{"id":4922,"date":"2024-10-20T22:19:39","date_gmt":"2024-10-20T14:19:39","guid":{"rendered":"https:\/\/nullthought.net\/?p=4922"},"modified":"2024-12-25T14:30:34","modified_gmt":"2024-12-25T06:30:34","slug":"visrag%ef%bc%9a%e6%8a%8arag%e6%89%a9%e5%b1%95%e5%88%b0%e5%9b%be%e7%89%87%e5%92%8c%e8%a7%86%e8%a7%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=4922","title":{"rendered":"VisRAG\uff1a\u628aRAG\u6269\u5c55\u5230\u56fe\u7247\u548c\u89c6\u89c9"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2410.10594\" target=\"_blank\" rel=\"noreferrer noopener\">VisRAG: Vision-based Retrieval-Augmented Generation on Multi-Modality Documents<\/a><\/strong>\u4e3b\u8981\u63a2\u8ba8\u4e86\u4e00\u4e2a\u540d\u4e3aVisRAG\u7684\u65b0\u578b\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u7ba1\u9053\uff0c\u65e8\u5728\u89e3\u51b3\u4f20\u7edfRAG\u4ec5\u57fa\u4e8e\u6587\u672c\u7684\u5c40\u9650\u6027\uff0c\u65e0\u6cd5\u6709\u6548\u5229\u7528\u6587\u6863\u4e2d\u7684\u89c6\u89c9\u4fe1\u606f\uff08\u5982\u5e03\u5c40\u548c\u56fe\u7247\uff09\u3002\u901a\u8fc7\u5f15\u5165<strong><a href=\"https:\/\/nullthought.net\/?p=4920\" target=\"_blank\" rel=\"noreferrer noopener\">\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff08VLM\uff09<\/a><\/strong>\uff0cVisRAG\u80fd\u591f\u76f4\u63a5\u4ee5\u89c6\u89c9\u5f62\u5f0f\u5d4c\u5165\u6574\u4e2a\u6587\u6863\uff0c\u4ece\u800c\u4fdd\u7559\u539f\u59cb\u6587\u6863\u4e2d\u7684\u6240\u6709\u4fe1\u606f\uff0c\u907f\u514d\u6587\u672c\u89e3\u6790\u8fc7\u7a0b\u4e2d\u5f15\u5165\u7684\u4fe1\u606f\u635f\u5931\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aShi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, Maosong Sun\uff0c\u6765\u81eaTsinghua University\uff08\u6e05\u534e\u5927\u5b66\uff09\uff0cModelBest Inc.\uff08\u9762\u58c1\u667a\u80fd\uff09\uff0cRice University\u548cNortheastern University\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"609\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-20-1024x609.png\" alt=\"\" class=\"wp-image-4923\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-20-1024x609.png 1024w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-20-300x179.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-20-768x457.png 768w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-20.png 1040w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2410.10594\" target=\"_blank\" rel=\"noreferrer noopener\">VisRAG: Vision-based Retrieval-Augmented Generation on Multi-Modality Documents<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u4ee5\u4e0b\u4e3a\u8bba\u6587\u5185\u5bb9\u4ecb\u7ecd\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u80cc\u666f\u4e0e\u7814\u7a76\u52a8\u673a<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">1.1 \u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u7684\u9650\u5236<\/h4>\n\n\n\n<p>\u4f20\u7edf\u7684RAG\u7cfb\u7edf\uff08\u5982TextRAG\uff09\u5e7f\u6cdb\u5e94\u7528\u4e8e\u6587\u672c\u9886\u57df\uff0c\u5176\u4e3b\u8981\u601d\u60f3\u662f\u901a\u8fc7\u68c0\u7d22\u5916\u90e8\u77e5\u8bc6\u5e93\u4e2d\u7684\u4fe1\u606f\u6765\u589e\u5f3a\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u751f\u6210\u80fd\u529b\u3002\u8fd9\u4e9b\u7cfb\u7edf\u5728\u5904\u7406\u7eaf\u6587\u672c\u77e5\u8bc6\u65f6\u8868\u73b0\u826f\u597d\uff0c\u5c24\u5176\u662f\u5728\u9488\u5bf9\u50cfWikipedia\u8fd9\u6837\u7ecf\u8fc7\u6574\u7406\u548c\u6e05\u6d17\u7684\u7eaf\u6587\u672c\u6570\u636e\u65f6\u3002\u7136\u800c\uff0c\u5728\u73b0\u5b9e\u573a\u666f\u4e2d\uff0c\u77e5\u8bc6\u5e38\u5e38\u4ee5\u591a\u6a21\u6001\u5f62\u5f0f\u5448\u73b0\uff0c\u4f8b\u5982\u4e66\u7c4d\u3001\u624b\u518c\u3001\u56fe\u8868\u7b49\uff0c\u8fd9\u4e9b\u6587\u6863\u4e0d\u4ec5\u5305\u542b\u6587\u672c\uff0c\u8fd8\u5305\u62ec\u5e03\u5c40\u3001\u56fe\u50cf\u3001\u8868\u683c\u548c\u56fe\u8868\u7b49\u5176\u4ed6\u89c6\u89c9\u5143\u7d20\u200b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1.2 \u591a\u6a21\u6001\u4fe1\u606f\u7684\u6311\u6218<\/h4>\n\n\n\n<p>\u5728\u591a\u6a21\u6001\u6587\u6863\u5904\u7406\u4e2d\uff0c\u4f20\u7edfRAG\u7cfb\u7edf\u9762\u4e34\u4fe1\u606f\u4e22\u5931\u7684\u95ee\u9898\u3002\u4f8b\u5982\uff0c\u6587\u6863\u89e3\u6790\u9700\u8981\u4f7f\u7528\u5e03\u5c40\u8bc6\u522b\u3001\u5149\u5b66\u5b57\u7b26\u8bc6\u522b\uff08OCR\uff09\u7b49\u6b65\u9aa4\uff0c\u8fd9\u4e9b\u6b65\u9aa4\u5bb9\u6613\u5f15\u5165\u9519\u8bef\u548c\u566a\u58f0\uff0c\u4ece\u800c\u5f71\u54cd\u68c0\u7d22\u4e0e\u751f\u6210\u7684\u8d28\u91cf\u3002\u6b64\u5916\uff0c\u8fd9\u4e9b\u89e3\u6790\u8fc7\u7a0b\u53ea\u80fd\u63d0\u53d6\u6587\u672c\u4fe1\u606f\uff0c\u800c\u65e0\u6cd5\u6709\u6548\u5229\u7528\u6587\u6863\u7684\u89c6\u89c9\u5143\u7d20\uff0c\u8fd9\u4e9b\u5143\u7d20\u5bf9\u7406\u89e3\u6587\u6863\u4e0a\u4e0b\u6587\u6709\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\u200b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1.3 VisRAG\u7684\u63d0\u51fa<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3aVisRAG\uff08Vision-based Retrieval-Augmented Generation\uff09\u7684\u65b0\u578b\u7ba1\u9053\u3002VisRAG\u5229\u7528\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff08VLM\uff09\u76f4\u63a5\u5bf9\u6587\u6863\u7684\u56fe\u50cf\u8fdb\u884c\u68c0\u7d22\u548c\u751f\u6210\uff0c\u65e0\u9700\u6587\u672c\u89e3\u6790\u8fc7\u7a0b\uff0c\u4ece\u800c\u6700\u5927\u5316\u4fdd\u7559\u6587\u6863\u4e2d\u7684\u539f\u59cb\u4fe1\u606f\u3002VisRAG\u901a\u8fc7\u76f4\u63a5\u5c06\u6587\u6863\u5d4c\u5165\u4e3a\u89c6\u89c9\u5f62\u5f0f\uff0c\u5e76\u5229\u7528\u8fd9\u4e9b\u5d4c\u5165\u8fdb\u884c\u68c0\u7d22\u548c\u751f\u6210\uff0c\u4f7f\u5176\u5728\u5904\u7406\u591a\u6a21\u6001\u6587\u6863\u65f6\uff0c\u8868\u73b0\u51fa\u66f4\u597d\u7684\u6027\u80fd\u548c\u9c81\u68d2\u6027\u200b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. VisRAG\u7684\u7cfb\u7edf\u6846\u67b6\u4e0e\u65b9\u6cd5\u7ec6\u8282<\/h3>\n\n\n\n<p>VisRAG\u7cfb\u7edf\u7531\u4e24\u5927\u6838\u5fc3\u7ec4\u4ef6\u6784\u6210\uff1a<strong>VisRAG-Ret<\/strong>\uff08\u7528\u4e8e\u68c0\u7d22\uff09\u548c<strong>VisRAG-Gen<\/strong>\uff08\u7528\u4e8e\u751f\u6210\uff09\uff0c\u4e24\u8005\u5747\u57fa\u4e8e\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2.1 VisRAG-Ret\uff1a\u57fa\u4e8e\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7684\u68c0\u7d22<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\">2.1.1 \u53cc\u7f16\u7801\u5668\u7ed3\u6784<\/h5>\n\n\n\n<p>VisRAG-Ret\u91c7\u7528\u4e86\u4e0e\u6587\u672c\u68c0\u7d22\u76f8\u4f3c\u7684\u53cc\u7f16\u7801\u5668\u7ed3\u6784\uff0c\u4f46\u4e0e\u4f20\u7edf\u7684\u6587\u672c\u7f16\u7801\u5668\u4e0d\u540c\uff0cVisRAG-Ret\u4f7f\u7528VLM\u5bf9\u67e5\u8be2\u548c\u6587\u6863\u76f4\u63a5\u8fdb\u884c\u89c6\u89c9\u7f16\u7801\u3002\u8be5\u6a21\u578b\u9996\u5148\u5c06\u67e5\u8be2\u7f16\u7801\u4e3a\u6587\u672c\u5d4c\u5165\uff0c\u5c06\u6587\u6863\u56fe\u50cf\u7f16\u7801\u4e3a\u89c6\u89c9\u5d4c\u5165\u3002\u7136\u540e\uff0c\u901a\u8fc7\u52a0\u6743\u5747\u503c\u6c60\u5316\uff08weighted mean pooling\uff09\u5bf9\u6700\u7ec8\u5c42\u7684VLM\u9690\u72b6\u6001\u8fdb\u884c\u5904\u7406\uff0c\u4ee5\u751f\u6210\u7528\u4e8e\u68c0\u7d22\u7684\u5d4c\u5165\u3002\u52a0\u6743\u7684\u65b9\u5f0f\u662f\u6839\u636e\u4f4d\u7f6e\u8d4b\u4e88\u540e\u7eedtoken\u66f4\u9ad8\u7684\u6743\u91cd\uff0c\u8fd9\u6837\u80fd\u591f\u66f4\u597d\u5730\u6355\u6349\u89c6\u89c9\u548c\u6587\u672c\u7684\u76f8\u5173\u6027\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2.1.2 \u4f18\u5316\u76ee\u6807<\/h5>\n\n\n\n<p>VisRAG-Ret\u901a\u8fc7\u4fe1\u606f\u5bf9\u6bd4\u635f\u5931\uff08InfoNCE loss\uff09\u8fdb\u884c\u8bad\u7ec3\uff0c\u7528\u4e8e\u6700\u5927\u5316\u67e5\u8be2\u4e0e\u6b63\u6837\u672c\u6587\u6863\u4e4b\u95f4\u7684\u76f8\u4f3c\u6027\uff0c\u6700\u5c0f\u5316\u4e0e\u8d1f\u6837\u672c\u4e4b\u95f4\u7684\u76f8\u4f3c\u6027\u3002\u5177\u4f53\u7684\u635f\u5931\u51fd\u6570\u5982\u4e0b\uff1a\u200b<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"707\" height=\"80\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-21.png\" alt=\"\" class=\"wp-image-4924\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-21.png 707w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-21-300x34.png 300w\" sizes=\"auto, (max-width: 707px) 100vw, 707px\" \/><\/figure>\n\n\n\n<p>\u5176\u4e2d\uff0cd<sup>+<\/sup>\u8868\u793a\u6b63\u6837\u672c\u6587\u6863\uff0cD<sup>\u2212<\/sup>\u4e3a\u8d1f\u6837\u672c\u6587\u6863\u96c6\uff0cs(q,d)\u4e3a\u67e5\u8be2\u548c\u6587\u6863\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u5f97\u5206\uff0c\u03c4\u4e3a\u6e29\u5ea6\u53c2\u6570\uff0c\u7528\u4e8e\u8c03\u6574\u5bf9\u6bd4\u5ea6\u200b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2.2 VisRAG-Gen\uff1a\u57fa\u4e8e\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7684\u751f\u6210<\/h4>\n\n\n\n<p>VisRAG-Gen\u7684\u76ee\u6807\u662f\u5728\u7ed9\u5b9a\u67e5\u8be2\u548c\u68c0\u7d22\u5230\u7684\u6587\u6863\u7684\u60c5\u51b5\u4e0b\uff0c\u751f\u6210\u6b63\u786e\u7684\u7b54\u6848\u3002VisRAG-Gen\u53ef\u4ee5\u5904\u7406\u5355\u5e45\u56fe\u50cf\u8f93\u5165\u548c\u591a\u5e45\u56fe\u50cf\u8f93\u5165\uff0c\u5177\u4f53\u7b56\u7565\u5305\u62ec\u4ee5\u4e0b\u51e0\u79cd\uff1a<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2.2.1 \u9875\u9762\u62fc\u63a5\uff08Page Concatenation\uff09<\/h5>\n\n\n\n<p>\u5bf9\u4e8e\u53ea\u80fd\u63a5\u53d7\u5355\u5e45\u56fe\u50cf\u8f93\u5165\u7684\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff08\u5982MiniCPM-V 2.0\uff09\uff0cVisRAG-Gen\u901a\u8fc7\u5c06\u591a\u9875\u6587\u6863\u8fdb\u884c\u6c34\u5e73\u62fc\u63a5\u7684\u65b9\u5f0f\uff0c\u5c06\u5b83\u4eec\u5408\u5e76\u4e3a\u4e00\u5e45\u56fe\u50cf\u8fdb\u884c\u8f93\u5165\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2.2.2 \u52a0\u6743\u9009\u62e9\uff08Weighted Selection\uff09<\/h5>\n\n\n\n<p>\u5bf9\u4e8e\u591a\u4e2a\u68c0\u7d22\u5230\u7684\u6587\u6863\uff0cVisRAG-Gen\u8fd8\u91c7\u7528\u52a0\u6743\u9009\u62e9\u7684\u7b56\u7565\uff0c\u5373\u5bf9\u6bcf\u4e00\u9875\u751f\u6210\u4e00\u4e2a\u7b54\u6848\uff0c\u7136\u540e\u57fa\u4e8e\u751f\u6210\u7684\u7f6e\u4fe1\u5ea6\u6765\u9009\u62e9\u6700\u7ec8\u7b54\u6848\u3002\u751f\u6210\u7f6e\u4fe1\u5ea6\u901a\u8fc7\u751f\u6210\u6982\u7387\u4e0e\u68c0\u7d22\u5f97\u5206\u7684\u52a0\u6743\u8ba1\u7b97\u5f97\u5230\uff0c\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"354\" height=\"54\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-22.png\" alt=\"\" class=\"wp-image-4925\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-22.png 354w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-22-300x46.png 300w\" sizes=\"auto, (max-width: 354px) 100vw, 354px\" \/><\/figure>\n\n\n\n<p>\u5176\u4e2d\uff0cP(a\u2223q,d)\u662f\u751f\u6210\u7b54\u6848\u7684\u6982\u7387\uff0c\u03bb(q,d)\u662f\u68c0\u7d22\u5f97\u5206\u7684\u5f52\u4e00\u5316\u6743\u91cd\u200b\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2.2.3 \u591a\u56fe\u50cf\u8f93\u5165\u7684VLM<\/h5>\n\n\n\n<p>\u5bf9\u4e8e\u80fd\u591f\u63a5\u53d7\u591a\u5e45\u56fe\u50cf\u8f93\u5165\u7684\u6a21\u578b\uff08\u5982MiniCPM-V 2.6\u548cGPT-4o\uff09\uff0cVisRAG-Gen\u76f4\u63a5\u5c06\u591a\u4e2a\u68c0\u7d22\u5230\u7684\u6587\u6863\u8f93\u5165\u6a21\u578b\uff0c\u8fd9\u6837\u53ef\u4ee5\u8fdb\u884c\u8de8\u9875\u9762\u7684\u63a8\u7406\uff0c\u7279\u522b\u662f\u5728\u9700\u8981\u591a\u8df3\uff08multi-hop\uff09\u63a8\u7406\u7684\u60c5\u5883\u4e2d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u6570\u636e\u6784\u5efa\u4e0e\u5b9e\u9a8c\u8bbe\u7f6e<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">3.1 \u6570\u636e\u96c6\u6784\u5efa<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u8bc4\u4f30VisRAG\u5728\u591a\u6a21\u6001\u6587\u6863\u4e2d\u7684\u8868\u73b0\uff0c\u8bba\u6587\u6784\u5efa\u4e86\u4e00\u4e2a\u5305\u542b\u591a\u79cd\u89c6\u89c9\u95ee\u7b54\uff08VQA\uff09\u6570\u636e\u96c6\u548c\u5408\u6210\u6570\u636e\u7684\u7efc\u5408\u6570\u636e\u96c6\u3002\u8fd9\u4e9b\u6570\u636e\u96c6\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>MP-DocVQA<\/strong>\uff1a\u7528\u4e8e\u5de5\u4e1a\u6587\u6863\u7684\u95ee\u7b54\u3002<\/li>\n\n\n\n<li><strong>ArXivQA\u3001ChartQA\u3001InfographicsVQA\u3001PlotQA<\/strong>\uff1a\u5206\u522b\u7528\u4e8e\u79d1\u5b66\u6587\u732e\u56fe\u8868\u3001\u56fe\u8868\u3001\u4fe1\u606f\u56fe\u548c\u79d1\u5b66\u56fe\u8868\u7684\u95ee\u7b54\u3002<\/li>\n\n\n\n<li><strong>SlideVQA<\/strong>\uff1a\u7528\u4e8e\u6f14\u793a\u6587\u7a3f\u5e7b\u706f\u7247\u7684\u95ee\u7b54\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">3.2 \u67e5\u8be2\u8fc7\u6ee4<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u4fdd\u8bc1\u68c0\u7d22\u7684\u6709\u6548\u6027\uff0cVisRAG\u5bf9\u4eceVQA\u6570\u636e\u96c6\u4e2d\u63d0\u53d6\u7684\u67e5\u8be2\u8fdb\u884c\u4e86\u8fc7\u6ee4\uff0c\u53bb\u9664\u4e86\u90a3\u4e9b\u4f9d\u8d56\u4e0a\u4e0b\u6587\u4e14\u7f3a\u4e4f\u660e\u786e\u6307\u5411\u6027\u7684\u67e5\u8be2\uff08\u5982\u201c\u4f1a\u8bae\u5728\u54ea\u91cc\u4e3e\u884c\uff1f\u201d\u8fd9\u7c7b\u95ee\u9898\uff09\u3002\u901a\u8fc7\u8fc7\u6ee4\uff0c\u4fdd\u7559\u4e86\u90a3\u4e9b\u5177\u6709\u8f83\u5f3a\u6587\u6863\u7279\u5f02\u6027\u7684\u67e5\u8be2\uff0c\u4ee5\u4fbf\u66f4\u6709\u6548\u5730\u8fdb\u884c\u5f00\u653e\u68c0\u7d22\u4efb\u52a1\u200b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3.3 \u5b9e\u9a8c\u8bbe\u7f6e<\/h4>\n\n\n\n<p>\u8bba\u6587\u8bbe\u8ba1\u4e86\u591a\u7ec4\u5b9e\u9a8c\u6765\u9a8c\u8bc1VisRAG\u5728\u68c0\u7d22\u548c\u751f\u6210\u4e2d\u7684\u6027\u80fd\u3002\u5b9e\u9a8c\u5206\u4e3a\u4e09\u7c7b\u8bbe\u7f6e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Off-the-shelf\u6a21\u578b<\/strong>\uff1a\u76f4\u63a5\u6d4b\u8bd5\u73b0\u6709\u7684\u6587\u672c\u548c\u89c6\u89c9\u68c0\u7d22\u6a21\u578b\u7684\u6027\u80fd\u3002<\/li>\n\n\n\n<li><strong>Out-of-domain\u8bbe\u7f6e<\/strong>\uff1a\u6a21\u578b\u4ec5\u5728\u5408\u6210\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u8bc4\u4f30\u5176\u5728VQA\u6570\u636e\u96c6\u4e0a\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n\n\n\n<li><strong>In-domain\u8bbe\u7f6e<\/strong>\uff1a\u6a21\u578b\u5728VQA\u6570\u636e\u96c6\u548c\u5408\u6210\u6570\u636e\u4e0a\u8fdb\u884c\u8054\u5408\u8bad\u7ec3\uff0c\u4ee5\u8bc4\u4f30\u5176\u5728\u57df\u5185\u76d1\u7763\u6761\u4ef6\u4e0b\u7684\u6027\u80fd\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">4. \u5b9e\u9a8c\u7ed3\u679c\u4e0e\u5206\u6790<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">4.1 \u68c0\u7d22\u6027\u80fd<\/h4>\n\n\n\n<p>VisRAG-Ret\u5728\u68c0\u7d22\u6027\u80fd\u4e0a\u663e\u8457\u4f18\u4e8e\u4f20\u7edf\u7684\u6587\u672c\u548c\u89c6\u89c9\u68c0\u7d22\u6a21\u578b\u3002\u5728Out-of-domain\u7684\u60c5\u51b5\u4e0b\uff0cVisRAG-Ret\u8d85\u8fc7\u4e86\u73b0\u6709\u7684\u6700\u5148\u8fdb\u6587\u672c\u68c0\u7d22\u6a21\u578bNV-Embed-v2\uff0c\u5c3d\u7ba1\u540e\u8005\u4f7f\u7528\u4e86\u66f4\u591a\u7684\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\u3002\u8fd9\u8868\u660eVisRAG-Ret\u5177\u6709\u5f88\u5f3a\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u7279\u522b\u662f\u5728\u5904\u7406\u672a\u89c1\u8fc7\u7684\u591a\u6a21\u6001\u6570\u636e\u65f6\uff0c\u8868\u73b0\u51fa\u8272\u200b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">4.2 \u751f\u6210\u6027\u80fd<\/h4>\n\n\n\n<p>VisRAG-Gen\u5728\u751f\u6210\u9636\u6bb5\u540c\u6837\u8868\u73b0\u4f18\u5f02\u3002\u901a\u8fc7\u76f4\u63a5\u5904\u7406\u6587\u6863\u56fe\u50cf\uff0cVisRAG-Gen\u76f8\u6bd4\u4f20\u7edf\u7684\u57fa\u4e8e\u6587\u672c\u7684\u751f\u6210\u5668\uff0c\u5982MiniCPM\uff08OCR\uff09\u548cGPT-4o\uff08OCR\uff09\uff0c\u8868\u73b0\u51fa\u66f4\u9ad8\u7684\u51c6\u786e\u7387\u3002\u5728\u5904\u7406\u5305\u542b\u591a\u5e45\u56fe\u50cf\u7684\u573a\u666f\u4e0b\uff0c\u591a\u56fe\u50cf\u8f93\u5165\u7684VLM\uff08\u5982GPT-4o\uff09\u8868\u73b0\u51fa\u4e86\u968f\u68c0\u7d22\u5230\u7684\u6587\u6863\u6570\u91cf\u589e\u52a0\u800c\u6027\u80fd\u63d0\u5347\u7684\u8d8b\u52bf\uff0c\u8fd9\u8868\u660e\u8fd9\u4e9b\u6a21\u578b\u5728\u591a\u9875\u9762\u63a8\u7406\u4e0a\u5177\u6709\u6f5c\u529b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">4.3 \u7aef\u5230\u7aef\u6027\u80fd<\/h4>\n\n\n\n<p>\u5728\u7aef\u5230\u7aef\u7684\u6027\u80fd\u6bd4\u8f83\u4e2d\uff0cVisRAG\u76f8\u6bd4TextRAG\u5b9e\u73b0\u4e86\u5927\u5e45\u63d0\u5347\u3002\u5728\u68c0\u7d22\u9636\u6bb5\uff0cVisRAG-Ret\u7684\u8868\u73b0\u660e\u663e\u4f18\u4e8eTextRAG\u7684\u68c0\u7d22\u5668\uff0c\u800c\u5728\u751f\u6210\u9636\u6bb5\uff0cVisRAG-Gen\u80fd\u591f\u66f4\u597d\u5730\u5229\u7528\u68c0\u7d22\u5230\u7684\u6587\u6863\u4fe1\u606f\uff0c\u663e\u8457\u63d0\u9ad8\u4e86\u6700\u7ec8\u7684\u95ee\u7b54\u51c6\u786e\u7387\u3002\u7279\u522b\u662f\u5728InfographicsVQA\u6570\u636e\u96c6\u4e0a\u7684\u6d4b\u8bd5\u4e2d\uff0cVisRAG\u7684\u6574\u4f53\u6027\u80fd\u8f83TextRAG\u63d0\u5347\u4e86\u63a5\u8fd1\u4e24\u500d\u200b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5. \u672a\u6765\u5de5\u4f5c\u4e0e\u6311\u6218<\/h3>\n\n\n\n<p>\u8bba\u6587\u8fd8\u8ba8\u8bba\u4e86VisRAG\u672a\u6765\u7684\u53d1\u5c55\u65b9\u5411\uff0c\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u63d0\u9ad8\u591a\u9875\u9762\u63a8\u7406\u80fd\u529b<\/strong>\uff1a\u5f53\u524d\u7684\u591a\u56fe\u50cf\u63a8\u7406\u80fd\u529b\u867d\u7136\u5df2\u7ecf\u6709\u663e\u8457\u63d0\u5347\uff0c\u4f46\u4ecd\u5b58\u5728\u8fdb\u4e00\u6b65\u4f18\u5316\u7684\u7a7a\u95f4\uff0c\u7279\u522b\u662f\u5728\u5904\u7406\u9700\u8981\u590d\u6742\u8de8\u9875\u9762\u903b\u8f91\u63a8\u7406\u7684\u95ee\u9898\u65f6\u3002<\/li>\n\n\n\n<li><strong>\u6269\u5c55\u6570\u636e\u96c6\u89c4\u6a21<\/strong>\uff1a\u4e3a\u4e86\u8fdb\u4e00\u6b65\u9a8c\u8bc1VisRAG\u7684\u6cdb\u5316\u80fd\u529b\u548c\u9c81\u68d2\u6027\uff0c\u672a\u6765\u9700\u8981\u5728\u66f4\u5927\u89c4\u6a21\u7684\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u5b9e\u9a8c\u200b\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>VisRAG on GitHub: <a href=\"https:\/\/github.com\/openbmb\/visrag\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/openbmb\/visrag<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587VisRAG: Vision-based Retrieval-Augmented Generation o [&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":"","ast-disable-related-posts":"","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":"","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":""}},"footnotes":""},"categories":[8],"tags":[39,63,76,90,86,80],"class_list":{"0":"post-4922","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"hentry","6":"category-tech","7":"tag-ai","9":"tag-rag","10":"tag-multimodal","11":"tag-llm","12":"tag-cv"},"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587VisRAG: Vision-based Retrieval-Augmented Generation o&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4922","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4922"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4922\/revisions"}],"predecessor-version":[{"id":4927,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4922\/revisions\/4927"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4922"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4922"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4922"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}