{"id":15437,"date":"2024-12-13T19:28:47","date_gmt":"2024-12-13T11:28:47","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=15437"},"modified":"2024-12-13T19:28:47","modified_gmt":"2024-12-13T11:28:47","slug":"juanqilailezhangban","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/juanqilailezhangban\/","title":{"rendered":"\u5377\u8d77\u6765\u4e86\uff01\u957f\u6587\u672c\u5411\u91cf\u6a21\u578b\u5206\u5757\u7b56\u7565\u5927\u6bd4\u62fc"},"content":{"rendered":"

<\/p>\n

\u957f\u6587\u672c\u5411\u91cf\u6a21\u578b\u80fd\u591f\u5c06\u5341\u9875\u957f\u7684\u6587\u672c\u7f16\u7801\u4e3a\u5355\u4e2a\u5411\u91cf\uff0c\u542c\u8d77\u6765\u5f88\u5f3a\u5927\uff0c\u4f46\u771f\u7684\u5b9e\u7528\u5417\uff1f<\/p>\n

\u5f88\u591a\u4eba\u89c9\u5f97… \u672a\u5fc5\u3002<\/p>\n

\u76f4\u63a5\u7528\u884c\u4e0d\u884c\uff1f\u8be5\u4e0d\u8be5\u5206\u5757\uff1f\u600e\u4e48\u5206\u624d\u6700\u9ad8\u6548\uff1f\u672c\u6587\u5c06\u5e26\u4f60\u6df1\u5165\u63a2\u8ba8\u957f\u6587\u672c\u5411\u91cf\u6a21\u578b\u7684\u4e0d\u540c\u5206\u5757\u7b56\u7565\uff0c\u5206\u6790\u5229\u5f0a\uff0c\u5e2e\u4f60\u907f\u5751\u3002<\/p>\n

 <\/p>\n

\u957f\u6587\u672c\u5411\u91cf\u5316\u7684\u95ee\u9898<\/h2>\n

\u9996\u5148\uff0c\u8ba9\u6211\u4eec\u770b\u770b\u5c06\u6574\u7bc7\u6587\u7ae0\u538b\u7f29\u6210\u5355\u4e00\u5411\u91cf\u4f1a\u5b58\u5728\u54ea\u4e9b\u95ee\u9898\u3002<\/p>\n

\u4ee5\u6784\u5efa\u6587\u6863\u641c\u7d22\u7cfb\u7edf\u4e3a\u4f8b\uff0c\u5355\u7bc7\u6587\u7ae0\u53ef\u80fd\u5305\u542b\u591a\u4e2a\u4e3b\u9898\u3002<\/strong> \u6bd4\u5982\u8fd9\u7bc7\u5173\u4e8e ICML 2024 \u53c2\u4f1a\u62a5\u544a\u7684\u535a\u5ba2\uff0c\u5c31\u5305\u542b\u4f1a\u8bae\u4ecb\u7ecd\u3001Jina AI \u7684\u5de5\u4f5c\u5c55\u793a (jina-clip-v1<\/code>) \u4ee5\u53ca\u5176\u4ed6\u7814\u7a76\u8bba\u6587\u603b\u7ed3\u3002\u5982\u679c\u5c06\u6574\u7bc7\u6587\u7ae0\u5411\u91cf\u5316\u6210\u5355\u4e00\u5411\u91cf\uff0c\u8be5\u5411\u91cf\u5c06\u6df7\u5408\u4e09\u4e2a\u4e0d\u540c\u4e3b\u9898\u7684\u4fe1\u606f\uff1a<\/p>\n

<\/p>\n

\u8fd9\u4f1a\u5bfc\u81f4\u4ee5\u4e0b\u95ee\u9898\uff1a<\/p>\n

1. \u8868\u793a\u7a00\u91ca(Representation Dilution)<\/strong><\/p>\n

\u8868\u793a\u7a00\u91ca\u4f1a\u524a\u5f31\u6587\u672c\u5411\u91cf\u7684\u7cbe\u5ea6\u3002\u867d\u7136\u535a\u5ba2\u6587\u7ae0\u5305\u542b\u4e86\u591a\u4e2a\u4e3b\u9898\uff0c\u4f46\u7528\u6237\u7684\u641c\u7d22\u67e5\u8be2\u5f80\u5f80\u53ea\u5173\u6ce8\u5176\u4e2d\u4e00\u4e2a<\/strong>\u3002\u7528\u5355\u4e2a\u5411\u91cf\u8868\u793a\u6574\u7bc7\u6587\u7ae0\uff0c\u76f8\u5f53\u4e8e\u5c06\u6240\u6709\u4e3b\u9898\u4fe1\u606f\u538b\u7f29\u5230\u5411\u91cf\u7a7a\u95f4\u4e2d\u7684\u4e00\u4e2a\u70b9\u3002\u968f\u7740\u66f4\u591a\u6587\u672c\u6dfb\u52a0\u5230\u6a21\u578b\u7684\u8f93\u5165\u4e2d\uff0c\u8fd9\u4e2a\u5411\u91cf\u4f1a\u9010\u6e10\u4ee3\u8868\u6587\u7ae0\u7684\u6574\u4f53\u4e3b\u9898\uff0c\u7a00\u91ca\u7279\u5b9a\u6bb5\u843d\u6216\u4e3b\u9898\u7684\u7ec6\u8282\u3002\u8fd9\u5c31\u50cf\u628a\u591a\u79cd\u989c\u6599\u6df7\u5408\u6210\u4e86\u4e00\u79cd\u989c\u8272\uff0c\u7528\u6237\u60f3\u627e\u5230\u67d0\u79cd\u7279\u5b9a\u989c\u8272\u65f6\uff0c\u5c31\u5f88\u96be\u4ece\u6df7\u5408\u8272\u4e2d\u8bc6\u522b\u51fa\u6765\u3002<\/p>\n

2. \u5bb9\u91cf\u6709\u9650<\/strong><\/p>\n

\u6a21\u578b\u751f\u6210\u7684\u5411\u91cf\u7ef4\u5ea6\u662f\u56fa\u5b9a\u7684\uff0c\u957f\u6587\u672c\u5305\u542b\u5927\u91cf\u4fe1\u606f\uff0c\u8f6c\u5316\u8fc7\u7a0b\u4e2d\u5fc5\u7136\u4f1a\u5bfc\u81f4\u4fe1\u606f\u635f\u5931\u3002\u5c31\u597d\u6bd4\u628a\u9ad8\u6e05\u5730\u56fe\u538b\u7f29\u6210\u90ae\u7968\uff0c\u5f88\u591a\u7ec6\u8282\u90fd\u770b\u4e0d\u6e05\u4e86\u3002<\/p>\n

3. \u4fe1\u606f\u4e22\u5931<\/strong><\/p>\n

\u5f88\u591a\u957f\u6587\u672c\u6a21\u578b\u6700\u591a\u53ea\u80fd\u5904\u7406 8192 token\u3002\u66f4\u597d\u7684\u6587\u672c\u5f97\u622a\u65ad\u4e00\u90e8\u5206\uff0c\u901a\u5e38\u662f\u622a\u65ad\u540e\u9762\u7684\u5185\u5bb9\uff0c\u8981\u662f\u5173\u952e\u4fe1\u606f\u5728\u6587\u6863\u672b\u5c3e\uff0c\u5c31\u53ef\u80fd\u5bfc\u81f4\u68c0\u7d22\u5931\u8d25\u4e86\u3002<\/p>\n

4. \u5206\u5757\u9700\u6c42<\/strong><\/p>\n

\u6709\u4e9b\u5e94\u7528\u53ea\u9700\u8981\u5bf9\u6587\u672c\u7684\u7279\u5b9a\u7247\u6bb5\u505a\u5411\u91cf\u5316\u5904\u7406\uff0c\u6bd4\u5982\u95ee\u7b54\u7cfb\u7edf\uff0c\u53ea\u9700\u63d0\u53d6\u5305\u542b\u7b54\u6848\u7684\u6bb5\u843d\u8fdb\u884c\u5411\u91cf\u5316\u3002\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u8fd8\u662f\u9700\u8981\u5bf9\u6587\u672c\u8fdb\u884c\u5206\u5757\u5904\u7406\u3002<\/p>\n

 <\/p>\n

3 \u79cd\u957f\u6587\u672c\u5904\u7406\u7b56\u7565<\/h2>\n

\u5728\u5b9e\u9a8c\u5f00\u59cb\u4e4b\u524d\uff0c\u4e3a\u4e86\u907f\u514d\u6982\u5ff5\u6df7\u6dc6\uff0c\u6211\u4eec\u9996\u5148\u5b9a\u4e49\u4e09\u4e2a\u5206\u5757\u7684\u7b56\u7565\uff1a<\/p>\n

1. \u4e0d\u5206\u5757(No Chunking)\uff1a<\/strong>\u5c06\u6574\u4e2a\u6587\u672c\u76f4\u63a5\u7f16\u7801\u6210\u5355\u4e00\u5411\u91cf\u3002<\/p>\n

2. \u6734\u7d20\u5206\u5757(Naive Chunking)\uff1a<\/strong>\u5148\u5c06\u6587\u672c\u5206\u6210\u591a\u4e2a\u6587\u672c\u5757\uff0c\u5e76\u5206\u522b\u8fdb\u884c\u5411\u91cf\u5316\u3002\u5e38\u7528\u7684\u65b9\u6cd5\u6709\u56fa\u5b9a\u5927\u5c0f\u5206\u5757\uff0c\u5c06\u6587\u672c\u5206\u5272\u6210\u56fa\u5b9a token<\/a> \u6570\u91cf\u7684\u5757\uff1b\u57fa\u4e8e\u53e5\u5b50\u7684\u5206\u5757: \u4ee5\u53e5\u5b50\u4e3a\u5355\u4f4d\u8fdb\u884c\u5206\u5757\uff1b\u57fa\u4e8e\u8bed\u4e49\u7684\u5206\u5757: \u6839\u636e\u8bed\u4e49\u4fe1\u606f\u8fdb\u884c\u5206\u5757\u3002\u5728\u672c\u6b21\u5b9e\u9a8c\uff0c\u4f7f\u7528\u7684\u662f\u56fa\u5b9a\u5927\u5c0f\u5206\u5757\u3002<\/p>\n

3. \u8fdf\u5206(Late Chunking)\uff1a<\/strong>\u662f\u4e00\u79cd\u5148\u901a\u8bfb\u5168\u6587\u518d\u5206\u5757\u7684\u65b0\u65b9\u6cd5\uff0c\u5305\u542b\u4e24\u4e2a\u4e3b\u8981\u6b65\u9aa4\uff1a<\/p>\n

    \n
  1. \u7f16\u7801\u5168\u6587<\/strong>\uff1a\u5148\u7f16\u7801\u6574\u4e2a\u6587\u6863\uff0c\u5f97\u5230\u6bcf\u4e2a token \u7684\u5411\u91cf\u8868\u793a\uff0c\u4fdd\u7559\u5b8c\u6574\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/li>\n
  2. \u5206\u5757\u6c60\u5316<\/strong>\uff1a\u6839\u636e\u5206\u5757\u8fb9\u754c\uff0c\u5bf9\u540c\u4e00\u4e2a\u6587\u672c\u5757\u7684 token \u5411\u91cf\u8fdb\u884c\u5e73\u5747\u6c60\u5316\uff0c\u751f\u6210\u6bcf\u4e2a\u6587\u672c\u5757\u7684\u5411\u91cf\u3002\u7531\u4e8e\u6bcf\u4e2a token \u7684\u5411\u91cf\u662f\u5728\u5168\u6587\u7684\u8bed\u5883\u4e0b\u751f\u6210\u7684\uff0c\u56e0\u6b64\u8fdf\u5206\u53ef\u4ee5\u4fdd\u7559\u5757\u4e4b\u95f4\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/li>\n<\/ol>\n
    \"\"

    \u8fdf\u5206\u4e0e\u6734\u7d20\u5206\u5757\u7684\u5bf9\u6bd4<\/p><\/div>\n

     <\/p>\n

    \u5bf9\u4e8e\u8d85\u8fc7\u6a21\u578b\u6700\u5927\u8f93\u5165\u957f\u5ea6 (\u4f8b\u5982 8192 tokens<\/a>) \u7684\u8d85\u957f\u6587\u672c\uff0c\u6211\u4eec\u91c7\u7528 \u957f\u6587\u672c\u8fdf\u5206(Long Late Chunking)<\/strong>\uff0c\u5728\u8fdf\u5206\u7684\u57fa\u7840\u4e0a\u589e\u52a0\u4e86\u9884\u5206\u5272\u6b65\u9aa4\uff0c\u5148\u5c06\u6587\u6863\u5206\u5272\u6210\u591a\u4e2a\u91cd\u53e0\u7684\u5b8f\u5757\uff0c\u6bcf\u4e2a\u5b8f\u5757\u7684\u957f\u5ea6\u90fd\u5728\u6a21\u578b\u53ef\u5904\u7406\u8303\u56f4\u5185\u3002\u7136\u540e\uff0c\u5728\u6bcf\u4e2a\u5b8f\u5757\u5185\u90e8\u5e94\u7528\u6807\u51c6\u7684\u8fdf\u5206\u7b56\u7565\uff08\u7f16\u7801\u548c\u6c60\u5316\uff09\u3002\u5b8f\u5757\u4e4b\u95f4\u7684\u91cd\u53e0\u90e8\u5206\u7528\u4e8e\u786e\u4fdd\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u8fde\u7eed\u6027\u3002<\/p>\n

    \u8fdf\u5206\u5177\u4f53\u5b9e\u73b0\u4ee3\u7801\uff1ahttps:\/\/github.com\/jina-ai\/late-chunking\u5728 Notebook \u4f53\u9a8c\uff1ahttps:\/\/colab.research.google.com\/drive\/1iz3ACFs5aLV2O_uZEjiR1aHGlXqj0HY7?usp=sharing<\/p>\n

    \u90a3\u4e48\uff0c\u54ea\u79cd\u65b9\u6cd5\u6700\u597d\u5462\uff1f<\/strong><\/p>\n

    \u4e3a\u4e86\u5bf9\u6bd4\uff0c\u6211\u4eec\u5728 5 \u4e2a\u6570\u636e\u96c6\u4e0a\uff0c\u4f7f\u7528 jina-embeddings-v3<\/code> \u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u6240\u6709\u957f\u6587\u672c\u90fd\u88ab\u622a\u65ad\u81f3\u6a21\u578b\u6700\u5927\u8f93\u5165\u957f\u5ea6 (8192 Tokens)\uff0c\u5e76\u6309\u6bcf 64 \u4e2a Token \u4e3a\u4e00\u4e2a\u6587\u672c\u5757\u8fdb\u884c\u5206\u5272\u3002<\/p>\n

    \"\"

    5 \u4e2a\u6d4b\u8bd5\u6570\u636e\u96c6\u4e5f\u5bf9\u5e94 5 \u79cd\u4e0d\u540c\u7684\u68c0\u7d22\u4efb\u52a1<\/p><\/div>\n

     <\/p>\n

    \u4e0b\u56fe\u5c55\u793a\u4e86 3 \u79cd\u65b9\u6cd5\u5728\u4e0d\u540c\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u5dee\u5f02\uff0c\u5e76\u6ca1\u6709\u4e00\u79cd\u65b9\u6cd5\u5728\u6240\u6709\u60c5\u51b5\u4e0b\u90fd\u6700\u597d\uff0c\u9009\u62e9\u54ea\u79cd\u65b9\u6cd5\u53d6\u51b3\u4e8e\u5177\u4f53\u7684\u4efb\u52a1\u3002<\/p>\n

    \"\"

    \u4e0d\u5206\u5757 vs \u6734\u7d20\u5206\u5757 vs \u8fdf\u5206<\/p><\/div>\n

     <\/p>\n

    \ud83d\udc69\u200d\ud83c\udfeb\u00a0\u627e\u5177\u4f53\u4e8b\u5b9e\uff0c\u6734\u7d20\u5206\u5757\u597d\u3002\u3000<\/strong><\/p>\n

    \u5982\u679c\u9700\u8981\u4ece\u6587\u672c\u4e2d\u63d0\u53d6\u5177\u4f53\u7684\u3001\u5c40\u90e8\u7684\u4e8b\u5b9e\u4fe1\u606f\uff08\u4f8b\u5982\uff0c\u201c\u8c01\u5077\u4e86\u4e1c\u897f\uff1f\u201d\uff09\uff0c\u50cf QMSum\u3001NarrativeQA \u548c 2WikiMultiHopQA \u8fd9\u51e0\u4e2a\u6570\u636e\u96c6\uff0c\u6734\u7d20\u5206\u5757\u6bd4\u628a\u6574\u4e2a\u6587\u6863\u5411\u91cf\u5316\u8868\u73b0\u8981\u597d\u3002\u56e0\u4e3a\u7b54\u6848\u901a\u5e38\u4f4d\u4e8e\u6587\u672c\u7684\u67d0\u4e2a\u7279\u5b9a\u90e8\u5206\uff0c\u6734\u7d20\u5206\u5757\u80fd\u591f\u66f4\u7cbe\u51c6\u5730\u5b9a\u4f4d\u5230\u5305\u542b\u7b54\u6848\u7684\u6587\u672c\u5757\uff0c\u800c\u4e0d\u4f1a\u88ab\u5176\u4ed6\u65e0\u5173\u4fe1\u606f\u5e72\u6270\u3002<\/p>\n

    \u4f46\u662f\u6734\u7d20\u5206\u5757\u4e5f\u4f1a\u56e0\u4e3a\u5207\u65ad\u4e86\u4e0a\u4e0b\u6587\uff0c\u53ef\u80fd\u4e22\u5931\u5168\u5c40\u4fe1\u606f\uff0c\u65e0\u6cd5\u6b63\u786e\u89e3\u6790\u6587\u672c\u4e2d\u7684\u6307\u4ee3\u5173\u7cfb\u548c\u5f15\u7528\u3002<\/p>\n

    \ud83d\udc69\u200d\ud83c\udfeb \u6587\u7ae0\u4e3b\u9898\u8fde\u8d2f\uff0c\u8fdf\u5206\u66f4\u597d\u3002\u3000<\/strong><\/p>\n

    \u5982\u679c\u6587\u7ae0\u4e3b\u9898\u5f88\u6e05\u6670\uff0c\u4e14\u7bc7\u7ae0\u7ed3\u6784\u8fde\u8d2f\uff0c\u8fdf\u5206\u65b9\u6cd5\u4f1a\u66f4\u6709\u6548\u3002\u56e0\u4e3a\u8fdf\u5206\u8003\u8651\u4e86\u4e0a\u4e0b\u6587\uff0c\u80fd\u66f4\u597d\u5730\u7406\u89e3\u6bcf\u4e2a\u90e8\u5206\u7684\u610f\u4e49\u548c\u5173\u8054\uff0c\u5305\u62ec\u957f\u6587\u672c\u91cc\u9762\u7684\u6307\u4ee3\u5173\u7cfb\u3002<\/p>\n

    \u4f46\u5982\u679c\u6587\u7ae0\u91cc\u5939\u6742\u7740\u5f88\u591a\u548c\u4e3b\u9898\u65e0\u5173\u7684\u5185\u5bb9\uff0c\u8fdf\u5206\u53cd\u800c\u4f1a\u56e0\u4e3a\u8003\u8651\u4e86\u8fd9\u4e9b\u201c\u566a\u97f3\u201d\u5bfc\u81f4\u6027\u80fd\u56de\u9000\uff0c\u51c6\u786e\u7387\u4e0b\u964d\u3002\u50cf NarrativeQA\u30012WikiMultiHopQA\uff0c\u8fdf\u5206\u8868\u73b0\u4e0d\u5982\u6734\u7d20\u5206\u5757\uff0c\u5c31\u662f\u56e0\u4e3a\u8fd9\u4e9b\u6587\u7ae0\u91cc\u65e0\u5173\u4fe1\u606f\u592a\u591a\u4e86\u3002<\/p>\n

     <\/p>\n

    \u5206\u5757\u5927\u5c0f\u662f\u5426\u6709\u5f71\u54cd\uff1f<\/h3>\n

    \u4e0b\u56fe\u5c55\u793a\u4e86\u4e0d\u540c\u5206\u5757\u5927\u5c0f\u4e0b\uff0c\u6734\u7d20\u5206\u5757\u548c\u8fdf\u5206\u65b9\u6cd5\u5728\u4e0d\u540c\u6570\u636e\u96c6\u4e0a\u7684\u8868\u73b0\uff1a<\/p>\n

    \"\"

    \u4e0d\u540c\u5206\u5757\u5927\u5c0f\u4e0b\u6734\u7d20\u5206\u5757\u548c\u8fdf\u5206\u7684\u6027\u80fd\u5bf9\u6bd4<\/p><\/div>\n

     <\/p>\n

    \u4ece\u56fe\u4e2d\u6211\u4eec\u53ef\u4ee5\u770b\u51fa\uff0c\u6700\u4f73\u7684\u5206\u5757\u5927\u5c0f\u5176\u5b9e\u8981\u770b\u5177\u4f53\u7684\u6570\u636e\u96c6\u662f\u4ec0\u4e48\u6837\u7684\u3002<\/p>\n

    \u5bf9\u4e8e\u8fdf\u5206\u65b9\u6cd5\uff0c\u8f83\u5c0f\u7684\u5206\u5757\u80fd\u66f4\u597d\u5730\u6293\u4f4f\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u6240\u4ee5\u6548\u679c\u66f4\u597d\u3002\u5c24\u5176\u662f\uff0c\u5982\u679c\u6570\u636e\u96c6\u91cc\u6709\u5f88\u591a\u548c\u4e3b\u9898\u65e0\u5173\u7684\u5185\u5bb9\uff08\u5c31\u50cf NarrativeQA \u6570\u636e\u96c6\u90a3\u6837\uff09\uff0c\u8fc7\u5927\u7684\u4e0a\u4e0b\u6587\u53cd\u800c\u4f1a\u5f15\u5165\u566a\u97f3\uff0c\u635f\u5bb3\u6027\u80fd\u3002<\/p>\n

    \u5bf9\u4e8e\u6734\u7d20\u5206\u5757\uff0c\u5927\u4e00\u4e9b\u7684\u5206\u5757\u6709\u65f6\u6548\u679c\u4f1a\u66f4\u597d\uff0c\u56e0\u4e3a\u5305\u542b\u7684\u4fe1\u606f\u66f4\u5168\u9762\uff0c\u635f\u5931\u4e5f\u66f4\u5c0f\u3002\u4f46\u6709\u65f6\u5019\uff0c\u5206\u5757\u592a\u5927\uff0c\u4fe1\u606f\u53c8\u592a\u6742\u4e71\uff0c\u53cd\u800c\u964d\u4f4e\u4e86\u68c0\u7d22\u7684\u51c6\u786e\u6027\u3002\u6240\u4ee5\uff0c\u6700\u4f73\u7684\u5206\u5757\u5927\u5c0f\uff0c\u9700\u8981\u6839\u636e\u5177\u4f53\u7684\u6570\u636e\u96c6\u548c\u4efb\u52a1\u6765\u8c03\u6574\uff0c\u6ca1\u6709\u4e00\u4e2a\u4e00\u5200\u5207\u7684\u7b54\u6848\u3002<\/p>\n

    \u7406\u89e3\u4e86\u4e0d\u540c\u5206\u5757\u7b56\u7565\u7684\u4f18\u7f3a\u70b9\u540e\uff0c\u6211\u4eec\u8be5\u5982\u4f55\u9009\u62e9\u5408\u9002\u7684\u7b56\u7565\u5462\uff1f<\/p>\n

    1. \u5168\u6587\u5411\u91cf\u5316\uff08\u4e0d\u5206\u5757\uff09\u9002\u5408\u54ea\u4e9b\u60c5\u51b5\uff1f<\/strong><\/p>\n

      \n
    • \u4e3b\u9898\u5355\u4e00\uff0c\u5173\u952e\u4fe1\u606f\u96c6\u4e2d\u5728\u5f00\u5934\uff1a<\/strong>\u6bd4\u5982\u7ed3\u6784\u5316\u7684\u65b0\u95fb\u62a5\u9053\uff0c\u5173\u952e\u4fe1\u606f\u5f80\u5f80\u5728\u6807\u9898\u548c\u5f00\u5934\u6bb5\u843d\u4e2d\u3002\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u76f4\u63a5\u4f7f\u7528\u5168\u6587\u5411\u91cf\u5316\u901a\u5e38\u80fd\u591f\u83b7\u5f97\u4e0d\u9519\u7684\u6548\u679c\uff0c\u56e0\u4e3a\u6a21\u578b\u80fd\u591f\u6355\u6349\u5230\u4e3b\u8981\u4fe1\u606f\u3002<\/li>\n
    • \u901a\u5e38\u6765\u8bf4\uff0c\u628a\u5c3d\u53ef\u80fd\u591a\u7684\u6587\u672c\u5185\u5bb9\u653e\u8fdb\u6a21\u578b\u4e0d\u4f1a\u5f71\u54cd\u68c0\u7d22\u7ed3\u679c\u3002\u4f46\u662f\uff0c\u957f\u6587\u672c\u6a21\u578b\u5f80\u5f80\u66f4\u5173\u6ce8\u5f00\u5934\u90e8\u5206\uff08\u6807\u9898\u3001\u5f15\u8a00\u7b49\uff09\uff0c\u4e2d\u95f4\u548c\u7ed3\u5c3e\u90e8\u5206\u7684\u4fe1\u606f\u53ef\u80fd\u4f1a\u88ab\u5ffd\u7565\u3002\u6240\u4ee5\uff0c\u5982\u679c\u5173\u952e\u4fe1\u606f\u5728\u6587\u7ae0\u4e2d\u95f4\u6216\u7ed3\u5c3e\uff0c\u8fd9\u79cd\u65b9\u6cd5\u7684\u6548\u679c\u5c31\u4f1a\u5927\u6253\u6298\u6263\u3002<\/li>\n
    • \u8be6\u7ec6\u5b9e\u9a8c\u7ed3\u679c\u89c1\uff1ahttps:\/\/jina.ai\/news\/still-need-chunking-when-long-context-models-can-do-it-all<\/em><\/li>\n<\/ul>\n

      2. \u6734\u7d20\u5206\u5757\uff08Naive Chunking\uff09\u9002\u5408\u54ea\u4e9b\u60c5\u51b5\uff1f<\/strong><\/p>\n

        \n
      • \u4e3b\u9898\u591a\u6837\uff0c\u9700\u8981\u68c0\u7d22\u7279\u5b9a\u4fe1\u606f<\/strong>\uff1a\u5982\u679c\u4f60\u7684\u6587\u672c\u5305\u542b\u591a\u4e2a\u4e3b\u9898\uff0c\u6216\u8005\u7528\u6237\u67e5\u8be2\u7684\u76ee\u6807\u662f\u6587\u672c\u4e2d\u7684\u67d0\u4e2a\u5177\u4f53\u4e8b\u5b9e\uff0c\u90a3\u4e48\u6734\u7d20\u5206\u5757\u662f\u4e00\u4e2a\u4e0d\u9519\u7684\u9009\u62e9\u3002\u5b83\u80fd\u591f\u6709\u6548\u907f\u514d\u4fe1\u606f\u7a00\u91ca\uff0c\u63d0\u9ad8\u68c0\u7d22\u7279\u5b9a\u4fe1\u606f\u7684\u51c6\u786e\u7387\u3002<\/li>\n
      • \u9700\u8981\u5c55\u793a\u5c40\u90e8\u6587\u672c\u7247\u6bb5<\/strong>\uff1a\u7c7b\u4f3c\u4e8e\u641c\u7d22\u5f15\u64ce\uff0c\u9700\u8981\u5728\u7ed3\u679c\u4e2d\u5c55\u793a\u4e0e\u67e5\u8be2\u76f8\u5173\u7684\u6587\u672c\u7247\u6bb5\uff0c\u5c31\u5fc5\u987b\u91c7\u7528\u5206\u5757\u7b56\u7565\u3002<\/li>\n
      • \u53e6\u5916\uff0c\u5206\u5757\u4e5f\u4f1a\u56e0\u4e3a\u8981\u5411\u91cf\u5316\u66f4\u591a\u6587\u672c\u5757\u800c\u5f71\u54cd\u5b58\u50a8\u7a7a\u95f4\u548c\u5904\u7406\u65f6\u95f4\u3002<\/li>\n<\/ul>\n

        3. \u8fdf\u5206\uff08Late Chunking\uff09\u9002\u5408\u54ea\u4e9b\u60c5\u51b5\uff1f<\/strong><\/p>\n

          \n
        • \u4e3b\u9898\u8fde\u8d2f\uff0c\u9700\u8981\u4e0a\u4e0b\u6587\u4fe1\u606f<\/strong>\uff1a\u5bf9\u4e8e\u4e3b\u9898\u8fde\u8d2f\u7684\u957f\u6587\u672c\uff0c\u4f8b\u5982\u8bba\u6587\u3001\u957f\u7bc7\u62a5\u9053\u7b49\uff0c\u8fdf\u5206\u65b9\u6cd5\u80fd\u591f\u6709\u6548\u4fdd\u7559\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u4ece\u800c\u66f4\u597d\u5730\u7406\u89e3\u6587\u672c\u7684\u6574\u4f53\u8bed\u4e49\u3002\u5b83\u5c24\u5176\u9002\u7528\u4e8e\u9700\u8981\u7406\u89e3\u6587\u672c\u4e2d\u4e0d\u540c\u90e8\u5206\u4e4b\u95f4\u5173\u7cfb\u7684\u4efb\u52a1\uff0c\u4f8b\u5982\u9605\u8bfb\u7406\u89e3\u548c\u957f\u6587\u672c\u8bed\u4e49\u5339\u914d\u3002<\/li>\n
        • \u9700\u8981\u5e73\u8861\u5c40\u90e8\u7ec6\u8282\u548c\u5168\u5c40\u8bed\u4e49<\/strong>\uff1a\u8fdf\u5206\u65b9\u6cd5\u5728\u8f83\u5c0f\u7684\u5206\u5757\u5927\u5c0f\u4e0b\uff0c\u80fd\u591f\u6709\u6548\u5e73\u8861\u5c40\u90e8\u7ec6\u8282\u548c\u5168\u5c40\u8bed\u4e49\uff0c\u5728\u8bb8\u591a\u60c5\u51b5\u4e0b\u80fd\u591f\u53d6\u5f97\u6bd4\u5176\u4ed6\u4e24\u79cd\u65b9\u6cd5\u66f4\u597d\u7684\u6548\u679c\u3002\u4f46\u8981\u6ce8\u610f\uff0c\u5982\u679c\u6587\u7ae0\u91cc\u6709\u5f88\u591a\u65e0\u5173\u7d27\u8981\u7684\u5185\u5bb9\uff0c\u8fdf\u5206\u53cd\u800c\u4f1a\u56e0\u4e3a\u8003\u8651\u4e86\u8fd9\u4e9b\u65e0\u5173\u4fe1\u606f\u800c\u5f71\u54cd\u6548\u679c\u3002<\/li>\n<\/ul>\n

           <\/p>\n

          \u7ed3\u8bba<\/h2>\n

          \u957f\u6587\u672c\u5411\u91cf\u5316\u7b56\u7565\u7684\u9009\u62e9\u662f\u4e00\u4e2a\u590d\u6742\u7684\u95ee\u9898\uff0c\u6ca1\u6709\u4e00\u5200\u5207\u7684\u6700\u4f73\u65b9\u6848\uff0c\u9700\u8981\u8003\u8651\u6570\u636e\u7279\u70b9\u548c\u68c0\u7d22\u76ee\u6807\uff0c\u5305\u62ec\u524d\u9762\u63d0\u5230\u7684\u6587\u672c\u957f\u5ea6\u3001\u4e3b\u9898\u6570\u91cf\u3001\u5173\u952e\u4fe1\u606f\u4f4d\u7f6e\u3002<\/p>\n

          \u672c\u6587\u5e0c\u671b\u63d0\u4f9b\u4e00\u4e2a\u5173\u4e8e\u4e0d\u540c\u5206\u5757\u7b56\u7565\u7684\u6bd4\u8f83\u5206\u6790\u6846\u67b6\uff0c\u5e76\u901a\u8fc7\u5b9e\u9a8c\u7ed3\u679c\u63d0\u4f9b\u4e86\u4e00\u4e9b\u53c2\u8003\u3002\u5728\u5b9e\u9645\u5e94\u7528\u65f6\uff0c\u5927\u5bb6\u53ef\u4ee5\u591a\u5bf9\u6bd4\u5b9e\u9a8c\uff0c\u9009\u62e9\u6700\u9002\u5408\u81ea\u5df1\u573a\u666f\u7684\u7b56\u7565\u3002<\/p>\n

          \u5982\u679c\u4f60\u5bf9\u957f\u6587\u672c\u5411\u91cf\u5316\u611f\u5174\u8da3\uff0cjina-embeddings-v3<\/code>\u63d0\u4f9b\u4e86\u5148\u8fdb\u7684\u957f\u6587\u672c\u5904\u7406\u80fd\u529b\u3001\u591a\u8bed\u8a00\u652f\u6301\u4ee5\u53ca\u8fdf\u5206\u529f\u80fd\uff0c\u503c\u5f97\u5c1d\u8bd5\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

          \u957f\u6587\u672c\u5411\u91cf\u6a21\u578b\u80fd\u591f\u5c06\u5341\u9875\u957f\u7684\u6587\u672c\u7f16\u7801\u4e3a\u5355\u4e2a\u5411\u91cf\uff0c\u542c\u8d77\u6765\u5f88\u5f3a\u5927\uff0c\u4f46\u771f\u7684\u5b9e\u7528\u5417\uff1f \u5f88\u591a\u4eba\u89c9\u5f97… \u672a\u5fc5\u3002 \u76f4\u63a5\u7528\u884c\u4e0d\u884c\uff1f\u8be5\u4e0d\u8be5\u5206\u5757\uff1f\u600e\u4e48\u5206\u624d\u6700\u9ad8\u6548\uff1f\u672c\u6587\u5c06\u5e26\u4f60\u6df1\u5165\u63a2\u8ba8\u957f\u6587\u672c\u5411\u91cf\u6a21\u578b\u7684\u4e0d\u540c\u5206\u5757\u7b56\u7565\uff0c\u5206\u6790\u5229\u5f0a\uff0c\u5e2e\u4f60\u907f\u5751\u3002   \u957f…<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-15437","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/15437","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=15437"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/15437\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=15437"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=15437"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=15437"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}