b63\u5f0f\u516c\u5e03 Gemini 2.0 \u5b9a\u4ef7\uff0c\u4f46 Gemini 1.5 \u6a21\u578b\u53ef\u4ee5\u4f5c\u4e3a\u53c2\u8003\u3002\u5728\u5927\u591a\u6570\u516c\u5f00\u9884\u89c8\u4e2d\uff0c\u9002\u7528\u4ee5\u4e0b\u8d39\u7387\uff1a<\/p>\n<ul>\n<li>\u8f93\u5165\u4ee4\u724c\uff1a\u6bcf 100 \u4e07\u4e2a\u4ee4\u724c 0.075 \u7f8e\u5143<\/li>\n<li>\u8f93\u51fa\u4ee4\u724c\uff1a\u6bcf 100 \u4e07\u4e2a\u4ee4\u724c 0.30 \u7f8e\u5143<\/li>\n<\/ul>\n<p>\u4f8b\u5982\uff0c\u5982\u679c\u4f60\u53d1\u9001 800 \u4e2a\u6587\u672c\u4ee4\u724c\uff08\u63d0\u793a\uff09\u5e76\u6536\u5230 500 \u4e2a\u4ee4\u724c\uff08\u6a21\u578b\u54cd\u5e94\uff09\uff0c\u5219\u603b\u5171\u9700\u8981 1,300 \u4e2a\u4ee4\u724c\uff1a<\/p>\n<p>\u603b\u8d39\u7528 = \u6bcf\u9875 0.00021 \u7f8e\u5143\u3002<\/p>\n<blockquote><p>\n  \u6ce8\u610f\uff1a\u8f93\u51fa\u4ee4\u724c\u6bd4\u8f93\u5165\u4ee4\u724c\u8d35 4 \u500d\u3002Gemini \u7684\u5b9e\u9645\u4ef7\u683c2.0 \u53ef\u80fd\u56e0\u5730\u533a\u3001\u5c42\u7ea7\u6216 Google \u7684\u65b0\u516c\u544a\u800c\u5f02\u3002\n<\/p><\/blockquote>\n<h2>6\u3001\u9009\u62e9Document AI \u6216 OCR<\/h2>\n<p>\u6839\u636e\u4f60\u7684\u9700\u6c42\uff0c\u53ef\u4ee5\u5c06 Gemini \u4e0e\u4ee5\u4e0b\u4ea7\u54c1\u914d\u5bf9\uff1a<\/p>\n<p>Tesseract \u6216\u5176\u4ed6 OCR\uff1a<\/p>\n<ul>\n<li>\u8d39\u7528\uff1a\u901a\u5e38\u4e3a 0.00 \u7f8e\u5143\uff08\u5f00\u6e90\uff09\u3002<\/li>\n<li>\u5982\u679c\u4e0e\u7ea6 1,300 \u4e2a\u4ee4\u724c\u7684 LLM \u8bf7\u6c42\u76f8\u7ed3\u5408\uff0c\u4f60\u7684\u6bcf\u9875\u603b\u8d39\u7528\u4ecd\u7ea6\u4e3a 0.0002 \u7f8e\u5143 &#8211; \u5bf9\u4e8e\u5927\u6279\u91cf\u6765\u8bf4\u975e\u5e38\u4fbf\u5b9c<\/li>\n<\/ul>\n<p>Document AI OCR \u6216\u5e03\u5c40\uff1a<\/p>\n<ul>\n<li>\u603b\u8ba1\u53ef\u80fd\u4e3a\u6bcf\u9875 0.0017 \u7f8e\u5143 &#8211; 0.0102 \u7f8e\u5143\uff0c\u5177\u4f53\u53d6\u51b3\u4e8e\u6587\u6863 AI \u5904\u7406\u5668\u3002<\/li>\n<\/ul>\n<p>\u6587\u6863 AI \u4e13\u7528\u89e3\u6790\u5668\uff08\u4f8b\u5982\uff0c\u5e03\u5c40\uff09\uff1a<\/p>\n<ul>\n<li>\u6bcf 10 \u9875\u7ea6 0.10 \u7f8e\u5143\uff080.01 \u7f8e\u5143\/\u9875\uff09\uff0c\u5982\u679c\u5185\u7f6e\u5b57\u6bb5\u8db3\u591f\uff0c\u5219\u4e0d\u9700\u8981 LLM\u3002<\/li>\n<li>\u5982\u679c\u4f60\u4ecd\u9700\u8981 LLM \u6765\u5904\u7406\u989d\u5916\u5b57\u6bb5\u6216\u9a8c\u8bc1\uff0c\u8bf7\u6bcf\u9875\u6dfb\u52a0 0.0002 \u7f8e\u5143\uff08Gemini \u4ee4\u724c\uff09\u3002<\/li>\n<li>\u6bcf\u9875\u603b\u8ba1\u7ea6 0.0102 \u7f8e\u5143\u3002<\/li>\n<\/ul>\n<p>\u5373\u4f7f\u662f\u8f83\u4f4e\u7ea7\u522b\u7684 Document AI \u8ba1\u5212\uff0cGoogle \u7684 OCR \u4e5f\u662f\u6700\u5148\u8fdb\u7684\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u8d28\u91cf\u8f83\u5dee\u7684\u626b\u63cf\u6216\u624b\u5199\u3002\u5982\u679c\u4f60\u7684\u6587\u6863\u662f\u7b80\u5355\u7684\u56fe\u50cf\uff08\u4f8b\u5982\uff0c\u952e\u5165\u7684\u6587\u672c\uff0c\u5bf9\u6bd4\u5ea6\u597d\uff09\uff0c\u514d\u8d39\u7684 OCR \u52a0\u4e0a Gemini \u4ee4\u724c\u53ef\u4ee5\u8282\u7701\u5927\u91cf\u6210\u672c\u2014\u2014\u6bcf\u9875 0.0002 \u7f8e\u5143\u4e0e 0.01 \u7f8e\u5143\u5728\u89c4\u6a21\u4e0a\u76f8\u5dee\u7ea6 50 \u500d\u3002<\/p>\n<p>\u5bf9\u4e8e\u8bb8\u591a\u7528\u4f8b\uff0c\u5177\u6709\u5f00\u6e90 OCR \u6216\u76f4\u63a5 LLM \u89c6\u89c9\u7684 ExtractThinker \u662f\u6700\u5177\u6210\u672c\u6548\u76ca\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u540c\u65f6\u4ecd\u5141\u8bb8\u4f60\u5728\u9700\u8981\u65f6\u7ed3\u5408 Document AI \u7684\u9ad8\u7ea7 OCR \u6765\u5904\u7406\u7b7e\u540d\u6216\u5177\u6709\u6311\u6218\u6027\u7684\u626b\u63cf\u3002<\/p>\n<h2>7\u3001\u7ed3\u675f\u8bed<\/h2>\n<p>\u901a\u8fc7\u5c06 ExtractThinker \u4e0e Gemini 2.0 \u6a21\u578b\u76f8\u7ed3\u5408\uff0c\u4f60\u53ef\u4ee5\u6784\u5efa\u4e00\u4e2a\u5168\u9762\u7684 IDP \u5de5\u4f5c\u6d41\u7a0b\uff0c\u4f18\u96c5\u5730\u5e73\u8861\u901f\u5ea6\u3001\u6210\u672c\u548c\u51c6\u786e\u6027\u3002\u6587\u6863\u52a0\u8f7d\uff08\u901a\u8fc7 DocumentLoader\uff09\u3001\u5206\u7c7b\uff08\u5982\u679c\u7f6e\u4fe1\u5ea6\u4f4e\uff0c\u5219\u4f7f\u7528\u540e\u5907\u5c42\uff09\u3001\u63d0\u53d6\uff08\u4f7f\u7528\u57fa\u4e8e Pydantic \u7684\u5408\u540c\uff09\u548c\u62c6\u5206\u90fd\u96c6\u4e2d\u5728\u4e00\u4e2a\u7b80\u5316\u7684\u7cfb\u7edf\u4e2d\u3002<\/p>\n<p>\u65e0\u8bba\u4f60\u5904\u7406\u7684\u662f\u626b\u63cf\u7684\u53d1\u7968\u3001\u8bb8\u53ef\u8bc1\u3001\u591a\u9875 PDF \u8fd8\u662f\u5176\u4ed6\u6587\u6863\u7c7b\u578b\uff0cExtractThinker \u90fd\u53ef\u4ee5\u5728\u6587\u6863\u667a\u80fd\u4e16\u754c\u4e2d\u4e3a\u4f60\u63d0\u4f9b\u652f\u6301\uff01<\/p>\n<hr>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8 Google \u7684 Gemini 2.0 \u6a21\u578b\u5982\u4f55\u4e0e ExtractThinker\uff08\u4e00\u79cd\u65e8\u5728\u534f\u8c03 OCR\u3001\u5206\u7c7b\u3001\u6587\u6863\u62c6\u5206\u548c\u6570\u636e\u63d0\u53d6\u7ba1\u9053\u7684\u5f00\u6e90\u6846\u67b6\uff09\u7ed3\u5408\u4f7f\u7528\uff0c\u4ece\u800c\u589e\u5f3a\u667a\u80fd\u6587\u6863\u5904\u7406 (IDP)\u3002\u6211\u4eec\u5c06\u4ecb\u7ecd Google Document AI \u5982\u4f55\u878d\u5165\u5176\u4e2d\uff0c\u4ee5\u53ca Gemini 2.0 Flash \u7684\u65b0\u529f\u80fd\uff0c\u5e76\u901a\u8fc7\u4ee3\u7801\u793a\u4f8b\u548c\u5b9a\u4ef7\u89c1\u89e3\u603b\u7ed3\u6240\u6709\u5185\u5bb9\u3002 1\u3001\u7b80\u4ecb \u667a\u80fd\u6587\u6863\u5904\u7406 (IDP) \u662f\u5c06\u975e\u7ed3\u6784\u5316\u6570\u636e\uff08\u5982\u53d1\u7968\u3001\u9a7e\u9a76\u6267\u7167\u548c\u62a5\u544a\uff09\u8f6c\u6362\u4e3a\u7ed3\u6784\u5316\u3001\u53ef\u64cd\u4f5c\u4fe1\u606f\u7684\u5173\u952e\u5de5\u4f5c\u6d41\u7a0b\u3002\u867d\u7136\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u73b0\u5728\u53ef\u4ee5\u76f4\u63a5\u5904\u7406\u56fe\u50cf\u548c PDF\uff0c\u4f46\u4ec5\u4ec5\u5c06\u56fe\u50cf\u8f93\u5165 LLM \u5e76\u5e0c\u671b\u83b7\u5f97\u5b8c\u7f8e\u7684\u7ed3\u679c\u901a\u5e38\u662f\u4e0d\u591f\u7684\u3002\u76f8\u53cd\uff0c\u5f3a\u5927\u7684 IDP \u7ba1\u9053\u7ed3\u5408\u4e86\uff1a OCR \u6216\u5176\u4ed6\u5e03\u5c40\u63d0\u53d6\u5de5\u5177\uff08\u5982 Google Document AI\u3001Tesseract \u6216 PyPDF\uff09\u3002 \u5206\u7c7b\u4ee5\u8bc6\u522b\u6587\u6863\u7c7b\u578b\uff08\u53d1\u7968\u3001\u5408\u540c\u3001\u8bb8\u53ef\u8bc1\u7b49\uff09\u3002 \u62c6\u5206\u4ee5\u5904\u7406\u5927\u578b\u7ec4\u5408\u6587\u4ef6\u5e76\u5c06\u5176\u5206\u89e3\u4e3a\u903b\u8f91\u90e8\u5206\u3002 \u63d0\u53d6\u4ee5\u5c06\u4fe1\u606f\u6620\u5c04\u5230\u7ed3\u6784\u5316\u7684 Pydantic \u6a21\u578b\u4e2d \u2014 \u4f8b\u5982\u63d0\u53d6\u53d1\u7968\u53f7\u3001\u65e5\u671f\u3001\u603b\u91d1\u989d\u6216\u89e3\u91ca\u56fe\u8868\u6570\u636e\u3002 ExtractThinker \u662f\u4e00\u4e2a\u5f00\u7bb1\u5373\u7528\u7684\u5e93\uff0c\u53ef\u8ba9\u4f60\u5c06\u5b83\u4eec\u4e0e Google \u5168\u65b0\u7684 Gemini 2.0 \u6a21\u578b\u65e0\u7f1d\u96c6\u6210\u3002 2\u3001Google Document AI \u5728\u6df1\u5165\u7814\u7a76\u57fa\u4e8e LLM \u7684\u63d0\u53d6\u4e4b\u524d\uff0c\u8ba9\u6211\u4eec\u5148\u8c08\u8c08 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53734","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53734","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53734"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53734\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53734"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53734"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53734"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}