{"id":6361,"date":"2025-10-16T12:10:24","date_gmt":"2025-10-16T04:10:24","guid":{"rendered":"https:\/\/nullthought.net\/?p=6361"},"modified":"2025-10-16T12:10:25","modified_gmt":"2025-10-16T04:10:25","slug":"refrag-representation-for-rag-%ef%bc%9a%e8%a7%a3%e5%86%b3rag%e5%a4%84%e7%90%86%e5%a4%a7%e9%87%8f%e5%a4%96%e9%83%a8%e7%9f%a5%e8%af%86%e6%97%b6%e9%81%87%e5%88%b0%e7%9a%84%e9%80%9f%e5%ba%a6%e6%85%a2","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6361","title":{"rendered":"REFRAG (REpresentation For RAG) \uff1a\u89e3\u51b3RAG\u5904\u7406\u5927\u91cf\u5916\u90e8\u77e5\u8bc6\u65f6\u9047\u5230\u7684\u901f\u5ea6\u6162\u548c\u5185\u5b58\u5360\u7528\u5927\u95ee\u9898"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/www.arxiv.org\/abs\/2509.01092\" target=\"_blank\" rel=\"noreferrer noopener\">REFRAG: Rethinking RAG based Decoding<\/a><\/strong>\u4ecb\u7ecd\u4e86\u4e00\u79cd\u540d\u4e3a <strong>REFRAG (REpresentation For RAG)<\/strong> \u7684\u65b0\u578b\u9ad8\u6548\u89e3\u7801\u6846\u67b6\uff0c\u4e13\u4e3a<strong><a href=\"https:\/\/nullthought.net\/?tag=rag\" target=\"_blank\" rel=\"noreferrer noopener\">\u68c0\u7d22\u589e\u5f3a\u751f\u6210 (RAG)<\/a><\/strong> \u5e94\u7528\u8bbe\u8ba1\u3002\u7b80\u5355\u6765\u8bf4\uff0cREFRAG \u7684\u76ee\u6807\u662f\u89e3\u51b3\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLMs) \u5728\u4f7f\u7528 RAG \u5904\u7406\u5927\u91cf\u5916\u90e8\u77e5\u8bc6\u65f6\u9047\u5230\u7684\u901f\u5ea6\u6162\u548c\u5185\u5b58\u5360\u7528\u5927\u7684\u95ee\u9898\u3002\u5728 RAG \u5e94\u7528\u4e2d\uff0cLLMs \u4f1a\u5229\u7528\u68c0\u7d22\u5230\u7684\u5916\u90e8\u4fe1\u606f\u6765\u589e\u5f3a\u56de\u7b54\u3002\u7136\u800c\uff0c\u5f53\u8fd9\u4e9b\u68c0\u7d22\u5230\u7684\u4fe1\u606f\uff08\u5373\u4e0a\u4e0b\u6587\uff09\u5f88\u957f\u65f6\uff0c\u4f1a\u5bfc\u81f4\u7cfb\u7edf\u663e\u8457\u7684\u5ef6\u8fdf\u548c\u5de8\u5927\u7684\u5185\u5b58\u9700\u6c42\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5ef6\u8fdf\u95ee\u9898\uff1a<\/strong> \u63d0\u793a\uff08prompt\uff09\u8d8a\u957f\uff0c\u751f\u6210\u7b2c\u4e00\u4e2a token \u7684\u5ef6\u8fdf\uff08Time-to-First-Token, TTFT\uff09\u5c31\u4f1a<strong>\u4e8c\u6b21\u65b9\u7ea7<\/strong>\u589e\u52a0\uff0c\u800c\u751f\u6210\u540e\u7eed token \u7684\u5ef6\u8fdf\uff08Time-to-Iterative-Token, TTIT\uff09\u4f1a\u7ebf\u6027\u589e\u52a0\u3002<\/li>\n\n\n\n<li><strong>\u5185\u5b58\u95ee\u9898\uff1a<\/strong> \u957f\u63d0\u793a\u9700\u8981\u989d\u5916\u7684 Key-Value (KV) \u7f13\u5b58\u5185\u5b58\uff0c\u8fd9\u4e0e\u63d0\u793a\u957f\u5ea6\u5448\u7ebf\u6027\u5173\u7cfb\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8be5\u8bba\u6587\u6307\u51fa\uff0cRAG \u4e0a\u4e0b\u6587\u6709\u4e00\u4e2a\u72ec\u7279\u7684\u7ed3\u6784\uff0c\u4f7f\u5176\u6210\u4e3a\u4f18\u5316\u7684\u7406\u60f3\u76ee\u6807\uff1a\u5927\u90e8\u5206\u68c0\u7d22\u5230\u7684\u6bb5\u843d\u901a\u5e38\u8bed\u4e49\u76f8\u4f3c\u5ea6\u8f83\u4f4e\uff08\u53ef\u80fd\u662f\u7531\u4e8e\u591a\u6837\u6027\u6216\u53bb\u91cd\u64cd\u4f5c\uff09\uff0c\u5bfc\u81f4\u5728\u89e3\u7801\u8fc7\u7a0b\u4e2d\uff0cLLM \u5bf9\u8fd9\u4e9b\u4e0a\u4e0b\u6587\u7684\u6ce8\u610f\u529b\u6a21\u5f0f\u5448\u73b0\u51fa<strong>\u7a00\u758f\u7684\u3001\u5757\u5bf9\u89d2\u7ebf\u72b6<\/strong>\u7684\u7ed3\u6784\u3002\u8fd9\u610f\u5473\u7740\uff0c\u5728 RAG \u89e3\u7801\u8fc7\u7a0b\u4e2d\uff0c\u5bf9\u5927\u90e8\u5206\u4e0a\u4e0b\u6587\u8fdb\u884c\u7684\u8ba1\u7b97\u662f<strong>\u4e0d\u5fc5\u8981\u7684\u6d6a\u8d39<\/strong>\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aXiaoqiang Lin, Aritra Ghosh, Bryan Kian Hsiang Low, Anshumali Shrivastava, Vijai Mohan\uff0c\u6765\u81eaMeta Superintelligence Labs, National University of Singapore, Rice University\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1025\" height=\"505\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-8.png\" alt=\"\" class=\"wp-image-6364\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-8.png 1025w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-8-300x148.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-8-768x378.png 768w\" sizes=\"auto, (max-width: 1025px) 100vw, 1025px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/www.arxiv.org\/abs\/2509.01092\" target=\"_blank\" rel=\"noreferrer noopener\">REFRAG: Rethinking RAG based Decoding<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u4e00\u3001\u6838\u5fc3\u601d\u60f3\u4e0e\u7cfb\u7edf\u67b6\u6784\uff1a\u538b\u7f29\u2014\u611f\u77e5\u2014\u6269\u5c55\uff08compress\u2013sense\u2013expand\uff09<br>REFRAG\u7684\u5173\u952e\u505a\u6cd5\u662f\uff1a\u4e0d\u518d\u628a\u68c0\u7d22\u5230\u7684\u6bcf\u4e2a\u6bb5\u843d\u6309\u539f\u59cbtoken\u5e8f\u5217\u5582\u7ed9\u89e3\u7801\u5668\uff0c\u800c\u662f\u5148\u7528\u8f7b\u91cf\u7f16\u7801\u5668\u628a\u6bcf\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684chunk\u538b\u7f29\u6210\u4e00\u4e2a\u201cchunk embedding\u201d\uff0c\u7ecf\u6295\u5f71\u540e\u4e0e\u95ee\u9898\u7684token\u5d4c\u5165\u4e00\u8d77\u76f4\u63a5\u8f93\u5165\u5230\u539f\u6709\u7684\u89e3\u7801\u5668\u4e2d\u3002\u8fd9\u6837\u505a\u5e26\u6765\u4e09\u70b9\u6536\u76ca\uff1a\u2460\u628a\u201c\u89e3\u7801\u5668\u8f93\u5165\u957f\u5ea6\u201d\u7684\u5ea6\u91cf\u4ece\u201ctoken\u6570\u201d\u964d\u4e3a\u201cchunk\u6570\u201d\uff0c\u6ce8\u610f\u529b\u590d\u6742\u5ea6\u4ece\u6309token\u8ba1\u7f29\u51cf\u4e3a\u6309chunk\u8ba1\uff1b\u2461\u53ef\u91cd\u7528\u68c0\u7d22\u9636\u6bb5\u5df2\u7ecf\u53ef\u9884\u8ba1\u7b97\/\u7f13\u5b58\u7684chunk\u8868\u793a\uff0c\u907f\u514d\u91cd\u590d\u7b97\u529b\uff1b\u2462\u7531\u4e8e\u8f93\u5165\u5e8f\u5217\u663e\u8457\u7f29\u77ed\uff0cKV\u7f13\u5b58\u4e0eprefill\u8ba1\u7b97\u540c\u65f6\u964d\u4f4e\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0cREFRAG\u652f\u6301\u5728\u4efb\u610f\u4f4d\u7f6e\u8fdb\u884c\u201c\u6309\u9700\u5c55\u5f00\u201d\uff08\u4fdd\u7559\u539f\u59cbtoken\u800c\u975e\u538b\u7f29\uff09\uff0c\u5e76\u7528\u4e00\u4e2a\u8f7b\u91cfRL\u7b56\u7565\u5b66\u4e60\u201c\u54ea\u4e9bchunk\u9700\u8981\u4ee5\u539f\u6587token\u5f62\u5f0f\u4fdd\u7559\u3001\u54ea\u4e9b\u7528\u4f4e\u6210\u672c\u5d4c\u5165\u5373\u53ef\u201d\uff0c\u4ece\u800c\u5728\u4fdd\u8bc1\u81ea\u56de\u5f52\u56e0\u679c\u6027\u7684\u524d\u63d0\u4e0b\u5b9e\u73b0\u201c\u968f\u5904\u53ef\u538b\u7f29\u3001\u9009\u62e9\u6027\u6269\u5c55\u201d\u3002<\/p>\n\n\n\n<p>\u4e8c\u3001\u8bad\u7ec3\u8303\u5f0f\u4e0e\u914d\u65b9<br>1\uff09\u91cd\u5efa\u4efb\u52a1\uff08Reconstruction\uff09\u3002\u5148\u51bb\u7ed3\u89e3\u7801\u5668\uff0c\u4ec5\u8bad\u7ec3\u7f16\u7801\u5668\u4e0e\u6295\u5f71\u5c42\uff1a\u7ed9\u5b9a\u524ds\u4e2atoken\uff0c\u8981\u6c42\u89e3\u7801\u5668\u4ec5\u51ed\u8fd9\u4e9bchunk\u5d4c\u5165\u5c3d\u91cf\u201c\u91cd\u5efa\u201d\u539f\u6587\u3002\u8fd9\u4e00\u6b65\u628a\u201c\u538b\u7f29\u7684\u8bed\u4e49\u201d\u4e0e\u201c\u89e3\u7801\u5668\u7684token\u7a7a\u95f4\u201d\u5bf9\u9f50\uff0c\u8feb\u4f7f\u6a21\u578b\u7528\u201c\u5916\u90e8\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u201d\u800c\u975e\u53c2\u6570\u8bb0\u5fc6\u6765\u590d\u539f\u4fe1\u606f\u3002\u5b9e\u8bc1\u8868\u660e\uff0c\u6ca1\u6709\u8fd9\u4e00\u6b65\u521d\u59cb\u5316\uff0c\u540e\u7eed\u6301\u7eed\u9884\u8bad\u7ec3\uff08CPT\uff09\u5f88\u96be\u6536\u655b\u5230\u8db3\u591f\u4f4e\u7684\u56f0\u60d1\u5ea6\u3002<br>2\uff09\u8bfe\u7a0b\u5b66\u4e60\uff08Curriculum\uff09\u3002\u4ece\u201c\u4e00\u4e2achunk\u7684\u91cd\u5efa\u201d\u8d77\u6b65\uff0c\u9010\u6b65\u589e\u52a0\u8981\u91cd\u5efa\u7684chunk\u6570\u4e0e\u96be\u5ea6\uff0c\u5e76\u5728\u6570\u636e\u6df7\u5408\u4e0a\u968f\u8bad\u7ec3\u63a8\u8fdb\u4ece\u201c\u7b80\u5355\u6837\u672c\u201d\u8fc7\u6e21\u5230\u201c\u56f0\u96be\u6837\u672c\u201d\u3002\u6d88\u878d\u663e\u793a\u8bfe\u7a0b\u5b66\u4e60\u5bf9\u91cd\u5efa\u4efb\u52a1\u81f3\u5173\u91cd\u8981\u3002<br>3\uff09\u6301\u7eed\u9884\u8bad\u7ec3\uff08CPT\uff0cNext-Paragraph Prediction\uff09\u3002\u5728\u5b8c\u6210\u91cd\u5efa\u5bf9\u9f50\u540e\uff0c\u89e3\u51bb\u89e3\u7801\u5668\uff0c\u4ee4\u7f16\u7801\u5668\u8f93\u51fa\u8f85\u52a9\u89e3\u7801\u5668\u9884\u6d4b\u540e\u7eedo\u4e2atoken\uff0c\u4f7f\u201c\u7528\u538b\u7f29\u8868\u5f81\u8fdb\u884c\u751f\u6210\u201d\u4e0e\u201c\u7528\u539f\u6587token\u751f\u6210\u201d\u7684\u5206\u5e03\u5c3d\u53ef\u80fd\u4e00\u81f4\u3002<br>4\uff09\u9009\u62e9\u6027\u538b\u7f29\u7684RL\u7b56\u7565\u3002\u4ee5\u201c\u4e0b\u4e00\u6bb5\u9884\u6d4b\u7684\u56f0\u60d1\u5ea6\u201d\u4e3a\u8d1f\u5956\u52b1\uff0c\u7b56\u7565\u7f51\u7edc\u6839\u636echunk\u5d4c\u5165\u4e0e\u63a9\u7801\uff0c\u987a\u5e8f\u51b3\u7b56\u54ea\u4e9bchunk\u9700\u8981\u201c\u5c55\u5f00\u201d\u4e3a\u539f\u59cbtoken\u3001\u54ea\u4e9b\u4fdd\u6301\u538b\u7f29\uff0c\u4ece\u800c\u5728\u65e2\u5b9a\u538b\u7f29\u7387\u4e0b\u6700\u5c0f\u5316\u4fe1\u606f\u635f\u5931\u3002\u5b9e\u9a8c\u663e\u793a\u8be5\u7b56\u7565\u7a33\u5b9a\u4f18\u4e8e\u57fa\u4e8e\u56f0\u60d1\u5ea6\u7684\u542f\u53d1\u5f0f\u4e0e\u968f\u673a\u9009\u62e9\u3002<\/p>\n\n\n\n<p>\u4e09\u3001\u590d\u6742\u5ea6\u4e0e\u7cfb\u7edf\u6536\u76ca\uff08\u76f4\u89c9\uff09<br>\u5728RAG\u91cc\uff0c\u95ee\u9898token q\u226a\u4e0a\u4e0b\u6587token s\uff0c\u4e14\u4e0a\u4e0b\u6587\u88ab\u5206\u5757\u540e\uff0c\u6ce8\u610f\u529b\u7684\u8ba1\u7b97\u4e0e\u7f13\u5b58\u5747\u6309\u201c\u5757\u6570\u201d\u800c\u4e0d\u662f\u201ctoken\u603b\u91cf\u201d\u6269\u5c55\u3002\u82e5\u6bcfk\u4e2atoken\u538b\u62101\u4e2a\u5d4c\u5165\uff0cprefill\/TTFT\u7684\u5173\u952e\u8def\u5f84\u663e\u8457\u7f29\u77ed\uff1b\u5f53\u4e0a\u4e0b\u6587\u5f88\u957f\u65f6\uff0c\u6309\u6587\u4e2d\u5206\u6790\uff0c\u4e0a\u9650\u53ef\u8fbe\u201c\u63a5\u8fd1k\u6216\u66f4\u9ad8\u9636\u201d\u7684\u52a0\u901f\u3002\u8bba\u6587\u62a5\u544a\u5728\u4e2d\u957f\u4e0a\u4e0b\u6587\uff08\u598216k\u7ea7\u522b\uff09\u4e0b\uff0ck=16\u65f6TTFT\u52a0\u901f\u8d85\u8fc716\u00d7\uff0c\u5e76\u968f\u7f13\u5b58\u7b56\u7565\u53d8\u5316\u4e0d\u540c\u3002<\/p>\n\n\n\n<p>\u56db\u3001\u5b9e\u9a8c\u8bbe\u7f6e\u4e0e\u57fa\u7ebf<br>\u6301\u7eed\u9884\u8bad\u7ec3\u4ee5Slimpajama\u7684ArXiv\u4e0eBooks\u57df\u7b49\u957f\u6587\u672c\u4e3a\u4e3b\uff0c\u8bc4\u4f30\u9664\u81ea\u7559\u57df\u5916\u8fd8\u542bPG19\u4e0eProof-pile\uff1b\u4e0b\u6e38\u6db5\u76d6RAG\u95ee\u7b54\u3001\u957f\u6587\u6458\u8981\u3001\u591a\u8f6e\u5bf9\u8bdd\u7b49\uff0c\u57fa\u7ebf\u5305\u62ecLLaMA-NoContext\u3001Full-Context\u3001LLaMA-32K\u3001CEPE\u3001REPLUG\u7b49\uff0c\u7edf\u4e00\u5728LLaMA-2-7B\u4e0a\u8fdb\u884c\u53ef\u6bd4\u6027\u6d4b\u8bd5\u3002<\/p>\n\n\n\n<p>\u4e94\u3001\u5173\u952e\u7ed3\u679c\u4e0e\u91cf\u5316\u6536\u76ca<br>1\uff09\u9996token\u65f6\u5ef6\u4e0e\u541e\u5410\u3002REFRAG\u5728\u4e0d\u727a\u7272\u56f0\u60d1\u5ea6\u7684\u60c5\u51b5\u4e0b\uff0c\u628aTTFT\u52a0\u901f\u523030.85\u00d7\uff08\u76f8\u5bf9LLaMA\uff09\uff0c\u8f83\u6b64\u524dSOTA\uff08\u5982CEPE\uff09\u518d\u63d0\u5347\u7ea63.75\u00d7\uff1b\u538b\u7f29\u7387k=32\u65f6\u4f9d\u7136\u4fdd\u6301\u4e0eCEPE\u76f8\u5f53\u7684\u56f0\u60d1\u5ea6\u3002<br>2\uff09\u51c6\u786e\u6027\u4e0e\u56f0\u60d1\u5ea6\u3002\u88681\u663e\u793a\u5728s=2048\u3001o\u2208{512,1024,2048}\u8bbe\u7f6e\u4e0b\uff0cREFRAG-8\/16\u5728\u7edd\u5927\u591a\u6570\u4efb\u52a1\u4e0a\u4f18\u4e8eCEPE\u4e0eLLaMA\u622a\u65ad\u7b49\u57fa\u7ebf\uff1b\u540c\u7b49\u201c\u89e3\u7801\u5668token\u9884\u7b97\u201d\u4e0b\uff0cREFRAG-8\u660e\u663e\u80dc\u8fc7\u4ec5\u4fdd\u7559\u672b256 token\u7684LLaMA-256\uff0c\u8bf4\u660e\u201c\u538b\u7f29\u5d4c\u5165\u201d\u7684\u4fe1\u606f\u5bc6\u5ea6\u66f4\u9ad8\u3002<br>3\uff09\u8d85\u51fa\u539f\u751f\u4e0a\u4e0b\u6587\u7684\u201c\u6709\u6548\u6269\u7a97\u201d\u3002\u88682\u5728s\u2208{4096,8192,16384}, o=2048\u4e0b\u8bc4\u4f30\uff1aREFRAG-8\/16\u5728\u8bad\u7ec3\u4ec5\u52306k\u4e0a\u4e0b\u6587\u7684\u524d\u63d0\u4e0b\uff0c\u4f9d\u9760chunk\u5d4c\u5165\u5b9e\u73b0\u201c\u5916\u63a8\u201d\uff0c\u5728\u66f4\u957f\u4e0a\u4e0b\u6587\u4ecd\u4fdd\u6301\u4f18\u4e8eCEPE\u7684\u56f0\u60d1\u5ea6\u3002<br>4\uff09\u9009\u62e9\u6027\u538b\u7f29\u4f18\u4e8e\u56fa\u5b9a\u538b\u7f29\u3002\u7528\u76f8\u540c\u201c\u6709\u6548\u538b\u7f29\u7387\u201d\u6bd4\u8f83\u65f6\uff0c\u57fa\u4e8eRL\u7684\u201c\u6309\u9700\u5c55\u5f00\u201d\u7248\u672c\uff08\u5982REFRAG-16+RL\uff0c\u5728\u76ee\u6807\u538b\u7f29\u7387\u22488\u65f6\uff09\u666e\u904d\u4f18\u4e8e\u76f4\u63a5\u8bad\u7ec3\u7684REFRAG-8\uff0c\u4e14\u4f18\u4e8e\u56f0\u60d1\u5ea6\u542f\u53d1\u5f0f\u6216\u968f\u673a\u7b56\u7565\u3002<br>5\uff09\u7aef\u5230\u7aefRAG\u6027\u80fd\/\u5ef6\u8fdf\u7b49\u4ef7\u5bf9\u6bd4\u3002\u5728\u5f3a\u68c0\u7d22\u5668\u573a\u666f\u4e0b\uff0c\u7ed9\u5b9a\u76f8\u540c\u68c0\u7d22\u6bb5\u6570\uff0cREFRAG\u4e0eLLaMA\u51c6\u786e\u7387\u76f8\u5f53\uff1b\u4f46\u5728\u201c\u540c\u7b49\u5ef6\u8fdf\u9884\u7b97\u201d\uff08REFRAG\u53ef\u7eb3\u5165\u66f4\u591a\u6bb5\u843d\uff09\u65f6\uff0cREFRAG\u5e73\u5747\u63d0\u5347\u7ea61\u20132\u4e2a\u767e\u5206\u70b9\uff1b\u5f31\u68c0\u7d22\u5668\u573a\u666f\u4e0b\u4f18\u52bf\u66f4\u660e\u663e\uff0c\u56e0\u4e3a\u80fd\u5728\u76f8\u540c\u5ef6\u8fdf\u5185\u7eb3\u5165\u66f4\u591a\u4e0a\u4e0b\u6587\u4ee5\u5bf9\u51b2\u68c0\u7d22\u566a\u58f0\u3002<br>6\uff09\u591a\u8f6eRAG\u4e0e\u957f\u6587\u6458\u8981\u3002REFRAG\u5728\u591a\u8f6e\u5bf9\u8bdd\u6570\u636e\u96c6\uff08TopiOCQA\/ORConvQA\/QReCC\uff09\u4e0a\uff0c\u968f\u7740\u8f6e\u6b21\u4e0e\u6bb5\u6570\u589e\u52a0\uff0c\u76f8\u6bd4\u57fa\u4e8e4k\u4e0a\u4e0b\u6587\u7684LLaMA\u5fae\u8c03\u7248\u66f4\u4e0d\u6613\u201c\u622a\u65ad\u5386\u53f2\u201d\uff0c\u8868\u73b0\u66f4\u7a33\uff1b\u5728ArXiv\/PubMed\u957f\u6587\u6458\u8981\u4e0a\uff0c\u5728\u201c\u76f8\u540c\u89e3\u7801\u5668token\u9884\u7b97\uff08\u540c\u7b49\u65f6\u5ef6\uff09\u201d\u4e0b\u8fbe\u5230\u6700\u597d\u5206\u6570\u3002<\/p>\n\n\n\n<p>\u516d\u3001\u4e0e\u76f8\u5173\u5de5\u4f5c\u7684\u5173\u7cfb<br>\u4e0eCEPE\u76f8\u6bd4\uff1aCEPE\u7528\u4ea4\u53c9\u6ce8\u610f\u529b\u63a5\u5165\u4e0a\u4e0b\u6587token\u5d4c\u5165\uff0c\u964d\u4f4eKV\u5f00\u9500\uff0c\u4f46\u7834\u574f\u4e86\u4e25\u683c\u7684\u56e0\u679c\u7ed3\u6784\uff0c\u4e3b\u8981\u9002\u914d\u201c\u524d\u7f00\u4e0a\u4e0b\u6587\u201d\uff0c\u96be\u4ee5\u652f\u6301\u201c\u4efb\u610f\u4f4d\u7f6e\u538b\u7f29\/\u5c55\u5f00\u201d\u7684\u591a\u8f6eRAG\u4e0e\u957f\u6587\u6458\u8981\uff1b\u4e14CEPE\u4e0d\u505atoken\u2192chunk\u7ea7\u538b\u7f29\uff0cprefill\u65f6\u5ef6\u964d\u5e45\u6709\u9650\u3002REFRAG\u5728\u201c\u4e0d\u6539\u89e3\u7801\u5668\u67b6\u6784\u3001\u4e0d\u5f15\u5165\u65b0\u89e3\u7801\u5668\u53c2\u6570\u201d\u7684\u524d\u63d0\u4e0b\uff0c\u4ee5chunk\u7ea7\u538b\u7f29+\u6309\u9700\u5c55\u5f00\u53d6\u5f97\u66f4\u5927TTFT\u6536\u76ca\u4e0e\u66f4\u5e7f\u9002\u7528\u6027\u3002<br>\u4e0ePrompt\u538b\u7f29\uff08LLMLingua\u7b49\uff09\u4e92\u8865\uff1a\u63d0\u793a\u538b\u7f29\u591a\u57fa\u4e8e\u201c\u5220\/\u6539token\u201d\uff0c\u800cREFRAG\u4ee5\u201c\u66ff\u4ee3\u8868\u793a\u201d\u8fdb\u5165\u89e3\u7801\u5668\uff0c\u4e8c\u8005\u53ef\u53e0\u52a0\u8fdb\u4e00\u6b65\u964d\u5ef6\u3002<\/p>\n\n\n\n<p>\u4e03\u3001\u4e3a\u4f55\u201c\u538b\u7f29\u5d4c\u5165\u201d\u4e0d\u964d\u51c6\uff1f<br>\u76f4\u89c9\u4e0a\uff0c\u628ak\u4e2atoken\u538b\u62101\u4e2a\u5411\u91cf\u4f1a\u635f\u5931\u7ec6\u8282\uff0c\u4f46RAG\u4e0a\u4e0b\u6587\u7684\u201c\u5757\u5bf9\u89d2\u7a00\u758f\u201d\u610f\u5473\u7740\u8bb8\u591a\u8de8\u5757\u4ea4\u4e92\u5e76\u4e0d\u91cd\u8981\uff1b\u518d\u52a0\u4e0a\u68c0\u7d22\u4e0e\u91cd\u6392\u672c\u8eab\u5df2\u63d0\u4f9b\u201c\u4e0e\u95ee\u9898\u5bf9\u9f50\u201d\u7684\u8868\u5f81\uff0c\u538b\u7f29\u540e\u4ecd\u4fdd\u7559\u4e86\u8db3\u591f\u591a\u7684\u53ef\u7528\u4fe1\u53f7\u3002\u8bfe\u7a0b\u5b66\u4e60\u4e0e\u91cd\u5efa\u4efb\u52a1\u628a\u201c\u538b\u7f29\u8bed\u4e49\u201d\u5f3a\u529b\u5bf9\u9f50\u5230\u89e3\u7801token\u7a7a\u95f4\uff0cRL\u518d\u5728\u63a8\u7406\u65f6\u8865\u4e0a\u201c\u5fc5\u987b\u4fdd\u7559\u539f\u6587\u201d\u7684\u5c40\u90e8\uff0c\u4ece\u800c\u5728\u4f4e\u6210\u672c\u4e0e\u9ad8\u8d28\u91cf\u95f4\u53d6\u5f97\u5e73\u8861\u3002<\/p>\n\n\n\n<p>\u516b\u3001\u5c40\u9650\u6027\u4e0e\u8fb9\u754c<br>1\uff09\u538b\u7f29\u7387\u4e0a\u9650\uff1a\u8bad\u7ec3\u66f2\u7ebf\u663e\u793a\u538b\u5230\u00d764\u5f00\u59cb\u660e\u663e\u9000\u5316\uff0c\u8bf4\u660e\u201c\u53ef\u538b\u7f29\u5bb9\u91cf\u201d\u5b58\u5728\u5b9e\u9645\u8fb9\u754c\uff1b\u00d732\u4ecd\u53ef\u7528\uff0c\u4f46\u7a33\u59a5\u533a\u95f4\u662f\u00d78\uff5e\u00d716\u3002<br>2\uff09\u7f16\u7801\u5668\u89c4\u6a21\u6536\u76ca\u6709\u9650\uff1a\u5728\u6570\u636e\u91cf\u4e0d\u5927\u7684\u6301\u7eed\u9884\u8bad\u7ec3\u8bbe\u5b9a\u4e0b\uff0c\u628aRoBERTa\u4eceBase\u6362\u5230Large\u6536\u76ca\u5e76\u4e0d\u663e\u8457\uff0c\u53cd\u800c\u5e94\u4f18\u5148\u6269\u5927\u89e3\u7801\u5668\u89c4\u6a21\uff087B\u219213B\uff09\u66f4\u5212\u7b97\u3002<br>3\uff09\u7f13\u5b58\u4f9d\u8d56\u4e0e\u5de5\u7a0b\u6298\u4e2d\uff1a\u5f53\u4e0d\u80fd\u590d\u7528\u7f13\u5b58\u7684chunk\u5d4c\u5165\u65f6\uff0c\u6574\u4f53\u63d0\u901f\u4ecd\u663e\u8457\u4f46\u5f31\u4e8e\u201c\u5e26\u7f13\u5b58\u201d\u7684\u4e0a\u9650\uff0c\u90e8\u7f72\u4e2d\u9700\u7ed3\u5408\u5b9e\u9645\u7f13\u5b58\u547d\u4e2d\u7387\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p>\u4e5d\u3001\u843d\u5730\u5efa\u8bae\uff08\u9762\u5411\u5de5\u7a0b\u4e0e\u4ea7\u54c1\uff09<br>1\uff09\u5728RAG\u670d\u52a1\u4fa7\uff0c\u628a\u68c0\u7d22\u9636\u6bb5\u7684chunk\u5d4c\u5165\u56fa\u5316\u5230\u5bf9\u8c61\u5e93\uff0c\u5e76\u6309\u95ee\u9898\u57df\u4e0e\u65f6\u95f4\u5206\u7247\u7f13\u5b58\uff0c\u547d\u4e2d\u5219\u96f6\u5f00\u9500\u590d\u7528\uff1b\u547d\u4e2d\u7387\u2014\u5ef6\u8fdf\u2014\u5185\u5b58\u4e09\u8005\u505aA\/B\u6743\u8861\u3002<br>2\uff09\u5bf9\u9ad8\u4ef7\u503c\/\u9ad8\u4e0d\u786e\u5b9a\u6027\u7684chunk\u542f\u7528\u201cRL\u6309\u9700\u5c55\u5f00\u201d\uff0c\u5176\u4f59\u9ed8\u8ba4\u538b\u7f29\uff1b\u65e9\u671f\u53ef\u7528\u56f0\u60d1\u5ea6\u542f\u53d1\u5f0f\u8fd1\u4f3c\uff0c\u540e\u7eed\u518d\u4ee5\u7b56\u7565\u5b66\u4e60\u66ff\u6362\u3002<br>3\uff09\u5728\u5f31\u68c0\u7d22\u6216\u77e5\u8bc6\u566a\u58f0\u8f83\u5927\u7684\u4e1a\u52a1\u4e0a\uff0c\u628a\u201c\u7701\u4e0b\u7684\u65f6\u5ef6\u9884\u7b97\u201d\u8f6c\u5316\u4e3a\u201c\u66f4\u591a\u68c0\u7d22\u6bb5\u6570\u201d\uff0c\u901a\u5e38\u6bd4\u7cbe\u70bc\u5c11\u91cf\u6bb5\u843d\u66f4\u7a33\u5065\u3002<br>4\uff09\u591a\u8f6e\/\u4ee3\u7406\u578b\u5e94\u7528\u4f18\u5148\u66ff\u6362\u4e3aREFRAG\u5f0f\u8f93\u5165\u8def\u5f84\uff0c\u907f\u514d\u56e04k\u4e0a\u4e0b\u6587\u4e0a\u9650\u5bfc\u81f4\u7684\u5386\u53f2\u622a\u65ad\u3002<br>5\uff09\u538b\u7f29\u7387\u5efa\u8bae\uff1a\u9ed8\u8ba4\u00d78\u8d77\u6b65\uff1b\u82e5\u5ef6\u8fdf\u7ea2\u7ebf\u66f4\u7d27\u5f20\u6216\u68c0\u7d22\u6bb5\u6570\u591a\uff0c\u53ef\u8bd5\u00d716\u5e76\u5f00\u542f\u201c\u5c11\u91cf\u5173\u952echunk\u5c55\u5f00\u201d\u3002\u00d732\u9700\u7ed3\u5408\u4efb\u52a1\u9a8c\u8bc1\uff0c\u907f\u514d\u8d28\u91cf\u56de\u9000\u3002<\/p>\n\n\n\n<p>\u5341\u3001\u7ed3\u8bba<br>REFRAG\u4ee5\u201c\u7528chunk\u5d4c\u5165\u66ff\u4ee3\u5927\u90e8\u5206\u4e0a\u4e0b\u6587token + \u6309\u9700\u5c55\u5f00\u201d\u7684\u8def\u5f84\uff0c\u7cbe\u51c6\u547d\u4e2d\u4e86RAG\u7684\u7ed3\u6784\u6027\u7a00\u758f\uff0c\u663e\u8457\u964d\u4f4eprefill\u4e0eKV\u6210\u672c\u3002\u5728\u591a\u79cd\u957f\u4e0a\u4e0b\u6587\u573a\u666f\uff08RAG\u3001\u591a\u8f6e\u5bf9\u8bdd\u3001\u957f\u6587\u6458\u8981\uff09\u4e0a\uff0c\u4e8e\u4e0d\u964d\u6216\u63d0\u5347\u51c6\u786e\u6027\u7684\u524d\u63d0\u4e0b\uff0c\u628aTTFT\u52a0\u901f\u81f330.85\u00d7\uff0c\u5e76\u628a\u201c\u6709\u6548\u4e0a\u4e0b\u6587\u201d\u5916\u63a8\u5230\u539f\u751f\u7a97\u53e3\u768416\u00d7\u91cf\u7ea7\uff0c\u4e14\u4fdd\u6301\u4e0e\u4e3b\u6d41LLaMA\u7cfb\u7684\u67b6\u6784\u517c\u5bb9\uff0c\u5177\u5907\u5f88\u5f3a\u7684\u5de5\u7a0b\u53ef\u843d\u5730\u6027\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587REFRAG: Rethinking RAG based Decoding\u4ecb\u7ecd\u4e86\u4e00\u79cd\u540d\u4e3a REFRAG ( [&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":[39,77,76,86],"class_list":["post-6361","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-meta","tag-rag","tag-llm"],"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\u6587REFRAG: Rethinking RAG based Decoding\u4ecb\u7ecd\u4e86\u4e00\u79cd\u540d\u4e3a REFRAG (&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6361","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=6361"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6361\/revisions"}],"predecessor-version":[{"id":6365,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6361\/revisions\/6365"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6361"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6361"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6361"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}