{"id":3668,"date":"2024-05-08T17:47:26","date_gmt":"2024-05-08T09:47:26","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=3668"},"modified":"2024-05-08T17:47:26","modified_gmt":"2024-05-08T09:47:26","slug":"rag-%e5%85%a5%e9%97%a8%e6%95%99%e7%a8%8bpdf-ollama","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=3668","title":{"rendered":"RAG \u5165\u95e8\u6559\u7a0b(PDF-Ollama)"},"content":{"rendered":"\n<p>\u672c\u6559\u7a0b\u4f7f\u7528\u4e86<strong><a href=\"https:\/\/github.com\/VikParuchuri\/surya\">surya-ocr<\/a><\/strong>\u5e93\uff0c\u5b9e\u73b0\u672c\u5730RAG\uff0c<\/p>\n\n\n\n<p>\u4f7f\u7528\u4e86\u5d4c\u5165\u6a21\u578b bert-base-multilingual-cased\uff08\u652f\u6301\u591a\u8bed\u8a00\uff09<\/p>\n\n\n\n<p>\u548c Ollama \u73af\u5883\u4e0b\u7684\u63a8\u7406\u6a21\u578b Qwen1.5-1.8B-Chat<\/p>\n\n\n\n<p>Surya \u662f\u4e00\u4e2a\u6587\u6863 OCR \u5de5\u5177\u5305\uff0c\u53ef\u4ee5\u5904\u7406pdf\u6587\u4ef6\u548c\u56fe\u7247\u7b49<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>90+ \u79cd\u8bed\u8a00\u7684 OCR\uff0c\u4e0e\u4e91\u670d\u52a1\u76f8\u6bd4\u5177\u6709\u4f18\u52bf<\/li>\n\n\n\n<li>\u4efb\u4f55\u8bed\u8a00\u7684\u884c\u7ea7\u6587\u672c\u68c0\u6d4b<\/li>\n\n\n\n<li>\u5e03\u5c40\u5206\u6790\uff08\u8868\u683c\u3001\u56fe\u50cf\u3001\u9875\u7709\u7b49\u68c0\u6d4b\uff09<\/li>\n\n\n\n<li>\u8bfb\u53d6\u987a\u5e8f\u68c0\u6d4b<\/li>\n<\/ul>\n\n\n\n<p>\u6d4b\u8bd5\u73af\u5883\uff1aWindows<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>1. \u5b89\u88c5\u5fc5\u8981\u7684\u5e93<\/strong><\/h2>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >!pip install surya-ocr\n!pip install python-magic\n\n!pip install -U sentence_transformers\n!pip install -U numpy<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>2. \u5f15\u5165\u6240\u6709\u7684\u5e93<\/strong><\/h2>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >from sentence_transformers import SentenceTransformer\nimport faiss\nimport numpy as np\nfrom typing import List, Dict, Any\nimport json\nimport requests\n\nimport io\nimport pypdfium2\n\nfrom surya.detection import batch_text_detection\nfrom surya.layout import batch_layout_detection\nfrom surya.model.detection.segformer import load_model, load_processor\nfrom surya.model.recognition.model import load_model as load_rec_model\nfrom surya.model.recognition.processor import load_processor as load_rec_processor\nfrom surya.model.ordering.processor import load_processor as load_order_processor\nfrom surya.model.ordering.model import load_model as load_order_model\nfrom surya.ordering import batch_ordering\nfrom surya.postprocessing.heatmap import draw_polys_on_image\nfrom surya.ocr import run_ocr\nfrom surya.postprocessing.text import draw_text_on_image\nfrom PIL import Image\nfrom surya.languages import CODE_TO_LANGUAGE\nfrom surya.input.langs import replace_lang_with_code\nfrom surya.schema import OCRResult, TextDetectionResult, LayoutResult, OrderResult\nfrom surya.settings import settings\n<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>3. \u5904\u7406 pdf \u7684\u51fd\u6570\u5b9a\u4e49<\/strong><\/h2>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >def open_pdf(pdf_file):\n    # \u6253\u5f00\u6587\u4ef6\u5e76\u8bfb\u53d6\u5185\u5bb9\u5230\u5185\u5b58\n    with open(pdf_file, 'rb') as file:\n        pdf_data = file.read()\n    stream = io.BytesIO(pdf_data)\n    return pypdfium2.PdfDocument(stream)\n\ndef page_count(pdf_file):\n    doc = open_pdf(pdf_file)\n    return len(doc)\n\ndef get_page_image(pdf_file, page_num, dpi=96):\n    doc = open_pdf(pdf_file)\n    renderer = doc.render(\n        pypdfium2.PdfBitmap.to_pil,\n        page_indices=[page_num - 1],\n        scale=dpi \/ 72,\n    )\n    png = list(renderer)[0]\n    png_image = png.convert(\"RGB\")\n    return png_image\n\ndef ocr(img, langs: List[str]) -&gt; (Image.Image, OCRResult):\n    replace_lang_with_code(langs)\n    img_pred = run_ocr([img], [langs], det_model, det_processor, rec_model, rec_processor)[0]\n\n    bboxes = [l.bbox for l in img_pred.text_lines]\n    text = [l.text for l in img_pred.text_lines]\n    rec_img = draw_text_on_image(bboxes, text, img.size, langs, has_math=\"_math\" in langs)\n    return rec_img, img_pred\n\ndef load_det_cached():\n    checkpoint = settings.DETECTOR_MODEL_CHECKPOINT\n    return load_model(checkpoint=checkpoint), load_processor(checkpoint=checkpoint)\n\ndef load_rec_cached():\n    return load_rec_model(), load_rec_processor()\n<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>4. ollama \u5bf9\u8bdd\u51fd\u6570<\/strong><\/h2>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >def chat(messages: List[Dict[str, Any]], model: str = '', stream: bool = False) -&gt; requests.Response:\n    json_data = {\n        'model': model,\n        'messages': messages,\n        'stream': stream,\n    }\n\n    response = requests.post(\n        'http:\/\/127.0.0.1:11434\/api\/chat',\n        json=json_data\n    )\n    #print(\"Request Headers:\", response.request.headers)\n    #print(\"Request Body:\", response.request.body)\n    #print(\"Response Status Code:\", response.status_code)\n    #print(\"Response Body:\", response.text)\n\n    response_object = json.loads(response.text)\n    return response_object['message']['content']\n<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>5. \u5d4c\u5165\u6a21\u578b<\/strong><\/h2>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" ># \u5c0f\u6a21\u578b\u7528\u4e8e\u521b\u5efa\u5d4c\u5165\n#embedder = SentenceTransformer('Qwen\/Qwen1.5-0.5B-Chat')\nembedder = SentenceTransformer('bert-base-multilingual-cased')<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>6. \u53c2\u6570\u521d\u59cb\u5316<\/strong><\/h2>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >languages=[\"English\"]\n\n# Initialize an empty list to store the embeddings\nembeddings_list = []\ndocuments = []\n\ndet_model, det_processor = load_det_cached()\nrec_model, rec_processor = load_rec_cached()<\/pre><\/div>\n\n\n\n<p>languages=[\u201cEnglish\u201d]\uff0c\u652f\u6301\u591a\u8bed\u8a00\uff0c\u53ef\u4ee5\u81ea\u884c\u52a0\u5165\u5176\u4ed6\u8bed\u8a00,\u6bd4\u5982\u52a0\u5165\u4e2d\u6587<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >languages=[\"English\", \"Chinese\"]<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>7. \u5904\u7406 pdf \u6587\u4ef6<\/strong><\/h2>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >in_file = \"data\/Learning to Model the World with Language.pdf\"\n\npage_count = page_count(in_file)\nprint(f\"page_count=\", page_count)\n\n# \u5faa\u73af\u904d\u5386\u6bcf\u4e00\u9875\nfor page_number in range(page_count):\n    pil_image = get_page_image(in_file, page_number + 1)\n    rec_img, pred = ocr(pil_image, languages)\n    document = \"\\n\".join([p.text for p in pred.text_lines])\n\n    embeddings = embedder.encode(document)   \n    embeddings_list.append(embeddings)\n    print(f\"page {page_number + 1},{len(document)}:\", document)\n    # print(f\"embeddings:{len(embeddings)},\", embeddings)\n    documents.append(document)<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>8. \u521b\u5efa FAISS \u7d22\u5f15\u548c\u63a8\u7406<\/strong><\/h2>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" ># \u521b\u5efaFAISS\u7d22\u5f15\nif embeddings_list:\n    embeddings_array = np.vstack(embeddings_list)\n    index = faiss.IndexFlatL2(embeddings_array.shape[1])\n    index.add(embeddings_array.astype('float32'))\n    \n    # \u7528\u6237\u95ee\u9898\u5904\u7406\u4e0e\u63a8\u7406\n    #question = \"What is the theme of the document? \"\n    question = \"\u8fd9\u4efd\u6587\u6863\u7684\u4e3b\u9898\u662f\u4ec0\u4e48\uff1f\"\n    query_embedding = embedder.encode([question])[0].astype('float32')\n    \n    # \u68c0\u7d22\u6700\u76f8\u5173\u7684\u51e0\u4e2a\u6587\u6863\u6bb5\u843d\n    combined_segments = \"\"\n    k = 3  # \u4f60\u5e0c\u671b\u68c0\u7d22\u7684\u76f8\u5173\u6587\u6863\u6570\u91cf\n    D, I = index.search(np.array([query_embedding]), k=k)\n    print(\"D:\", D)\n    print(\"I:\", I)\n    #print(\"Top\", k, \"most relevant document segments:\")\n    for idx, segment_index in enumerate(I[0]):\n        most_relevant_segment = documents[segment_index]\n        #print(f\"{idx+1}: {most_relevant_segment}\\n\")\n        combined_segments += \" \" + most_relevant_segment\n    \n    prompt = combined_segments + \"\\n\\n###\\n\\n\" + question + \"\\n\\n\u7528\u4e2d\u6587\u56de\u7b54\"\n    messages = [\n        {\"role\": \"system\", \"content\": \"You are a helpful assistant.\"},\n        {\"role\": \"user\", \"content\": prompt}\n    ]\n    response = chat(messages=messages, model=\"qwen:1.8b-chat\")\n    print(\"Answer to the question:\", response)\nelse:\n    print(\"No embeddings found. Please check your data.\")<\/pre><\/div>\n\n\n\n<p>\u663e\u793a\u90e8\u5206\u7ed3\u679c\u5185\u5bb9\uff1a<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >D: [[143.52664 143.77243 145.9426 ]]\nI: [[1 0 2]]\nAnswer to the question: \u8fd9\u7bc7\u6587\u6863\u7684\u4e3b\u9898\u662f\u591a\u6a21\u6001\u8bed\u8a00\u7406\u89e3\u7cfb\u7edf\u7684\u5b9e\u73b0\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u8be5\u7cfb\u7edf\u91c7\u7528\u4e86\u591a\u6a21\u6001\u7684\u8f93\u5165\u5904\u7406\u65b9\u5f0f\uff0c\u80fd\u591f\u4ece\u89c6\u89c9\u3001\u542c\u89c9\u7b49\u591a\u79cd\u9014\u5f84\u83b7\u53d6\u4eba\u7c7b\u8bed\u8a00\u4fe1\u606f\u3002\n\n\u5728\u591a\u6a21\u6001\u8bed\u8a00\u7406\u89e3\u7cfb\u7edf\u4e2d\uff0c\u6587\u672c\u4f5c\u4e3a\u8f93\u5165\u4fe1\u53f7\u88ab\u8f6c\u5316\u4e3a\u89c6\u89c9\u6216\u542c\u89c9\u7279\u5f81\uff0c\u5982\u56fe\u50cf\u3001\u97f3\u9891\u7b49\u3002\u8fd9\u4e9b\u7279\u5f81\u901a\u8fc7\u591a\u6a21\u6001\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u878d\u5408\u548c\u6620\u5c04\uff0c\u4ece\u800c\u5b9e\u73b0\u5bf9\u590d\u6742\u81ea\u7136\u8bed\u8a00\u73af\u5883\u7684\u7406\u89e3\u548c\u64cd\u4f5c\u3002\n\n\u4e3a\u4e86\u589e\u5f3a\u7cfb\u7edf\u7684\u6cdb\u5316\u80fd\u529b\u548c\u9002\u5e94\u6027\uff0c\u8be5\u7cfb\u7edf\u8fd8\u5f15\u5165\u4e86\u4e30\u5bcc\u7684\u77e5\u8bc6\u56fe\u8c31\u548c\u8bed\u4e49\u6a21\u578b\uff0c\u4ee5\u63d0\u5347\u7cfb\u7edf\u7684\u8de8\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u80fd\u529b\u3002\u6b64\u5916\uff0c\u4e3a\u4e86\u5145\u5206\u5229\u7528\u591a\u6a21\u6001\u8bed\u8a00\u7406\u89e3\u7cfb\u7edf\u5728\u89c6\u89c9\u3001\u542c\u89c9\u7b49\u591a\u79cd\u9014\u5f84\u83b7\u53d6\u4eba\u7c7b\u8bed\u8a00\u4fe1\u606f\u7684\u57fa\u7840\u4e0a\u5b9e\u73b0\u5bf9\u590d\u6742\u81ea\u7136\u8bed\u8a00\u73af\u5883\u7684\u7406\u89e3\u548c\u64cd\u4f5c\u7684\u4f18\u52bf\uff0c\u8be5\u7cfb\u7edf\u8fd8\u5728\u5176\u8bbe\u8ba1\u4e2d\u5145\u5206\u8003\u8651\u5230\u4e86\u591a\u6a21\u6001\u8bed\u8a00\u7406\u89e3\u7cfb\u7edf\u7684\u53ef\u6269\u5c55\u6027\u548c\u9002\u5e94\u6027\uff0c\u4ece\u800c\u4f7f\u5f97\u591a\u6a21\u6001\u8bed\u8a00\u7406\u89e3\u7cfb\u7edf\u7684\u6027\u80fd\u5f97\u4ee5\u5145\u5206\u53d1\u6325\u3002<\/pre><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6559\u7a0b\u4f7f\u7528\u4e86surya-ocr\u5e93\uff0c\u5b9e\u73b0\u672c\u5730RAG\uff0c \u4f7f\u7528\u4e86\u5d4c\u5165\u6a21\u578b bert-base-multilingua [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[444,443,442],"tags":[433,450,422,456],"class_list":["post-3668","post","type-post","status-publish","format-standard","hentry","category-ai","category-llm","category-llms","tag-ollama","tag-pdf","tag-rag","tag-surya-ocr"],"views":2240,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/3668","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3668"}],"version-history":[{"count":1,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/3668\/revisions"}],"predecessor-version":[{"id":3669,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/3668\/revisions\/3669"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3668"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3668"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3668"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}