{"id":34941,"date":"2024-12-15T13:59:05","date_gmt":"2024-12-15T05:59:05","guid":{"rendered":"https:\/\/17aitech.com\/?p=34941"},"modified":"2024-12-15T13:59:05","modified_gmt":"2024-12-15T05:59:05","slug":"%e5%bb%b6%e8%bf%9f%e4%ba%a4%e4%ba%92%e6%a8%a1%e5%9e%8b%ef%bc%8c%e4%b8%ba%e4%bb%80%e4%b9%88%e6%98%af%e4%b8%8b%e4%b8%80%e4%bb%a3rag%e7%9a%84%e6%a0%87%e9%85%8d%ef%bc%9f","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34941","title":{"rendered":"\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f"},"content":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f<\/a><\/p>\n

<\/a><\/section>\n
\n
\n
AIxiv\u4e13\u680f\u662f\u673a\u5668\u4e4b\u5fc3<\/mark>\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c\u673a\u5668\u4e4b\u5fc3<\/mark>AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/section>\n<\/section>\n<\/blockquote>\n

\u5f20\u9896\u5cf0\uff1a\u82f1\u98de\u6d41\u8054\u5408\u521b\u59cb\u4eba\uff0c\u591a\u5e74\u641c\u7d22\u3001AI\u3001Infra\u57fa\u7840\u8bbe\u65bd\u5f00\u53d1\u7ecf\u5386\uff0c\u76ee\u524d\u6b63\u81f4\u529b\u4e8e\u4e0b\u4e00\u4ee3 RAG \u6838\u5fc3\u4ea7\u54c1\u5efa\u8bbe\u3002<\/strong><\/p>\n

\u5728 RAG \u7cfb\u7edf\u5f00\u53d1\u4e2d\uff0c\u826f\u597d\u7684 Reranker \u6a21\u578b\u5904\u4e8e\u5fc5\u4e0d\u53ef\u5c11\u7684\u73af\u8282\uff0c\u4e5f\u603b\u662f\u88ab\u62ff\u6765\u653e\u5230\u5404\u7c7b\u8bc4\u6d4b\u5f53\u4e2d\uff0c\u8fd9\u662f\u56e0\u4e3a\u4ee5\u5411\u91cf\u641c\u7d22\u4e3a\u4ee3\u8868\u7684\u67e5\u8be2<\/mark>\uff0c\u4f1a\u9762\u4e34\u547d\u4e2d\u7387\u4f4e\u7684\u95ee\u9898\uff0c\u56e0\u6b64\u9700\u8981\u9ad8\u7ea7\u7684 Reranker \u6a21\u578b\u6765\u8865\u6551\uff0c\u8fd9\u6837\u5c31\u6784\u6210\u4e86\u4ee5\u5411\u91cf\u641c\u7d22\u4e3a\u7c97\u7b5b\uff0c\u4ee5 Reranker \u6a21\u578b\u4f5c\u7cbe\u6392\u7684\u4e24\u9636\u6bb5\u6392\u5e8f\u67b6\u6784\u3002<\/p>\n

\u76ee\u524d\u6392\u5e8f\u6a21\u578b\u7684\u67b6\u6784\u4e3b\u8981\u6709\u4e24\u7c7b\uff1a<\/p>\n

1. \u53cc\u7f16\u7801\u5668\u3002\u4ee5 BERT \u6a21\u578b\u4e3a\u4f8b\uff0c\u5b83\u9488\u5bf9\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u5206\u522b\u7f16\u7801\uff0c\u6700\u540e\u518d\u7ecf\u8fc7\u4e00\u4e2a Pooling \u5c42\uff0c\u4f7f\u5f97\u8f93\u51fa\u4ec5\u5305\u542b\u4e00\u4e2a\u5411\u91cf\u3002\u5728\u67e5\u8be2<\/mark>\u65f6\u7684 Ranking \u9636\u6bb5\uff0c\u53ea\u9700\u8981\u8ba1\u7b97\u4e24\u4e2a\u5411\u91cf\u76f8\u4f3c\u5ea6\u5373\u53ef\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u53cc\u7f16\u7801\u5668\u65e2\u53ef\u4ee5\u7528\u4e8e Ranking \u4e5f\u53ef\u4ee5\u7528\u4e8e Reranking \u9636\u6bb5\uff0c\u5411\u91cf\u641c\u7d22\u5b9e\u9645\u4e0a\u5c31\u662f\u8fd9\u79cd\u6392\u5e8f\u6a21\u578b\u3002\u7531\u4e8e\u53cc\u7f16\u7801\u5668\u9488\u5bf9\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u5206\u522b\u7f16\u7801\uff0c\u56e0\u6b64\u65e0\u6cd5\u6355\u83b7\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u7684 Token \u4e4b\u95f4\u7684\u590d\u6742\u4ea4\u4e92\u5173\u7cfb\uff0c\u5728\u8bed\u4e49\u4e0a\u4f1a\u6709\u5f88\u591a\u635f\u8017\uff0c\u4f46\u7531\u4e8e\u53ea\u9700\u8981\u5411\u91cf\u641c\u7d22\u5373\u53ef\u5b8c\u6210\u6392\u5e8f\u6253\u5206\u8ba1\u7b97\uff0c\u56e0\u6b64\u6267\u884c\u6548\u7387\u975e\u5e38\u9ad8\u3002<\/p>\n

<\/a><\/p>\n

2. \u4ea4\u53c9\u7f16\u7801\u5668\uff08Cross Encoder\uff09\u3002Cross-Encoder \u4f7f\u7528\u5355\u7f16\u7801\u5668\u6a21\u578b\u6765\u540c\u65f6\u7f16\u7801\u67e5\u8be2<\/mark>\u548c\u6587\u6863\uff0c\u5b83\u80fd\u591f\u6355\u6349\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u4e4b\u95f4\u7684\u590d\u6742\u4ea4\u4e92\u5173\u7cfb\uff0c\u56e0\u6b64\u80fd\u591f\u63d0\u4f9b\u66f4\u7cbe\u51c6\u7684\u641c\u7d22\u6392\u5e8f\u7ed3\u679c\u3002Cross-Encoder \u5e76\u4e0d\u8f93\u51fa\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u7684 Token \u6240\u5bf9\u5e94\u7684\u5411\u91cf\uff0c\u800c\u662f\u518d\u6dfb\u52a0\u4e00\u4e2a\u5206\u7c7b\u5668\u76f4\u63a5\u8f93\u51fa\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u7684\u76f8\u4f3c\u5ea6\u5f97\u5206\u3002\u5b83\u7684\u7f3a\u70b9\u5728\u4e8e\uff0c\u7531\u4e8e\u9700\u8981\u5728\u67e5\u8be2<\/mark>\u65f6\u5bf9\u6bcf\u4e2a\u6587\u6863\u548c\u67e5\u8be2<\/mark>\u5171\u540c\u7f16\u7801\uff0c\u8fd9\u4f7f\u5f97\u6392\u5e8f\u7684\u901f\u5ea6\u975e\u5e38\u6162\uff0c\u56e0\u6b64 Cross-Encoder \u53ea\u80fd\u7528\u4e8e\u6700\u7ec8\u7ed3\u679c\u7684\u91cd\u6392\u5e8f\u3002\u4f8b\u5982\u9488\u5bf9\u521d\u7b5b\u7ed3\u679c\u7684 Top 10 \u505a\u91cd\u6392\u5e8f\uff0c\u4ecd\u7136\u9700\u8981\u8017\u65f6\u79d2\u7ea7\u624d\u53ef\u4ee5\u5b8c\u6210\u3002<\/p>\n

<\/a><\/p>\n

\u4eca\u5e74\u4ee5\u6765\uff0c\u53e6\u4e00\u7c7b\u4ee5 ColBERT\u3010\u53c2\u8003\u6587\u732e1\u3011 \u4e3a\u4ee3\u8868\u7684\u5de5\u4f5c\uff0c\u5728 RAG \u5f00\u53d1\u793e\u533a\u5f15\u8d77\u4e86\u5e7f\u6cdb\u5173\u6ce8\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5b83\u5177\u5907\u4e00\u4e9b\u663e\u8457\u533a\u5206\u4e8e\u4ee5\u4e0a\u4e24\u7c7b\u6392\u5e8f\u6a21\u578b\u7684\u7279\u70b9\uff1a<\/p>\n

\u5176\u4e00\u662f\u76f8\u6bd4\u4e8e Cross Encoder\uff0cColBERT \u4ecd\u91c7\u7528\u53cc\u7f16\u7801\u5668\u7b56\u7565\uff0c\u5c06\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u5206\u522b\u91c7\u7528\u72ec\u7acb\u7684\u7f16\u7801\u5668\u7f16\u7801\uff0c\u56e0\u6b64\u67e5\u8be2<\/mark>\u7684 Token \u548c\u6587\u6863\u7684 Token \u5728\u7f16\u7801\u65f6\u4e92\u4e0d\u5f71\u54cd\uff0c\u8fd9\u79cd\u5206\u79bb\u4f7f\u5f97\u6587\u6863\u7f16\u7801\u53ef\u4ee5\u79bb\u7ebf\u5904\u7406\uff0c\u67e5\u8be2<\/mark>\u65f6\u4ec5\u9488\u5bf9 Query \u7f16\u7801\uff0c\u56e0\u6b64\u5904\u7406\u7684\u901f\u5ea6\u5927\u5927\u9ad8\u4e8e Cross Encoder\uff1b<\/p>\n

\u5176\u4e8c\u662f\u76f8\u6bd4\u4e8e\u53cc\u7f16\u7801\u5668\uff0cColBERT \u8f93\u51fa\u7684\u662f\u591a\u5411\u91cf\u800c\u975e\u5355\u5411\u91cf\uff0c\u8fd9\u662f\u4ece Transformer \u7684\u6700\u540e\u8f93\u51fa\u5c42\u76f4\u63a5\u83b7\u5f97\u7684\uff0c\u800c\u53cc\u7f16\u7801\u5668\u5219\u901a\u8fc7\u4e00\u4e2a Pooling \u5c42\u628a\u591a\u4e2a\u5411\u91cf\u8f6c\u6210\u4e00\u4e2a\u5411\u91cf\u8f93\u51fa\uff0c\u56e0\u6b64\u4e22\u5931\u4e86\u90e8\u5206\u8bed\u4e49\u3002<\/p>\n

\u5728\u6392\u5e8f\u8ba1\u7b97\u65f6\uff0cColBERT \u5f15\u5165\u4e86\u5ef6\u8fdf\u4ea4\u4e92\u8ba1\u7b97\u76f8\u4f3c\u5ea6\u51fd\u6570\uff0c\u5e76\u5c06\u5176\u547d\u540d\u4e3a\u6700\u5927\u76f8\u4f3c\u6027\uff08MaxSim\uff09\uff0c\u8ba1\u7b97\u65b9\u6cd5\u5982\u4e0b\uff1a\u5bf9\u4e8e\u6bcf\u4e2a\u67e5\u8be2<\/mark> Token \u7684\u5411\u91cf\u90fd\u8981\u4e0e\u6240\u6709\u6587\u6863 Token \u5bf9\u5e94\u7684\u5411\u91cf\u8fdb\u884c\u76f8\u4f3c\u5ea6\u8ba1\u7b97\uff0c\u5e76\u8ddf\u8e2a\u6bcf\u4e2a\u67e5\u8be2<\/mark> Token \u7684\u6700\u5927\u5f97\u5206\u3002\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u7684\u603b\u5206\u5c31\u662f\u8fd9\u4e9b\u6700\u5927\u4f59\u5f26\u5206\u6570\u7684\u603b\u548c\u3002\u4f8b\u5982\u5bf9\u4e8e\u4e00\u4e2a\u6709 32 \u4e2a Token \u5411\u91cf\u7684\u67e5\u8be2<\/mark>\uff08\u6700\u5927\u67e5\u8be2<\/mark>\u957f\u5ea6\u4e3a 32\uff09\u548c\u4e00\u4e2a\u6709 128 \u4e2a Token \u7684\u6587\u6863\uff0c\u9700\u8981\u6267\u884c 32*128 \u6b21\u76f8\u4f3c\u6027\u64cd\u4f5c\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n

\u56e0\u6b64\u76f8\u6bd4\u4e4b\u4e0b\uff0c Cross Encoder \u53ef\u4ee5\u79f0\u4f5c\u65e9\u671f\u4ea4\u4e92\u6a21\u578b \uff08Early Interaction Model\uff09<\/strong>\uff0c\u800c\u4ee5 ColBERT \u4e3a\u4ee3\u8868\u7684\u5de5\u4f5c\u53ef\u79f0\u4e3a\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff08Late Interaction Model\uff09\u3002<\/strong><\/p>\n

<\/a><\/p>\n

\u4e0b\u56fe\u4ece\u6027\u80fd\u548c\u6392\u5e8f\u8d28\u91cf\u4e0a\uff0c\u5206\u522b\u5bf9\u4ee5\u4e0a\u6392\u5e8f\u6a21\u578b\u8fdb\u884c\u5bf9\u6bd4\u3002\u7531\u4e8e\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\u6ee1\u8db3\u4e86\u5bf9\u6392\u5e8f\u8fc7\u7a0b\u4e2d\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u4e4b\u95f4\u590d\u6742\u4ea4\u4e92\u7684\u6355\u83b7\uff0c\u540c\u65f6\u4e5f\u907f\u514d\u4e86\u5bf9\u6587\u6863 Token \u7f16\u7801\u7684\u5f00\u9500\uff0c\u56e0\u6b64\u65e2\u80fd\u4fdd\u8bc1\u826f\u597d\u7684\u6392\u5e8f\u6548\u679c\uff0c\u4e5f\u80fd\u5b9e\u73b0\u8f83\u5feb\u7684\u6392\u5e8f\u6027\u80fd \u2014\u2014 \u76f8\u540c\u6570\u636e\u89c4\u6a21\u4e0b\uff0c ColBERT \u7684\u6548\u7387\u53ef\u8fbe Cross Encoder \u7684 100 \u500d\u4ee5\u4e0a\u3002\u56e0\u6b64\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\u662f\u4e00\u79cd\u975e\u5e38\u6709\u524d\u666f\u7684\u6392\u5e8f\u6a21\u578b\uff0c\u4e00\u4e2a\u5929\u7136\u7684\u60f3\u6cd5\u662f\uff1a\u80fd\u5426\u5728 RAG \u4e2d\u76f4\u63a5\u91c7\u7528\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\u66ff\u4ee3\u5411\u91cf\u641c\u7d22 + \u7cbe\u6392\u8fd9\u6837\u7684\u4e24\u9636\u6bb5\u6392\u5e8f\u67b6\u6784\uff1f<\/strong><\/section>\n

<\/a><\/p>\n

\u4e3a\u6b64\uff0c\u6211\u4eec\u9700\u8981\u8003\u8651 ColBERT \u5de5\u7a0b\u5316\u7684\u4e00\u4e9b\u95ee\u9898\uff1a<\/p>\n

1. ColBERT \u7684 MaxSim \u5ef6\u8fdf\u4ea4\u4e92\u76f8\u4f3c\u5ea6\u51fd\u6570\uff0c\u8ba1\u7b97\u6548\u7387\u5927\u5927\u9ad8\u4e8e Cross Encoder\uff0c\u4f46\u76f8\u6bd4\u666e\u901a\u5411\u91cf\u641c\u7d22\uff0c\u8ba1\u7b97\u5f00\u9500\u4ecd\u7136\u5f88\u5927\uff1a\u56e0\u4e3a\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\uff0c\u662f\u591a\u5411\u91cf\u8ba1\u7b97\uff0c\u56e0\u6b64 MaxSim \u7684\u5f00\u9500\u662f\u666e\u901a\u5411\u91cf\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u7684 M * N \u500d \uff08M \u4e3a\u67e5\u8be2<\/mark>\u7684 Token \u6570\uff0c N \u4e3a \u6587\u6863\u7684 Token \u6570\uff09\u3002\u9488\u5bf9\u8fd9\u4e9b\uff0cColBERT \u4f5c\u8005\u5728 2021 \u5e74\u63a8\u51fa\u4e86 ColBERT v2 \u3010\u53c2\u8003\u6587\u732e 2\u3011\uff0c\u901a\u8fc7 Cross Encoder \u548c\u6a21\u578b\u84b8\u998f\uff0c\u6539\u8fdb\u4e86\u751f\u6210\u7684 Embedding \u8d28\u91cf\uff0c\u5e76\u4e14\u91c7\u7528\u538b\u7f29\u6280\u672f\uff0c\u5bf9\u751f\u6210\u7684\u6587\u6863\u5411\u91cf\u8fdb\u884c\u91cf\u5316<\/mark>\uff0c\u4ece\u800c\u6539\u5584 MaxSim \u7684\u8ba1\u7b97\u6027\u80fd\u3002\u57fa\u4e8e ColBERT v2 \u5305\u88c5\u7684\u9879\u76ee RAGatouille \u3010\u53c2\u8003\u6587\u732e 3\u3011\u6210\u4e3a\u9ad8\u8d28\u91cf RAG \u6392\u5e8f\u7684\u89e3\u51b3\u65b9\u6848\u3002\u7136\u800c\uff0cColBERT v2 \u53ea\u662f\u4e00\u4e2a\u7b97\u6cd5\u5e93\uff0c\u7aef\u5230\u7aef\u7684\u8ba9\u5b83\u5728\u4f01\u4e1a\u7ea7 RAG \u7cfb\u7edf\u4f7f\u7528\uff0c\u4ecd\u7136\u662f\u4e00\u4ef6\u56f0\u96be\u7684\u4e8b\u60c5\u3002<\/p>\n

2. \u7531\u4e8e ColBERT \u662f\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u800c\u8bad\u7ec3\u6570\u636e\u6765\u81ea\u4e8e\u641c\u7d22\u5f15\u64ce\u7684\u67e5\u8be2<\/mark>\u548c\u8fd4\u56de\u7ed3\u679c\uff0c\u8fd9\u4e9b\u6587\u672c\u6570\u636e\u5e76\u4e0d\u5927\uff0c\u4f8b\u5982\u67e5\u8be2<\/mark> Token \u6570 32 \uff0c \u6587\u6863 Token \u6570 128 \u662f\u5178\u578b\u7684\u957f\u5ea6\u9650\u5236\u3002\u56e0\u6b64\u5c06 ColBERT \u7528\u4e8e\u771f\u5b9e\u6570\u636e\u65f6\uff0c \u8d85\u8fc7\u9650\u5236\u7684\u957f\u5ea6\u4f1a\u88ab\u622a\u65ad\uff0c\u8fd9\u5bf9\u4e8e\u957f\u6587\u6863\u68c0\u7d22\u5e76\u4e0d\u53cb\u597d\u3002<\/p>\n

\u57fa\u4e8e\u4ee5\u4e0a\u95ee\u9898\uff0c \u5f00\u6e90 AI \u539f\u751f\u6570\u636e\u5e93<\/mark> Infinity \u5728\u6700\u65b0\u7248\u672c\u4e2d\u63d0\u4f9b\u4e86 Tensor \u6570\u636e\u7c7b\u578b\uff0c\u5e76\u539f\u751f\u5730\u63d0\u4f9b\u7aef\u5230\u7aef\u7684 ColBERT \u65b9\u6848\u3002\u5f53 Tensor \u4f5c\u4e3a\u4e00\u79cd\u6570\u636e\u7c7b\u578b\uff0cColBERT \u7f16\u7801\u8f93\u51fa\u7684\u591a\u4e2a\u5411\u91cf\uff0c\u5c31\u53ef\u4ee5\u76f4\u63a5\u7528\u4e00\u4e2a Tensor \u6765\u5b58\u653e\uff0c\u56e0\u6b64 Tensor \u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u5c31\u53ef\u4ee5\u76f4\u63a5\u5f97\u51fa MaxSim \u6253\u5206\u3002\u9488\u5bf9 MaxSim \u8ba1\u7b97\u91cf\u5927\u7684\u95ee\u9898\uff0cInfinity \u7ed9\u51fa\u4e86 2 \u4e2a\u65b9\u6848\u6765\u4f18\u5316\uff1a\u5176\u4e00\u79cd\u662f binary \u91cf\u5316<\/mark>\uff0c\u5b83\u53ef\u4ee5\u8ba9\u539f\u59cb Tensor \u7684\u7a7a\u95f4\u53ea\u9700\u539f\u59cb\u5c3a\u5bf8\u7684 1\/32 \uff0c \u4f46\u5e76\u4e0d\u6539\u53d8 MaxSim \u8ba1\u7b97\u7684\u76f8\u5bf9\u6392\u5e8f\u7ed3\u679c\u3002\u8fd9\u79cd\u65b9\u6848\u4e3b\u8981\u7528\u4e8e Reranker\uff0c\u56e0\u4e3a\u9700\u8981\u6839\u636e\u524d\u4e00\u9636\u6bb5\u7c97\u7b5b\u7684\u7ed3\u679c\u53d6\u51fa\u5bf9\u5e94\u7684 Tensor \u3002\u53e6\u4e00\u79cd\u662f Tensor Index\uff0cColBERTv2 \u5b9e\u9645\u4e0a\u5c31\u662f ColBERT \u4f5c\u8005\u63a8\u51fa\u7684 Tensor Index \u5b9e\u73b0\uff0cInfinity \u91c7\u7528\u7684\u5219\u662f EMVB\u3010\u53c2\u8003\u6587\u732e 4\u3011\uff0c\u5b83\u53ef\u4ee5\u770b\u4f5c\u662f ColBERT v2 \u7684\u6539\u8fdb\uff0c\u4e3b\u8981\u901a\u8fc7\u91cf\u5316<\/mark>\u548c\u9884\u8fc7\u6ee4\u6280\u672f\uff0c\u5e76\u5728\u5173\u952e\u64cd\u4f5c\u4e0a\u5f15\u5165 SIMD \u6307\u4ee4\u6765\u52a0\u901f\u5b9e\u73b0\u3002Tensor Index \u53ea\u80fd\u7528\u6765\u670d\u52a1 Ranker \u800c\u975e Reranker\u3002\u6b64\u5916\uff0c\u9488\u5bf9\u8d85\u8fc7 Token \u9650\u5236\u7684\u957f\u6587\u672c\uff0cInfinity \u5f15\u5165\u4e86 Tensor Array \u7c7b\u578b\uff1a<\/p>\n

<\/a><\/p>\n

\u4e00\u7bc7\u8d85\u8fc7 ColBERT \u9650\u5236\u7684\u6587\u6863\uff0c\u4f1a\u88ab\u5207\u5206\u6210\u591a\u4e2a\u6bb5\u843d\uff0c\u5206\u522b\u7f16\u7801\u751f\u6210 Tensor \u540e\uff0c\u90fd\u8ddf\u539f\u59cb\u6587\u6863\u4fdd\u5b58\u5728\u4e00\u884c\u3002\u8ba1\u7b97 MaxSim \u7684\u65f6\u5019\uff0c\u67e5\u8be2<\/mark>\u8ddf\u8fd9\u4e9b\u6bb5\u843d\u5206\u522b\u8ba1\u7b97\uff0c\u7136\u540e\u53d6\u6700\u5927\u503c\u4f5c\u4e3a\u6574\u4e2a\u6587\u6863\u7684\u6253\u5206\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n

<\/a><\/p>\n

\u56e0\u6b64\uff0c\u91c7\u7528 Infinity\uff0c\u53ef\u4ee5\u7aef\u5230\u7aef\u5730\u5f15\u5165\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\u9ad8\u8d28\u91cf\u5730\u670d\u52a1 RAG\u3002\u90a3\u4e48\uff0c\u5e94\u8be5\u662f\u91c7\u7528 ColBERT \u4f5c\u4e3a Ranker \uff0c\u8fd8\u662f Reranker \u5462\uff1f\u4e0b\u8fb9\u6211\u4eec\u91c7\u7528 Infinity \u6765\u5728\u771f\u5b9e\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bc4\u6d4b\u3002\u7531\u4e8e Infinity \u7684\u6700\u65b0\u7248\u672c\u5b9e\u73b0\u4e86\u6709\u53f2\u4ee5\u6765\u6700\u5168\u7684\u6df7\u5408\u641c\u7d22\u65b9\u6848\uff0c\u53ec\u56de\u624b\u6bb5\u5305\u542b\u5411\u91cf\u641c\u7d22\u3001\u5168\u6587\u641c\u7d22\u3001\u7a00\u758f\u5411\u91cf\u641c\u7d22\uff0c\u4e0a\u6587\u6240\u8ff0\u7684 Tensor \uff0c\u4ee5\u53ca\u8fd9\u4e9b\u624b\u6bb5\u7684\u4efb\u610f\u7ec4\u5408\uff0c\u5e76\u4e14\u63d0\u4f9b\u4e86\u591a\u79cd Reranker \u624b\u6bb5\uff0c\u5982 RRF\uff0c\u4ee5\u53ca ColBERT Reranker \u7b49\uff0c\u56e0\u6b64\u6211\u4eec\u5728\u8bc4\u6d4b\u4e2d\u5305\u542b\u4e86\u5404\u79cd\u6df7\u5408\u641c\u7d22\u548c Reranker \u7684\u7ec4\u5408\u3002<\/p>\n

\u6211\u4eec\u91c7\u7528 MLDR \u6570\u636e\u96c6\u8fdb\u884c\u8bc4\u6d4b\u3002MLDR \u662f MTEB \u3010\u53c2\u8003\u6587\u732e 5\u3011\u7528\u6765\u8bc4\u6d4b Embedding \u6a21\u578b\u8d28\u91cf\u7684 benchmark \u96c6\uff0c\u5176\u4e2d MLDR \u662f\u5176\u4e2d\u4e00\u4e2a\u6570\u636e\u96c6\uff0c\u5168\u79f0\u4e3a Multi Long Document Retrieval\uff0c\u4e00\u5171\u5305\u542b 20 \u4e07\u957f\u6587\u672c\u6570\u636e\u3002\u8bc4\u6d4b\u91c7\u7528 BGE-M3\u3010\u53c2\u8003\u6587\u732e 6\u3011\u4f5c\u4e3a Embedding \u6a21\u578b\uff0c\u91c7\u7528 Jina-ColBERT \u3010\u53c2\u8003\u6587\u732e 7\u3011\u6765\u751f\u6210 Tensor\uff0c\u8bc4\u6d4b\u811a\u672c\u4e5f\u653e\u5230\u4e86 Infinity \u4ed3\u5e93\u3010\u53c2\u8003\u6587\u732e 8\u3011\u3002<\/p>\n

\u8bc4\u6d4b\u4e00\uff1aColBERT \u4f5c\u4e3a Reranker \u662f\u5426\u6709\u6548\u3002<\/strong>\u5c06 20 \u4e07 MLDR \u6570\u636e\u5206\u522b\u7528 BGE-M3 \u751f\u6210\u7a20\u5bc6\u5411\u91cf\u548c\u7a00\u758f\u5411\u91cf\uff0c\u5e76\u63d2\u5165\u5230 Infinity \u6570\u636e\u5e93<\/mark>\u4e2d\uff0c\u6570\u636e\u5e93<\/mark>\u5305\u542b 4 \u5217\uff0c\u5206\u522b\u4fdd\u5b58\u539f\u59cb\u6587\u672c\uff0c\u5411\u91cf\uff0c\u7a00\u758f\u5411\u91cf\uff0c\u4ee5\u53ca Tensor\uff0c\u5e76\u5206\u522b\u6784\u5efa\u76f8\u5e94\u5168\u6587\u7d22\u5f15\u3001\u5411\u91cf\u7d22\u5f15\u3001\u7a00\u758f\u5411\u91cf\u7d22\u5f15\u3002\u8bc4\u6d4b\u5305\u542b\u6240\u6709\u7684\u53ec\u56de\u7ec4\u5408\uff0c\u5305\u542b\u5355\u8def\u53ec\u56de\u3001\u53cc\u8def\u53ec\u56de\uff0c\u4ee5\u53ca\u4e09\u8def\u53ec\u56de\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n

<\/a><\/p>\n

\u8bc4\u6d4b\u6307\u6807\u91c7\u7528 nDCG@10\u3002\u5176\u4ed6\u53c2\u6570<\/mark>\uff1a\u91c7\u7528 RRF Reranker \u65f6\u7c97\u7b5b\u8fd4\u56de\u7684 Top N = 1000 \uff0c\u67e5\u8be2<\/mark>\u7d2f\u8ba1\u5171\u6709 800 \u6761\uff0c\u5e73\u5747\u6bcf\u6761\u67e5\u8be2<\/mark>\u957f\u5ea6\u5728 10 \u4e2a token \u5de6\u53f3\u3002<\/p>\n

<\/a><\/p>\n

\u4ece\u56fe\u4e2d\u770b\u5230\uff0c\u6240\u6709\u7684\u53ec\u56de\u65b9\u6848\uff0c\u5728\u91c7\u7528\u4e86 ColBERT Reranker \u4e4b\u540e\uff0c\u90fd\u6709\u660e\u663e\u7684\u6548\u679c\u63d0\u5347\u3002ColBERT \u4f5c\u4e3a\u4e00\u79cd\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u5b83\u53ef\u4ee5\u63d0\u4f9b\u8ddf\u5728 MTEB \u7684 Reranker \u6392\u884c\u699c\u4e0a\u4f4d\u5c45\u524d\u5217\u76f8\u63d0\u5e76\u8bba\u7684\u6392\u5e8f\u8d28\u91cf\uff0c\u4f46\u662f\u6027\u80fd\u5374\u662f\u5b83\u4eec\u7684 100 \u500d\uff0c\u6240\u4ee5\u53ef\u4ee5\u5728\u66f4\u5927\u7684\u8303\u56f4\u5185\u8fdb\u884c\u91cd\u6392\u5e8f\u3002\u56fe\u4e2d\u7ed9\u51fa\u7684\u7ed3\u679c\u662f\u9488\u5bf9 Top 100 \u8fdb\u884c Reranker\uff0c\u800c\u91c7\u7528 Top 1000 \u8fdb\u884c ColBERT \u91cd\u6392\u5e8f\uff0c\u6570\u503c\u6ca1\u6709\u660e\u663e\u53d8\u5316\uff0c\u6027\u80fd\u8fd8\u6709\u660e\u663e\u4e0b\u964d\uff0c\u56e0\u6b64\u4e0d\u63a8\u8350\u91c7\u7528\u3002\u4f20\u7edf\u4e0a\u91c7\u7528\u57fa\u4e8e Cross Encoder \u7684\u5916\u90e8 Reranker \uff0cTop 10 \u5c31\u4f1a\u6709\u79d2\u7ea7\u7684\u5ef6\u8fdf\uff0c\u800c Infinity \u5185\u90e8\u5b9e\u73b0\u4e86\u9ad8\u6027\u80fd\u7684 ColBERT Reranker\uff0c\u5373\u4f7f\u9488\u5bf9 Top 100 \u751a\u81f3 Top 1000 \u505a\u91cd\u6392\u5e8f\uff0c\u4e5f\u4e0d\u4f1a\u5f71\u54cd\u7528\u6237\u4f53\u9a8c\uff0c\u800c\u53ec\u56de\u7684\u8303\u56f4\u5374\u5927\u5927\u589e\u52a0\uff0c\u56e0\u6b64\u53ef\u4ee5\u663e\u8457\u6539\u8fdb\u6700\u7ec8\u7684\u6392\u5e8f\u6548\u679c\u3002\u6b64\u5916\uff0c\u8fd9\u79cd ColBERT Reranker \u8ba1\u7b97\u53ea\u9700\u5728\u7eaf CPU \u67b6\u6784\u4e0a\u5373\u53ef\u8fd0\u884c\uff0c\u8fd9\u4e5f\u5927\u5927\u964d\u4f4e\u4e86\u90e8\u7f72\u7684\u6210\u672c\u3002<\/p>\n

\u8bc4\u6d4b\u4e8c\uff1a\u5bf9\u6bd4\u57fa\u4e8e ColBERT \u4f5c\u4e3a Ranker \u800c\u4e0d\u662f Reranker\u3002<\/strong>\u56e0\u6b64\uff0c\u8fd9\u65f6\u9700\u8981\u9488\u5bf9 Tensor \u8fd9\u5217\u6570\u636e\u6784\u5efa Tensor Index\u3002\u540c\u65f6\uff0c\u4e3a\u4e86\u8bc4\u4f30 Tensor Index \u5f15\u5165\u7684\u7cbe\u5ea6\u635f\u8017\uff0c\u8fd8\u8fdb\u884c\u4e86\u66b4\u529b\u641c\u7d22\u3002<\/p>\n

<\/a><\/p>\n

\u53ef\u4ee5\u770b\u5230\uff0c\u76f8\u6bd4 Reranker \uff0c\u5373\u4f7f\u662f\u91c7\u7528\u6ca1\u6709\u7cbe\u5ea6\u635f\u5931\u7684\u66b4\u529b\u641c\u7d22\uff0c\u4e5f\u6ca1\u6709\u663e\u8457\u7684\u63d0\u5347\uff0c\u800c\u91c7\u7528\u57fa\u4e8e Tensor Index \u7684\u6392\u5e8f\u8d28\u91cf\u751a\u81f3\u4f4e\u4e8e\u91c7\u7528 Reranker\u3002\u7136\u800c\uff0c\u4f5c\u4e3a Ranker \u7684\u67e5\u8be2<\/mark>\u65f6\u95f4\u5374\u8981\u6162\u5f97\u591a\uff1aMLDR \u6570\u636e\u96c6\u5305\u542b 20 \u4e07\u6587\u6863\u6570\u636e\uff0c\u5927\u7ea6 2GB \u5de6\u53f3\uff0c\u91c7\u7528 Jina-ColBERT \u8f6c\u6210 Tensor \u6570\u636e\u540e\uff0c\u9ad8\u8fbe 320 G\uff0c\u8fd9\u662f\u56e0\u4e3a Tensor \u6570\u636e\u7c7b\u578b\u662f\u628a\u4e00\u7bc7\u6587\u6863\u7684\u6bcf\u4e2a Token \u5bf9\u5e94\u7684\u5411\u91cf\u90fd\u8981\u4fdd\u5b58\u4e0b\u6765\uff0c ColBERT \u6a21\u578b\u7684\u7ef4\u5ea6\u662f 128 \u7ef4\uff0c\u56e0\u6b64\u9ed8\u8ba4\u6570\u636e\u91cf\u4f1a\u81a8\u80c0 2 \u4e2a\u6570\u91cf\u7ea7\uff0c\u5373\u4f7f\u6784\u5efa\u4e86 Tensor Index\uff0c\u5728\u67e5\u8be2<\/mark>\u8fd9\u4e48\u591a\u6570\u636e\u7684\u65f6\u5019\uff0c\u4e5f\u9700\u8981\u5e73\u5747 7s \u624d\u80fd\u8fd4\u56de\u4e00\u4e2a\u67e5\u8be2<\/mark>\uff0c\u4f46\u5f97\u5230\u7684\u7ed3\u679c\u5374\u5e76\u6ca1\u6709\u66f4\u597d\u3002<\/section>\n

\u56e0\u6b64\uff0c\u5f88\u663e\u7136\uff0cColBERT \u4f5c\u4e3a Reranker \u7684\u6536\u76ca\u6bd4\u4f5c\u4e3a Ranker \u8981\u9ad8\u5f97\u591a\u3002\u5f53\u524d\u6700\u4f73\u7684 RAG \u68c0\u7d22\u65b9\u6848\uff0c\u662f\u5728 3 \u8def\u6df7\u5408\u641c\u7d22\uff08\u5168\u6587\u641c\u7d22 + \u5411\u91cf + \u7a00\u758f\u5411\u91cf\uff09\u7684\u57fa\u7840\u4e0a\u52a0 ColBERT Reranker\u3002\u6709\u4f19\u4f34\u53ef\u80fd\u4f1a\u95ee\u4e86\uff0c\u4e3a\u4e86\u91c7\u7528 ColBERT Reranker\uff0c\u5c31\u9700\u8981\u589e\u52a0\u5355\u72ec\u7684 Tensor \u5217\uff0c\u5e76\u4e14\u8be5\u5217\u4f1a\u76f8\u6bd4\u539f\u59cb\u6570\u636e\u96c6\u81a8\u80c0 2 \u4e2a\u6570\u91cf\u7ea7\uff0c\u8fd9\u6837\u505a\u662f\u5426\u503c\u5f97\uff1f\u9996\u5148\uff1aInfinity \u9488\u5bf9 Tensor \u63d0\u4f9b\u4e86 Binary \u91cf\u5316<\/mark>\u624b\u6bb5\uff0c\u4f5c\u4e3a Reranker\uff0c\u5b83\u5e76\u4e0d\u5f71\u54cd\u6392\u5e8f\u7ed3\u679c\u5f88\u591a\uff0c\u4f46\u5374\u53ef\u4ee5\u8ba9\u6700\u7ec8\u7684\u6570\u636e\u4ec5\u6709\u539f\u59cb Tensor \u5927\u5c0f\u7684 1\/32\u3002\u5176\u6b21\uff0c\u5373\u4fbf\u5982\u6b64\uff0c\u4e5f\u4f1a\u6709\u4eba\u8ba4\u4e3a\u8fd9\u6837\u7684\u5f00\u9500\u8fc7\u9ad8\u3002\u7136\u800c\u7ad9\u5728\u4f7f\u7528\u8005\u7684\u89c6\u89d2\uff0c\u7528\u66f4\u591a\u7684\u5b58\u50a8\uff0c\u6765\u6362\u53d6\u66f4\u9ad8\u7684\u6392\u5e8f\u8d28\u91cf\u548c\u66f4\u5ec9\u4ef7\u7684\u6210\u672c\uff08\u6392\u5e8f\u8fc7\u7a0b\u65e0\u9700 GPU\uff09\uff0c\u8fd9\u6837\u505a\u4f9d\u7136\u662f\u975e\u5e38\u503c\u5f97\u7684\u3002\u6700\u540e\uff0c\u76f8\u4fe1\u5f88\u5feb\u5c31\u53ef\u4ee5\u63a8\u51fa\u6548\u679c\u4e0a\u7565\u6709\u4e0b\u964d\uff0c\u4f46\u5b58\u50a8\u5f00\u9500\u5927\u5927\u964d\u4f4e\u7684 Late Interaction \u6a21\u578b\uff0c\u4f5c\u4e3a\u4e00\u6b3e Data Infra \u57fa\u7840\u8bbe\u65bd\uff0c \u5bf9\u8fd9\u4e9b\u53d8\u5316\u4fdd\u6301\u900f\u660e\uff0c\u628a\u8fd9\u4e9b Trade Off \u4ea4\u7ed9\u7528\u6237\u662f\u660e\u667a\u7684\u9009\u62e9\u3002<\/p>\n

\u4ee5\u4e0a\u662f\u57fa\u4e8e Infinity \u5728 MLDR \u6570\u636e\u96c6\u4e0a\u7684\u591a\u8def\u53ec\u56de\u8bc4\u6d4b\uff0c\u5728\u5176\u4ed6\u6570\u636e\u96c6\u7684\u8bc4\u6d4b\u7ed3\u679c\uff0c\u53ef\u80fd\u4f1a\u6709\u6240\u4e0d\u540c\uff0c\u4f46\u6574\u4f53\u4e0a\u7ed3\u8bba\u4e0d\u4f1a\u53d8 \u2014\u2014 3 \u8def\u6df7\u5408\u641c\u7d22 + \u57fa\u4e8e Tensor \u7684\u91cd\u6392\u5e8f\uff0c\u662f\u5f53\u524d\u641c\u7d22\u7ed3\u679c\u8d28\u91cf\u6700\u9ad8\u7684\u53ec\u56de\u624b\u6bb5\u3002<\/p>\n

\u7531\u6b64\u53ef\u4ee5\u770b\u5230\uff0cColBERT \u53ca\u5176\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u5728 RAG \u573a\u666f\u5177\u6709\u5f88\u5927\u7684\u5e94\u7528\u4ef7\u503c\uff0c\u4ee5\u4e0a\u662f\u5728\u6587\u672c\u5bf9\u8bdd\u5185\u5bb9\u751f\u6210\u7684\u76f8\u5173\u5de5\u4f5c\uff0c\u8fd1\u671f\uff0c\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\u5728\u591a\u6a21\u6001\u573a\u666f\uff0c\u4e5f\u5f97\u5230\u4e86 SOTA \u7684\u7ed3\u679c\u3002\u8fd9\u5c31\u662f ColPali\u3010\u53c2\u8003\u6587\u732e 9\u3011\uff0c\u5b83\u6539\u53d8\u4e86 RAG \u7684\u5de5\u4f5c\u6d41\u7a0b\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n

<\/a><\/p>\n

RAG \u5728\u9762\u4e34\u590d\u6742\u683c\u5f0f\u6587\u6863\u65f6\uff0c\u5f53\u4e0b\u7684 SOTA \uff0c\u662f\u91c7\u7528\u6587\u6863\u8bc6\u522b\u6a21\u578b\uff0c\u5bf9\u6587\u6863\u7684\u5e03\u5c40\u505a\u8bc6\u522b\uff0c\u5e76\u9488\u5bf9\u8bc6\u522b\u51fa\u7684\u90e8\u5206\u7ed3\u6784\uff0c\u4f8b\u5982\u56fe\u8868\uff0c\u56fe\u7247\u7b49\uff0c\u518d\u5206\u522b\u8c03\u7528\u76f8\u5e94\u7684\u6a21\u578b\uff0c\u5c06\u5b83\u4eec\u8f6c\u5316\u4e3a\u5bf9\u5e94\u7684\u6587\u5b57\uff0c\u518d\u7528\u5404\u79cd\u683c\u5f0f\u4fdd\u5b58\u5230 RAG \u914d\u5957\u7684\u6570\u636e\u5e93<\/mark>\u4e2d\u3002\u800c ColPali \u5219\u7701\u6389\u4e86\u8fd9\u4e9b\u6b65\u9aa4\uff0c\u5b83\u76f4\u63a5\u91c7\u7528\u591a\u6a21\u6001\u6a21\u578b\u751f\u6210 Embedding \u5185\u5bb9\u3002\u63d0\u95ee\u7684\u65f6\u5019\uff0c\u53ef\u4ee5\u76f4\u63a5\u9488\u5bf9\u6587\u6863\u4e2d\u7684\u56fe\u8868\u8fdb\u884c\u56de\u7b54\uff1a<\/p>\n

<\/a><\/p>\n

ColPali \u6a21\u578b\u7684\u8bad\u7ec3\u8ddf ColBERT \u7c7b\u4f3c\uff0c\u4e5f\u662f\u91c7\u7528\u67e5\u8be2<\/mark> – \u6587\u6863\u9875\u9762\u5bf9\u7684\u5f62\u5f0f\uff0c\u4ece\u800c\u6355\u83b7\u67e5\u8be2<\/mark>\u548c\u6587\u6863\u591a\u6a21\u6001\u6570\u636e\u4e4b\u95f4\u7684\u8bed\u4e49\u5173\u8054\uff0c\u53ea\u662f\u91c7\u7528 PaliGemma \u3010\u53c2\u8003\u6587\u732e 10\u3011\u7528\u6765\u751f\u6210\u591a\u6a21\u6001 Embedding \u3002\u76f8\u6bd4\u6ca1\u6709\u91c7\u7528 Late Interaction \u673a\u5236\u4f46\u540c\u6837\u91c7\u7528 PaliGemma \u751f\u6210 Embedding \u7684\u65b9\u6848 BiPali\uff0c\u5728 nDCG@5 \u7684\u8bc4\u6d4b\u6307\u6807\u5bf9\u6bd4\u662f 81.3 vs 58.8\uff0c\u8fd9\u79cd\u5dee\u8ddd\u662f\u5c31\u662f \u201c\u6781\u597d\u201d \u548c \u201c\u538b\u6839\u4e0d\u80fd\u5de5\u4f5c\u201d \u7684\u533a\u522b\u3002<\/p>\n

<\/a><\/p>\n

\u56e0\u6b64\uff0c\u5c3d\u7ba1 ColBERT \u51fa\u73b0\u81f3\u4eca\u5df2\u6709 4 \u5e74\u65f6\u95f4\uff0c\u53ef\u662f Late Interaction \u6a21\u578b\u5728 RAG \u7684\u5e94\u7528\u624d\u521a\u521a\u5f00\u59cb\uff0c\u5b83\u5fc5\u5c06\u6269\u5927 RAG \u7684\u4f7f\u7528\u573a\u666f\uff0c\u5728\u5305\u542b\u591a\u6a21\u6001\u5728\u5185\u7684\u590d\u6742 RAG \u573a\u666f\u63d0\u4f9b\u9ad8\u8d28\u91cf\u7684\u8bed\u4e49\u53ec\u56de\u3002\u800c Infinity \u5df2\u7ecf\u4e3a\u5b83\u7684\u7aef\u5230\u7aef\u5e94\u7528\u505a\u597d\u4e86\u51c6\u5907\uff0c\u6b22\u8fce\u5173\u6ce8\u548c Star Infinity\uff0chttps:\/\/github.com\/infiniflow\/infinity\uff0c \u81f4\u529b\u4e8e\u6210\u4e3a\u6700\u597d\u7684 AI \u539f\u751f\u6570\u636e\u5e93<\/mark>\uff01<\/p>\n

\u53c2\u8003\u6587\u732e\u00a0<\/sup><\/em><\/p>\n

1. Colbert: Efficient and effective passage search via contextualized late interaction over bert, SIGIR 2020.<\/em><\/sup><\/p>\n

2. Colbertv2: Effective and efficient retrieval via lightweight late interaction, arXiv:2112.01488, 2021.<\/em><\/sup><\/p>\n

3. RAGatouille https:\/\/github.com\/bclavie\/RAGatouille\u00a0<\/em><\/sup><\/p>\n

4. Efficient Multi-vector Dense Retrieval with Bit Vectors, ECIR 2024.<\/em><\/sup><\/p>\n

5. https:\/\/huggingface.co\/mteb<\/em><\/sup><\/p>\n

6. https:\/\/huggingface.co\/BAAI\/bge-m3<\/em><\/sup><\/p>\n

7. https:\/\/huggingface.co\/jinaai\/jina-colbert-v1-en<\/em><\/sup><\/p>\n

8. https:\/\/github.com\/infiniflow\/infinity\/tree\/main\/python\/benchmark\/mldr_benchmark<\/em><\/sup><\/p>\n

9. ColPali: Efficient Document Retrieval with Vision Language Models, arXiv:2407.01449, 2024.<\/em><\/sup><\/p>\n

10. https:\/\/github.com\/google-research\/big_vision\/tree\/main\/big_vision\/configs\/proj\/paligemma<\/sup><\/em><\/p>\n

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f […]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","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":"default","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":""}},"footnotes":""},"categories":[27],"tags":[71,73],"class_list":["post-34941","post","type-post","status-publish","format-standard","hentry","category-news","tag-rag","tag-73"],"yoast_head":"\n\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=34941\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34941\",\"url\":\"https:\/\/17aitech.com\/?p=34941\",\"name\":\"\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34941#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34941#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-f2aea2653c834f44d3aa0bf5ee2a95d4.png\",\"datePublished\":\"2024-12-15T05:59:05+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34941#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34941\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34941#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-f2aea2653c834f44d3aa0bf5ee2a95d4.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-f2aea2653c834f44d3aa0bf5ee2a95d4.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34941#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=34941","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34941","url":"https:\/\/17aitech.com\/?p=34941","name":"\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34941#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34941#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-f2aea2653c834f44d3aa0bf5ee2a95d4.png","datePublished":"2024-12-15T05:59:05+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34941#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34941"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34941#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-f2aea2653c834f44d3aa0bf5ee2a95d4.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-f2aea2653c834f44d3aa0bf5ee2a95d4.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34941#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u5ef6\u8fdf\u4ea4\u4e92\u6a21\u578b\uff0c\u4e3a\u4ec0\u4e48\u662f\u4e0b\u4e00\u4ee3RAG\u7684\u6807\u914d\uff1f"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34941","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=34941"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34941\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34941"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34941"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34941"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}