{"id":1019,"date":"2025-04-02T12:22:36","date_gmt":"2025-04-02T04:22:36","guid":{"rendered":"https:\/\/www.tyhlw.org\/?p=1019"},"modified":"2025-04-02T12:24:27","modified_gmt":"2025-04-02T04:24:27","slug":"2025%e5%b9%b4%e4%b8%bb%e6%b5%81%e5%b5%8c%e5%85%a5%e6%a8%a1%e5%9e%8b%ef%bc%88embedding-model%ef%bc%89%e7%9a%84%e8%a7%a3%e6%9e%90%e5%b5%8c%e5%85%a5%e6%a8%a1%e5%9e%8b%e5%a6%82%e4%bd%95%e9%80%89%e6%8b%a9","status":"publish","type":"post","link":"https:\/\/www.tyhlw.org\/?p=1019","title":{"rendered":"2025\u5e74\u4e3b\u6d41\u5d4c\u5165\u6a21\u578b\uff08Embedding Model\uff09\u7684\u89e3\u6790(\u5d4c\u5165\u6a21\u578b\u5982\u4f55\u9009\u62e9)"},"content":{"rendered":"<p>\u4ee5\u4e0b\u662f2025\u5e74\u4e3b\u6d41\u5d4c\u5165\u6a21\u578b\uff08Embedding Model\uff09\u7684\u89e3\u6790\uff0c\u7ed3\u5408\u6027\u80fd\u699c\u5355\u3001\u5e94\u7528\u573a\u666f\u548c\u6587\u6863\u9002\u914d\u6027\uff0c\u4ece\u5b9e\u8df5\u89d2\u5ea6\u4e3a\u4f60\u68b3\u7406\u5173\u952e\u4fe1\u606f\uff1a<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-1020 aligncenter\" src=\"https:\/\/www.tyhlw.org\/wp-content\/uploads\/2025\/04\/\u5c4f\u5e55\u622a\u56fe_2-4-2025_12213_i0.hdslb_.com_.jpeg\" alt=\"\" width=\"1236\" height=\"768\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>\u4e00\u3001\u5d4c\u5165\u6a21\u578b\u7684\u6838\u5fc3\u4f5c\u7528<\/h3>\n<p>\u5d4c\u5165\u6a21\u578b\u901a\u8fc7\u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u9ad8\u7ef4\u5411\u91cf\uff08\u5982\"\u5fd8\u8bb0\u5bc6\u7801\"\u2192[0.3, -0.8, 1.5...]\uff09\uff0c\u5b9e\u73b0\u8bed\u4e49\u5c42\u9762\u7684\u4fe1\u606f\u68c0\u7d22\u548c\u5339\u914d\u3002\u8fd9\u79cd\u80fd\u529b\u4f7f\u5f97\u8ba1\u7b97\u673a\u4e0d\u4ec5\u80fd\u8bc6\u522b\u5b57\u9762\u5173\u952e\u8bcd\uff0c\u66f4\u80fd\u7406\u89e3\"\u8d26\u53f7\u9501\u5b9a\"\u4e0e\"\u5bc6\u7801\u91cd\u7f6e\"\u7684\u6f5c\u5728\u5173\u8054\u6027\uff0c\u662fRAG\uff08\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff09\u3001\u77e5\u8bc6\u5e93\u95ee\u7b54\u7b49AI\u5e94\u7528\u7684\u57fa\u77f3\u3002<\/p>\n<h3>\u4e8c\u3001\u5f53\u524d\u4e3b\u6d41\u5d4c\u5165\u6a21\u578b\u5206\u7c7b<\/h3>\n<h4>\uff08\u4e00\uff09\u200b\u901a\u7528\u5168\u80fd\u578b<\/h4>\n<h4>1\u3001\u200bBGE-M3<\/h4>\n<p><strong>\u7279\u70b9\uff1a<\/strong>\u5317\u4eac\u667a\u6e90\u7814\u7a76\u9662\u63a8\u51fa\u7684\u591a\u8bed\u8a00\/\u591a\u529f\u80fd\/\u591a\u7c92\u5ea6\u4e09\u5408\u4e00\u6a21\u578b\uff0c\u652f\u6301100+\u8bed\u8a00\u30018K\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u5728MTEB\u4e2d\u6587\u699c\u5355\u4f4d\u5c45\u524d\u5217<br \/>\n<strong>\u4f18\u52bf\uff1a<\/strong>\u7a20\u5bc6\/\u7a00\u758f\/\u591a\u5411\u91cf\u6df7\u5408\u68c0\u7d22\u6a21\u5f0f\uff0c\u957f\u6587\u672c\u5904\u7406\u80fd\u529b\u7a81\u51fa\uff0c\u5f00\u6e90\u514d\u8d39<br \/>\n<strong>\u9002\u7528\u573a\u666f\uff1a<\/strong>\u4f01\u4e1a\u7ea7\u77e5\u8bc6\u5e93\uff08\u5c24\u5176\u591a\u8bed\u8a00\u6df7\u5408\u6587\u6863\uff09\u3001\u9700\u8981\u590d\u6742\u63a8\u7406\u7684RAG\u7cfb\u7edf<\/p>\n<h4>2\u3001\u200bNV-Embed-v2<\/h4>\n<p><strong>\u7279\u70b9\uff1a<\/strong>\u57fa\u4e8eMistral-7B\u67b6\u6784\uff0c\u68c0\u7d22\u4efb\u52a1\u5f97\u520662.65\uff082024\u5e748\u6708\u6570\u636e\uff09<br \/>\n<strong>\u4f18\u52bf\uff1a<\/strong>\u9ad8\u7cbe\u5ea6\u68c0\u7d22\uff0c\u9002\u5408\u5bf9\u53ec\u56de\u7387\u8981\u6c42\u4e25\u683c\u7684\u573a\u666f<br \/>\n<strong>\u5c40\u9650\uff1a<\/strong>7B\u53c2\u6570\u89c4\u6a21\u9700\u8f83\u9ad8\u8ba1\u7b97\u8d44\u6e90<\/p>\n<h4>\uff08\u4e8c\uff09\u200b\u5782\u76f4\u9886\u57df\u7279\u5316\u578b<\/h4>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b\u7c7b\u578b<\/th>\n<th>\u4ee3\u8868\u6a21\u578b<\/th>\n<th>\u6587\u6863\u9002\u914d\u573a\u666f<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u200b<strong>\u4e2d\u6587\u7279\u5316<\/strong><\/td>\n<td>BGE-large-zh-v1.5<\/td>\n<td>\u7eaf\u4e2d\u6587\u5408\u540c\/\u653f\u7b56\u6587\u4ef6<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>M3E-base<\/td>\n<td>\u4e2d\u6587\u5ba2\u670d\u5bf9\u8bdd\/\u793e\u4ea4\u5a92\u4f53\u6587\u672c\u5206\u6790<\/td>\n<\/tr>\n<tr>\n<td>\u200b<strong>\u82f1\u6587\u7279\u5316<\/strong><\/td>\n<td>bge-large-en-v1.5<\/td>\n<td>\u82f1\u6587\u79d1\u7814\u8bba\u6587\/\u6280\u672f\u6587\u6863<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>text-embedding-3-large<\/td>\n<td>\u9700\u4ed8\u8d39\u4f46\u6027\u80fd\u9876\u5c16\u7684\u82f1\u6587\u957f\u6587\u6863\u5904\u7406<\/td>\n<\/tr>\n<tr>\n<td>\u200b<strong>\u591a\u8bed\u8a00\u6df7\u5408<\/strong><\/td>\n<td>multilingual-e5-large<\/td>\n<td>\u8de8\u5883\u7535\u5546\u591a\u8bed\u5546\u54c1\u63cf\u8ff0<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>BGE-M3<\/td>\n<td>\u8de8\u56fd\u4f01\u4e1a\u591a\u8bed\u8a00\u77e5\u8bc6\u5e93<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>\uff08\u4e09\uff09\u200b**\u8f7b\u91cf\u5316\u90e8\u7f72\u578b<\/h4>\n<p><strong>\u200bnomic-embed-text<\/strong><\/p>\n<p>\u7279\u70b9\uff1aOllama\u5e73\u53f0\u70ed\u95e8\u6a21\u578b\uff0c\u4ec5768\u7ef4\u5411\u91cf<br \/>\n\u4f18\u52bf\uff1a\u672c\u5730\u90e8\u7f72\u53cb\u597d\uff0c\u63a8\u7406\u901f\u5ea6\u6bd4OpenAI\u5feb3\u500d<br \/>\n\u9002\u7528\uff1a\u8fb9\u7f18\u8bbe\u5907\/\u4e2d\u5c0f\u578b\u77e5\u8bc6\u5e93<\/p>\n<p><strong>\u200bgte-qwen2-1.5b-instruct<\/strong><\/p>\n<p>\u7279\u70b9\uff1a1.5B\u53c2\u6570\u89c4\u6a21\uff0c\u652f\u6301\u6307\u4ee4\u5fae\u8c03<br \/>\n\u4f18\u52bf\uff1a\u572816G\u663e\u5b58\u8bbe\u5907\u5373\u53ef\u6d41\u7545\u8fd0\u884c<br \/>\n\u9002\u7528\uff1a\u521d\u521b\u56e2\u961f\u539f\u578b\u9a8c\u8bc1<\/p>\n<h3>\u4e09\u3001\u6587\u6863\u7c7b\u578b\u4e0e\u6a21\u578b\u5339\u914d\u6307\u5357<\/h3>\n<h4>1. \u200b\u6cd5\u5f8b\/\u533b\u7597\u6587\u6863<\/h4>\n<p>\u200b\u63a8\u8350\u6a21\u578b\uff1aBGE-M3 + bge-reranker-v2.5<br \/>\n\u7406\u7531\uff1a\u4e13\u4e1a\u672f\u8bed\u5bc6\u96c6\u573a\u666f\u9700\u591a\u7c92\u5ea6\u68c0\u7d22\u80fd\u529b\uff0c\u6df7\u5408\u68c0\u7d22\u6a21\u5f0f\u63d0\u5347\u67e5\u51c6\u7387<\/p>\n<h4>2. \u200b\u591a\u6a21\u6001\u6587\u6863<\/h4>\n<p>\u200b\u63a8\u8350\u6a21\u578b\uff1aBGE-VL\uff08\u652f\u6301\u56fe\u6587\u8de8\u6a21\u6001\u68c0\u7d22\uff09<br \/>\n\u64cd\u4f5c\uff1a\u5c06\u56fe\u7247OCR\u6587\u672c\u4e0e\u56fe\u50cf\u7279\u5f81\u5411\u91cf\u8054\u5408\u7f16\u7801<\/p>\n<h4>3. \u200b\u957f\u6587\u672c\u6280\u672f\u6587\u6863<\/h4>\n<p><strong>\u200b\u63a8\u8350\u7ec4\u5408<\/strong><br \/>\n<strong>\u57fa\u7840\u5d4c\u5165\uff1a<\/strong>BGE-M3\uff08\u5904\u74068K tokens\uff09<br \/>\n<strong>\u589e\u5f3a\u65b9\u6848\uff1a<\/strong>Activation Beacon\u6280\u672f\u6269\u5c55\u81f332K\u4e0a\u4e0b\u6587<\/p>\n<h4>4. \u200b\u793e\u4ea4\u5a92\u4f53\u77ed\u6587\u672c<\/h4>\n<p><strong>\u200b\u63a8\u8350\u6a21\u578b\uff1a<\/strong>text2vec + Cohere\u91cd\u6392\u5e8f<br \/>\n<strong>\u4f18\u52bf\uff1a<\/strong>\u77ed\u6587\u672c\u8bed\u4e49\u6355\u6349\u7cbe\u51c6\uff0c\u5904\u7406\"yyds\"\u7b49\u7f51\u7edc\u7528\u8bed\u66f4\u7075\u6d3b<\/p>\n<h3>\u56db\u3001\u9009\u578b\u51b3\u7b56\u6811<\/h3>\n<p>\u200b1\u3001\u4e2d\u6587\u4e3a\u4e3b\u9009\uff1aBGE\u7cfb\u5217 &gt; M3E &gt; \u963f\u91cc\u4e91\u6a21\u578b<br \/>\n2\u3001\u200b\u591a\u8bed\u8a00\u9700\u6c42\u9009\uff1aBGE-M3 &gt; multilingual-e5<br \/>\n3\u3001\u200b\u9884\u7b97\u6709\u9650\u9009\uff1anomic-embed-text\uff08\u672c\u5730\uff09 \/ HuggingFace\u5f00\u6e90\u6a21\u578b<br \/>\n4\u3001\u200b\u8ffd\u6c42\u6781\u81f4\u6027\u80fd\u9009\uff1aOpenAI\u4ed8\u8d39API + \u6df7\u5408\u68c0\u7d22\u7b56\u7565<\/p>\n<h3>\u4e94\u3001\u6027\u80fd\u4f18\u5316\u6280\u5de7<\/h3>\n<p>1\u3001\u200b\u5411\u91cf\u7ef4\u5ea6\u538b\u7f29\uff1a\u4f7f\u7528PCA\u5c063072\u7ef4\u5411\u91cf\u964d\u81f3256\u7ef4\uff0c\u5b58\u50a8\u7a7a\u95f4\u51cf\u5c1192%<br \/>\n2\u3001\u6df7\u5408\u68c0\u7d22\u7b56\u7565\uff1aBGE-M3\u7684\u7a20\u5bc6\u68c0\u7d22 + BM25\u7b97\u6cd5\u7ed3\u5408\uff0cRecall@10\u63d0\u534718%<br \/>\n\u200b3\u3001\u91cd\u6392\u5e8f\u4f18\u5316\uff1a\u5728\u521d\u7b5b\u7ed3\u679c\u4e0a\u53e0\u52a0bge-reranker-v2.5\uff0cNDCG\u63d0\u534723%<\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u6700\u65b0\u8d8b\u52bf\uff1a2025\u5e74\u5d4c\u5165\u6a21\u578b\u5448\u73b0\u4e09\u5927\u53d1\u5c55\u65b9\u5411\uff1a<\/strong><\/p>\n<p>\u2460 \u591a\u6a21\u6001\u878d\u5408\uff08\u5982BGE-VL\u652f\u6301\u56fe\u6587\u4e92\u641c\uff09<br \/>\n\u2461 \u8d85\u957f\u4e0a\u4e0b\u6587\u5904\u7406\uff08\u7a81\u783432K token\u9650\u5236\uff09<br \/>\n\u2462 \u8f7b\u91cf\u5316\u63a8\u7406\uff081B\u53c2\u6570\u6a21\u578b\u8fbe\u52307B\u6a21\u578b90%\u6027\u80fd\uff09<\/p>\n<p>&nbsp;<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4ee5\u4e0b\u662f2025\u5e74\u4e3b\u6d41\u5d4c\u5165\u6a21\u578b\uff08Embedding Model\uff09\u7684\u89e3\u6790\uff0c\u7ed3\u5408\u6027\u80fd\u699c\u5355\u3001\u5e94\u7528\u573a\u666f\u548c\u6587\u6863\u9002\u914d\u6027\uff0c\u4ece\u5b9e [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[45],"tags":[],"class_list":["post-1019","post","type-post","status-publish","format-standard","hentry","category-ai"],"views":7681,"_links":{"self":[{"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=\/wp\/v2\/posts\/1019","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1019"}],"version-history":[{"count":3,"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=\/wp\/v2\/posts\/1019\/revisions"}],"predecessor-version":[{"id":1023,"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=\/wp\/v2\/posts\/1019\/revisions\/1023"}],"wp:attachment":[{"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1019"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1019"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tyhlw.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1019"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}