{"id":16331,"date":"2024-12-23T22:36:01","date_gmt":"2024-12-23T14:36:01","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=16331"},"modified":"2024-12-23T22:37:24","modified_gmt":"2024-12-23T14:37:24","slug":"late-chunkingmilvusrag","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/pt\/late-chunkingmilvusrag\/","title":{"rendered":"Late Chunking\u00d7Milvus\uff1a\u5982\u4f55\u63d0\u9ad8RAG\u51c6\u786e\u7387"},"content":{"rendered":"<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/baabf8e38340321.png\" \/><\/p>\n<p><strong>01.<\/strong><strong>\u80cc\u666f<\/strong><\/p>\n<p>\u5728RAG\u5e94\u7528\u5f00\u53d1\u4e2d\uff0c\u7b2c\u4e00\u6b65\u5c31\u662f\u5bf9\u4e8e\u6587\u6863\u8fdb\u884cchunking\uff08\u5206\u5757\uff09\uff0c\u9ad8\u6548\u7684\u6587\u6863\u5206\u5757\uff0c\u53ef\u4ee5\u6709\u6548\u7684\u63d0\u9ad8\u540e\u7eed\u7684\u53ec\u56de\u5185\u5bb9\u7684\u51c6\u786e\u6027\u3002\u800c\u5bf9\u4e8e\u5982\u4f55\u9ad8\u6548\u7684\u5206\u5757\u662f\u4e2a\u8ba8\u8bba\u7684\u70ed\u70b9\uff0c\u6709\u8bf8\u5982\u56fa\u5b9a\u5927\u5c0f\u5206\u5757\uff0c\u968f\u673a\u5927\u5c0f\u5206\u5757\uff0c\u6ed1\u52a8\u7a97\u53e3\u91cd\u65b0\u91c7\u6837\uff0c\u9012\u5f52\u5206\u5757\uff0c\u57fa\u4e8e\u5185\u5bb9\u8bed\u4e49\u5206\u5757\u7b49\u65b9\u6cd5\u3002\u800cJina AI\u63d0\u51fa\u7684Late Chunking\u4ece\u53e6\u5916\u4e00\u4e2a\u89d2\u5ea6\u6765\u5904\u7406\u5206\u5757\u95ee\u9898\uff0c\u8ba9\u6211\u4eec\u6765\u5177\u4f53\u770b\u770b\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>02.<\/strong><strong>Late Chunking\u662f\u4ec0\u4e48<\/strong><\/p>\n<p>\u4f20\u7edf\u7684\u5206\u5757\u5728\u5904\u7406\u957f\u6587\u6863\u65f6\u53ef\u80fd\u4f1a\u4e22\u5931\u6587\u6863\u4e2d\u957f\u8ddd\u79bb\u7684\u4e0a\u4e0b\u6587\u4f9d\u8d56\u5173\u7cfb\uff0c\u8fd9\u5bf9\u4e8e\u4fe1\u606f\u68c0\u7d22\u548c\u7406\u89e3\u662f\u4e00\u5927\u9690\u60a3\u3002\u5373\u5f53\u5173\u952e\u4fe1\u606f\u6563\u843d\u5728\u591a\u4e2a\u6587\u672c\u5757\u4e2d\uff0c\u8131\u79bb\u4e0a\u4e0b\u6587\u7684\u6587\u672c\u5206\u5757\u7247\u6bb5\u5f88\u53ef\u80fd\u5931\u53bb\u5176\u539f\u6709\u7684\u610f\u4e49\uff0c\u5bfc\u81f4\u540e\u7eed\u7684\u53ec\u56de\u6548\u679c\u6bd4\u8f83\u5dee\u3002<\/p>\n<p>\u4ee5Milvus 2.4.13 release note\u4e3a\u4f8b\uff0c\u5047\u5982\u5206\u4e3a\u5982\u4e0b\u4e24\u4e2a\u6587\u6863\u5757\uff0c\u5982\u679c\u6211\u4eec\u8981\u67e5\u8be2<code>Milvus 2.4.13\u6709\u54ea\u4e9b\u65b0\u529f\u80fd\uff1f<\/code>\uff0c\u76f4\u63a5\u76f8\u5173\u5185\u5bb9\u5728\u5206\u57572\u91cc\uff0c\u800cMilvus\u7248\u672c\u4fe1\u606f\u5728\u5206\u57571\u91cc\uff0c\u6b64\u65f6\uff0cEmbedding \u6a21\u578b\u5f88\u96be\u5c06\u8fd9\u4e9b\u6307\u4ee3\u6b63\u786e\u94fe\u63a5\u5230\u5b9e\u4f53\uff0c\u4ece\u800c\u4ea7\u751f\u8d28\u91cf\u4e0d\u9ad8\u7684Embedding\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36e43c6bce174d7.png\" \/><\/p>\n<p>\u7531\u4e8e\u529f\u80fd\u63cf\u8ff0\u4e0e\u7248\u672c\u4fe1\u606f\u4e0d\u5728\u540c\u4e00\u4e2a\u5206\u5757\u91cc\uff0c\u4e14\u7f3a\u4e4f\u66f4\u5927\u7684\u4e0a\u4e0b\u6587\u6587\u6863\uff0cLLM \u96be\u4ee5\u89e3\u51b3\u8fd9\u6837\u7684\u5173\u8054\u95ee\u9898\u3002\u5c3d\u7ba1\u6709\u4e00\u4e9b\u542f\u53d1\u5f0f\u7b97\u6cd5\u8bd5\u56fe\u7f13\u89e3\u8fd9\u4e00\u95ee\u9898\uff0c\u5982\u6ed1\u52a8\u7a97\u53e3\u91cd\u65b0\u91c7\u6837\u3001\u91cd\u53e0\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u957f\u5ea6\u4ee5\u53ca\u591a\u6b21\u6587\u6863\u626b\u63cf\u7b49\uff0c\u7136\u800c\uff0c\u50cf\u6240\u6709\u542f\u53d1\u5f0f\u7b97\u6cd5\u4e00\u6837\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u65f6\u7075\u65f6\u4e0d\u7075\uff0c\u5b83\u4eec\u53ef\u80fd\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u6709\u6548\uff0c\u4f46\u662f\u6ca1\u6709\u7406\u8bba\u4e0a\u7684\u4fdd\u8bc1\u3002<\/p>\n<p>\u4f20\u7edf\u7684\u5206\u5757\u91c7\u7528\u4e00\u79cd\u9884\u5148\u5206\u5757\u7684\u7b56\u7565\uff0c\u5373\u5148\u5206\u5757\uff0c\u518d\u8fc7 Embedding \u6a21\u578b\u3002\u9996\u5148\u4f9d\u636e\u53e5\u5b50\u3001\u6bb5\u843d\u6216\u9884\u8bbe\u7684\u6700\u5927\u957f\u5ea6\u7b49\u53c2\u6570\u5bf9\u6587\u672c\u8fdb\u884c\u5207\u5272\u3002\u7136\u540eEmbedding \u6a21\u578b\u4f1a\u5bf9\u8fd9\u4e9b\u5206\u5757\u9010\u4e00\u8fdb\u884c\u5904\u7406\uff0c\u901a\u8fc7\u5e73\u5747\u6c60\u5316\u7b49\u65b9\u6cd5\uff0c\u5c06 <a href=\"https:\/\/www.kdjingpai.com\/tokenization\/\">token<\/a> \u7ea7\u7684 Embedding \u805a\u5408\u6210\u5355\u4e00\u7684\u5757 Embedding \u5411\u91cf\u3002\u800cLate Chunking\u5219\u662f\u5148\u8fc7 Embedding \u6a21\u578b\u518d\u5206\u5757\uff08late\u7684\u542b\u4e49\u5c31\u662f\u5728\u4e8e\u6b64\uff0c\u5148\u5411\u91cf\u5316\u518d\u5206\u5757\uff09\uff0c\u6211\u4eec\u5148\u5c06 Embedding \u6a21\u578b\u7684 <a href=\"https:\/\/www.kdjingpai.com\/transformer\/\">transformer<\/a> \u5c42\u5e94\u7528\u5230\u6574\u4e2a\u6587\u672c\uff0c\u4e3a\u6bcf\u4e2a token \u751f\u6210\u4e00\u4e2a\u5305\u542b\u4e30\u5bcc\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u5411\u91cf\u8868\u793a\u5e8f\u5217\u3002\u7136\u540e\uff0c\u518d\u5bf9\u8fd9\u4e9b token \u5411\u91cf\u5e8f\u5217\u8fdb\u884c\u5e73\u5747\u6c60\u5316\uff0c\u6700\u7ec8\u5f97\u5230\u8003\u8651\u4e86\u6574\u4e2a\u6587\u672c\u4e0a\u4e0b\u6587\u7684\u5757 Embedding\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/b0aab587f375f5d.png\" \/><\/p>\n<p><em>(\u56fe\u7247\u6765\u6e90\uff1ahttps:\/\/jina.ai\/news\/late-chunking-in-long-context-embedding-models\/)<\/em><\/p>\n<p>Late Chunking\u751f\u6210\u7684\u5757Embedding\uff0c\u6bcf\u4e2a\u5757\u90fd\u7f16\u7801\u4e86\u66f4\u591a\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u7f16\u7801\u7684\u8d28\u91cf\u548c\u51c6\u786e\u6027\u3002\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u652f\u6301\u957f\u4e0a\u4e0b\u6587\u7684 Embedding \u6a21\u578b\uff0c\u5982 <code>jina-embeddings-v2-base-en<\/code>\uff0c\u5b83\u80fd\u591f\u5904\u7406\u957f\u8fbe8192\u4e2atoken \u7684\u6587\u672c(\u76f8\u5f53\u4e8e 10 \u9875 A4 \u7eb8)\uff0c\u57fa\u672c\u6ee1\u8db3\u4e86\u5927\u591a\u6570\u957f\u6587\u672c\u7684\u4e0a\u4e0b\u6587\u9700\u6c42\u3002<\/p>\n<p>\u7efc\u4e0a\u6240\u8ff0\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5230Late Chunking\u5728RAG\u5e94\u7528\u4e2d\u7684\u4f18\u52bf\uff1a<\/p>\n<ul>\n<li>\u63d0\u9ad8\u51c6\u786e\u6027\uff1a\u901a\u8fc7\u4fdd\u7559\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u4e0e\u7b80\u5355\u5206\u5757\u76f8\u6bd4\uff0cLate Chunking\u4e3a\u67e5\u8be2\u8fd4\u56de\u4e86\u76f8\u5173\u5ea6\u66f4\u9ad8\u7684\u5185\u5bb9\u3002<\/li>\n<li>\u9ad8\u6548\u7684LLM\u8c03\u7528\uff1aLate Chunking\u53ef\u4ee5\u51cf\u5c11\u4f20\u9012\u7ed9LLM\u7684\u6587\u672c\u91cf\uff0c\u56e0\u4e3a\u5b83\u8fd4\u56de\u7684\u5206\u5757\u66f4\u5c11\u4e14\u76f8\u5173\u5ea6\u66f4\u9ad8\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p><strong>03.<\/strong><strong>\u6d4b\u8bd5Late Chunking<\/strong><\/p>\n<p><strong>3.1. Late Chunking\u57fa\u7840\u5b9e\u73b0<\/strong><\/p>\n<p>\u51fd\u6570sentence_chunker\u5bf9\u4e8e\u539f\u59cb\u6587\u6863\u4ee5\u6bb5\u843d\u8fdb\u884c\u5206\u5757\uff0c\u8fd4\u56de\u5206\u5757\u5185\u5bb9\u4ee5\u53ca\u5206\u5757\u6807\u8bb0\u4fe1\u606fspan_annotations\uff08\u5373\u5206\u5757\u7684\u5f00\u59cb\u548c\u7ed3\u675f\u6807\u8bb0\uff09<\/p>\n<p>&nbsp;<\/p>\n<pre>def sentence_chunker(document, batch_size=10000):\r\n\u00a0\u00a0\u00a0\u00a0nlp\u00a0=\u00a0spacy.blank(\"en\")\r\n\u00a0\u00a0\u00a0\u00a0nlp.add_pipe(\"sentencizer\",\u00a0config={\"punct_chars\":\u00a0None})\r\n\u00a0\u00a0\u00a0\u00a0doc\u00a0=\u00a0nlp(document)\r\n\r\ndocs\u00a0=\u00a0[]\r\n\u00a0\u00a0\u00a0\u00a0for\u00a0i\u00a0in\u00a0range(0,\u00a0len(document),\u00a0batch_size):\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0batch\u00a0=\u00a0document[i\u00a0:\u00a0i\u00a0+\u00a0batch_size]\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0docs.append(nlp(batch))\r\n\r\ndoc\u00a0=\u00a0Doc.from_docs(docs)\r\n\r\nspan_annotations\u00a0=\u00a0[]\r\n\u00a0\u00a0\u00a0\u00a0chunks\u00a0=\u00a0[]\r\n\u00a0\u00a0\u00a0\u00a0for\u00a0i,\u00a0sent\u00a0in\u00a0enumerate(doc.sents):\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0span_annotations.append((sent.start,\u00a0sent.end))\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0chunks.append(sent.text)\r\n\r\nreturn chunks, span_annotations<\/pre>\n<p>&nbsp;<\/p>\n<p>\u51fd\u6570 document_to_token_embeddings \u901a\u8fc7\u6a21\u578b <code>jinaai\/jina-embeddings-v2-base-en<\/code> \u7684\u6a21\u578b\u4ee5\u53catokenizer\uff0c\u8fd4\u56de\u6574\u4e2a\u6587\u6863\u7684Embedding\u3002<\/p>\n<pre>def\u00a0document_to_token_embeddings(model,\u00a0tokenizer,\u00a0document,\u00a0batch_size=4096):\r\n\u00a0\u00a0\u00a0\u00a0tokenized_document\u00a0=\u00a0tokenizer(document,\u00a0return_tensors=\"pt\")\r\n\u00a0\u00a0\u00a0\u00a0<a href=\"https:\/\/www.kdjingpai.com\/tokenization\/\">tokens<\/a>\u00a0=\u00a0tokenized_document.tokens()\r\n\r\noutputs\u00a0=\u00a0[]\r\n\u00a0\u00a0\u00a0\u00a0for\u00a0i\u00a0in\u00a0range(0,\u00a0len(tokens),\u00a0batch_size):\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0start\u00a0=\u00a0i\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0end\u00a0\u00a0\u00a0=\u00a0min(i\u00a0+\u00a0batch_size,\u00a0len(tokens))\r\n\r\nbatch_inputs\u00a0=\u00a0{k:\u00a0v[:,\u00a0start:end]\u00a0for\u00a0k,\u00a0v\u00a0in\u00a0tokenized_document.items()}\r\n\r\nwith\u00a0torch.no_grad():\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0model_output\u00a0=\u00a0model(**batch_inputs)\r\n\r\noutputs.append(model_output.last_hidden_state)\r\n\r\nmodel_output\u00a0=\u00a0torch.cat(outputs,\u00a0dim=1)\r\n\u00a0\u00a0\u00a0\u00a0return\u00a0model_output<\/pre>\n<p>\u51fd\u6570 late_chunking \u5bf9\u6574\u4e2a\u6587\u6863\u7684Embedding\u4ee5\u53ca\u539f\u59cb\u5206\u5757\u7684\u6807\u8bb0\u4fe1\u606fspan_annotations\u8fdb\u884c\u5206\u5757\u3002<\/p>\n<pre>def\u00a0late_chunking(token_embeddings,\u00a0span_annotation,\u00a0max_length=None):\r\n\u00a0\u00a0\u00a0\u00a0outputs\u00a0=\u00a0[]\r\n\u00a0\u00a0\u00a0\u00a0for\u00a0embeddings,\u00a0annotations\u00a0in\u00a0zip(token_embeddings,\u00a0span_annotation):\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0if\u00a0(\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0max_length\u00a0is\u00a0not\u00a0None\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0):\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0annotations\u00a0=\u00a0[\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0(start,\u00a0min(end,\u00a0max_length\u00a0-\u00a01))\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0for\u00a0(start,\u00a0end)\u00a0in\u00a0annotations\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0if\u00a0start\u00a0&lt;\u00a0(max_length\u00a0-\u00a01)\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0]\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0pooled_embeddings\u00a0=\u00a0[]\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0for\u00a0start,\u00a0end\u00a0in\u00a0annotations:\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0if\u00a0(end\u00a0-\u00a0start)\u00a0&gt;=\u00a01:\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0pooled_embeddings.append(\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0embeddings[start:end].sum(dim=0)\u00a0\/\u00a0(end\u00a0-\u00a0start)\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0)\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0pooled_embeddings\u00a0=\u00a0[\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0embedding.detach().cpu().numpy()\u00a0for\u00a0embedding\u00a0in\u00a0pooled_embeddings\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0]\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0outputs.append(pooled_embeddings)\r\n\r\nreturn\u00a0outputs<\/pre>\n<p>\u5982\u4f7f\u7528\u6a21\u578b<code>jinaai\/jina-embeddings-v2-base-en<\/code>\u8fdb\u884cLate Chunking<\/p>\n<pre>tokenizer\u00a0=\u00a0AutoTokenizer.from_pretrained('jinaai\/jina-embeddings-v2-base-en',\u00a0trust_remote_code=True)\r\nmodel\u00a0\u00a0\u00a0\u00a0\u00a0=\u00a0AutoModel.from_pretrained('jinaai\/jina-embeddings-v2-base-en',\u00a0trust_remote_code=True)\r\n\r\n#\u00a0First\u00a0chunk\u00a0the\u00a0text\u00a0as\u00a0normal,\u00a0to\u00a0obtain\u00a0the\u00a0beginning\u00a0and\u00a0end\u00a0points\u00a0of\u00a0the\u00a0chunks.\r\nchunks,\u00a0span_annotations\u00a0=\u00a0sentence_chunker(document)\r\n#\u00a0Then\u00a0embed\u00a0the\u00a0full\u00a0document.\r\ntoken_embeddings\u00a0=\u00a0document_to_token_embeddings(model,\u00a0tokenizer,\u00a0document)\r\n#\u00a0Then\u00a0perform\u00a0the\u00a0late\u00a0chunking\r\nchunk_embeddings\u00a0=\u00a0late_chunking(token_embeddings,\u00a0[span_annotations])[0]<\/pre>\n<p><strong>3.2. \u4e0e\u4f20\u7edfEmbedding\u65b9\u6cd5\u5bf9\u6bd4<\/strong><\/p>\n<p>\u6211\u4eec\u4ee5milvus 2.4.13 release note \u8fd9\u4e00\u6bb5\u5185\u5bb9\u4e3a\u4f8b\uff0c<\/p>\n<blockquote><p>Milvus 2.4.13 introduces dynamic replica load, allowing users to adjust the number of collection replicas without needing to release and reload the collection.<\/p>\n<p>This version also addresses several critical bugs related to bulk importing, expression parsing, load balancing, and failure recovery.<\/p>\n<p>Additionally, significant improvements have been made to MMAP resource usage and import performance, enhancing overall system efficiency.<\/p>\n<p>We highly recommend upgrading to this release for better performance and stability.<\/p><\/blockquote>\n<p>\u5206\u522b\u8fdb\u884c\u4f20\u7edfEmbedding\uff0c\u5373\u5148\u5206\u5757\uff0c\u7136\u540e\u8fdb\u884cEmbedding\u3002\u4ee5\u53caLate Chunking\u65b9\u5f0fEmbedding\uff0c\u5373\u5148Embedding\uff0c\u7136\u540e\u518d\u5206\u5757\u3002\u7136\u540e\uff0c\u628a <code>milvus 2.4.13<\/code> \u5206\u522b\u4e0e\u8fd9\u4e24\u79cdEmbedding\u65b9\u5f0f\u7684\u7ed3\u679c\u8fdb\u884c\u5bf9\u6bd4<\/p>\n<pre>cos_sim\u00a0=\u00a0lambda\u00a0x,\u00a0y:\u00a0np.dot(x,\u00a0y)\u00a0\/\u00a0(np.linalg.norm(x)\u00a0*\u00a0np.linalg.norm(y))\r\n\r\nmilvus_embedding\u00a0=\u00a0model.encode('milvus\u00a02.4.13')\r\n\r\nfor\u00a0chunk,\u00a0late_chunking_embedding,\u00a0traditional_embedding\u00a0in\u00a0zip(chunks,\u00a0chunk_embeddings,\u00a0embeddings_traditional_chunking):\r\nprint(f'similarity_late_chunking(\"milvus\u00a02.4.13\",\u00a0\"{chunk}\")')\r\nprint('late_chunking:\u00a0',\u00a0cos_sim(milvus_embedding,\u00a0late_chunking_embedding))\r\nprint(f'similarity_traditional(\"milvus\u00a02.4.13\",\u00a0\"{chunk}\")')\r\nprint('traditional_chunking:\u00a0',\u00a0cos_sim(milvus_embedding,\u00a0traditional_embeddings))<\/pre>\n<p>\u4ece\u7ed3\u679c\u6765\u770b\uff0c\u8bcd\u8bed <code>milvus 2.4.13<\/code> \u4e0e\u5206\u5757\u6587\u6863Late Chunking\u7ed3\u679c\u76f8\u4f3c\u5ea6\u9ad8\u4e8e\u4f20\u7edfEmbedding\u3002\u539f\u56e0\u662fLate Chunking\u5148\u5bf9\u4e8e\u5168\u90e8\u6587\u672c\u6bb5\u843d\u8fdb\u884cEmbedding\uff0c\u4f7f\u5f97\u6574\u4e2a\u6587\u672c\u6bb5\u843d\u5f97\u5230\u4e86 <code>milvus 2.4.13<\/code> \u4fe1\u606f\uff0c\u8fdb\u800c\u5728\u540e\u7eed\u7684\u6587\u672c\u6bd4\u8f83\u4e2d\u663e\u8457\u7684\u63d0\u9ad8\u4e86\u76f8\u4f3c\u5ea6\u3002<\/p>\n<pre>similarity_late_chunking(\"milvus\u00a02.4.13\",\u00a0\"Milvus\u00a02.4.13\u00a0introduces\u00a0dynamic\u00a0replica\u00a0load,\u00a0allowing\u00a0users\u00a0to\u00a0adjust\u00a0the\u00a0number\u00a0of\u00a0collection\u00a0replicas\u00a0without\u00a0needing\u00a0to\u00a0release\u00a0and\u00a0reload\u00a0the\u00a0collection.\")\r\nlate_chunking:\u00a00.8785206\r\nsimilarity_traditional(\"milvus\u00a02.4.13\",\u00a0\"Milvus\u00a02.4.13\u00a0introduces\u00a0dynamic\u00a0replica\u00a0load,\u00a0allowing\u00a0users\u00a0to\u00a0adjust\u00a0the\u00a0number\u00a0of\u00a0collection\u00a0replicas\u00a0without\u00a0needing\u00a0to\u00a0release\u00a0and\u00a0reload\u00a0the\u00a0collection.\")\r\ntraditional_chunking:\u00a00.8354263\r\n\r\nsimilarity_late_chunking(\"milvus\u00a02.4.13\",\u00a0\"This\u00a0version\u00a0also\u00a0addresses\u00a0several\u00a0critical\u00a0bugs\u00a0related\u00a0to\u00a0bulk\u00a0importing,\u00a0expression\u00a0parsing,\u00a0load\u00a0balancing,\u00a0and\u00a0failure\u00a0recovery.\")\r\nlate_chunking:\u00a00.84828955\r\nsimilarity_traditional(\"milvus\u00a02.4.13\",\u00a0\"This\u00a0version\u00a0also\u00a0addresses\u00a0several\u00a0critical\u00a0bugs\u00a0related\u00a0to\u00a0bulk\u00a0importing,\u00a0expression\u00a0parsing,\u00a0load\u00a0balancing,\u00a0and\u00a0failure\u00a0recovery.\")\r\ntraditional_chunking:\u00a00.7222632\r\n\r\nsimilarity_late_chunking(\"milvus\u00a02.4.13\",\u00a0\"Additionally,\u00a0significant\u00a0improvements\u00a0have\u00a0been\u00a0made\u00a0to\u00a0MMAP\u00a0resource\u00a0usage\u00a0and\u00a0import\u00a0performance,\u00a0enhancing\u00a0overall\u00a0system\u00a0efficiency.\")\r\nlate_chunking:\u00a00.84942204\r\nsimilarity_traditional(\"milvus\u00a02.4.13\",\u00a0\"Additionally,\u00a0significant\u00a0improvements\u00a0have\u00a0been\u00a0made\u00a0to\u00a0MMAP\u00a0resource\u00a0usage\u00a0and\u00a0import\u00a0performance,\u00a0enhancing\u00a0overall\u00a0system\u00a0efficiency.\")\r\ntraditional_chunking:\u00a00.6907381\r\n\r\nsimilarity_late_chunking(\"milvus\u00a02.4.13\",\u00a0\"We\u00a0highly\u00a0recommend\u00a0upgrading\u00a0to\u00a0this\u00a0release\u00a0for\u00a0better\u00a0performance\u00a0and\u00a0stability.\")\r\nlate_chunking:\u00a00.85431844\r\nsimilarity_traditional(\"milvus\u00a02.4.13\",\u00a0\"We\u00a0highly\u00a0recommend\u00a0upgrading\u00a0to\u00a0this\u00a0release\u00a0for\u00a0better\u00a0performance\u00a0and\u00a0stability.\")\r\ntraditional_chunking:\u00a00.71859795<\/pre>\n<p><strong>3.3. Milvus\u4e2d\u6d4b\u8bd5Late Chunking<\/strong><\/p>\n<p><strong>\u5bfc\u5165Late Chunking\u6570\u636e\u5230Milvus<\/strong><\/p>\n<pre>batch_data=[]\r\nfor\u00a0i\u00a0in\u00a0range(len(chunks)):\r\n\u00a0\u00a0\u00a0\u00a0data\u00a0=\u00a0{\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\"content\":\u00a0chunks[i],\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\"embedding\":\u00a0chunk_embeddings[i].tolist(),\r\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0}\r\n\r\nbatch_data.append(data)\r\n\r\nres\u00a0=\u00a0client.insert(\r\ncollection_name=collection,\r\ndata=batch_data,\r\n)<\/pre>\n<p><strong>\u67e5\u8be2\u6d4b\u8bd5<\/strong><\/p>\n<p>\u6211\u4eec\u5b9a\u4e49cosine\u76f8\u4f3c\u5ea6\u67e5\u8be2\u65b9\u6cd5\uff0c\u4ee5\u53ca\u4f7f\u7528Milvus\u539f\u751f\u67e5\u8be2\u65b9\u6cd5\u5206\u522b\u5bf9\u4e8eLate Chunking\u8fdb\u884c\u67e5\u8be2\u3002<\/p>\n<pre>def\u00a0late_chunking_query_by_milvus(query,\u00a0top_k\u00a0=\u00a03):\r\n\u00a0\u00a0\u00a0\u00a0query_vector\u00a0=\u00a0model(**tokenizer(query,\u00a0return_tensors=\"pt\")).last_hidden_state.mean(1).detach().cpu().numpy().flatten()\r\n\r\nres\u00a0=\u00a0client.search(\r\ncollection_name=collection,\r\ndata=[query_vector.tolist()],\r\nlimit=top_k,\r\noutput_fields=[\"id\",\u00a0\"content\"],\r\n)\r\n\r\nreturn\u00a0[item.get(\"entity\").get(\"content\")\u00a0for\u00a0items\u00a0in\u00a0res\u00a0for\u00a0item\u00a0in\u00a0items]\r\n\r\ndef\u00a0late_chunking_query_by_cosine_sim(query,\u00a0k\u00a0=\u00a03):\r\ncos_sim\u00a0=\u00a0lambda\u00a0x,\u00a0y:\u00a0np.dot(x,\u00a0y)\u00a0\/\u00a0(np.linalg.norm(x)\u00a0*\u00a0np.linalg.norm(y))\r\nquery_vector\u00a0=\u00a0model(**tokenizer(query,\u00a0return_tensors=\"pt\")).last_hidden_state.mean(1).detach().cpu().numpy().flatten()\r\n\r\nresults\u00a0=\u00a0np.empty(len(chunk_embeddings))\r\nfor\u00a0i,\u00a0(chunk,\u00a0embedding)\u00a0in\u00a0enumerate(zip(chunks,\u00a0chunk_embeddings)):\r\nresults[i]\u00a0=\u00a0cos_sim(query_vector,\u00a0embedding)\r\n\r\nresults_order\u00a0=\u00a0results.argsort()[::-1]\r\nreturn\u00a0np.array(chunks)[results_order].tolist()[:k]<\/pre>\n<p>\u4ece\u7ed3\u679c\u6765\u770b\uff0c\u4e24\u4e2a\u65b9\u6cd5\u8fd4\u56de\u5185\u5bb9\u662f\u4e00\u81f4\u7684\uff0c\u8fd9\u8868\u660eMilvus\u4e2d\u5bf9\u4e8eLate Chunking\u67e5\u8be2\u7ed3\u679c\u662f\u51c6\u786e\u3002<\/p>\n<pre>&gt;\u00a0late_chunking_query_by_milvus(\"What\u00a0are\u00a0new\u00a0features\u00a0in\u00a0milvus\u00a02.4.13\",\u00a03)\r\n\r\n['nn###\u00a0Featuresnn-\u00a0Dynamic\u00a0replica\u00a0adjustment\u00a0for\u00a0loaded\u00a0collections\u00a0([#36417](https:\/\/github.com\/milvus-io\/milvus\/pull\/36417))n-\u00a0Sparse\u00a0vector\u00a0MMAP\u00a0in\u00a0growing\u00a0segment\u00a0types\u00a0([#36565](https:\/\/github.com\/milvus-io\/milvus\/pull\/36565))...\r\n\r\n&gt;\u00a0late_chunking_query_by_cosine_sim(\"What\u00a0are\u00a0new\u00a0features\u00a0in\u00a0milvus\u00a02.4.13\",\u00a03)\r\n\r\n['nn###\u00a0Featuresnn-\u00a0Dynamic\u00a0replica\u00a0adjustment\u00a0for\u00a0loaded\u00a0collections\u00a0([#36417](https:\/\/github.com\/milvus-io\/milvus\/pull\/36417))n-\u00a0Sparse\u00a0vector\u00a0MMAP\u00a0in\u00a0growing\u00a0segment\u00a0types\u00a0([#36565](https:\/\/github.com\/milvus-io\/milvus\/pull\/36565))...<\/pre>\n<p>&nbsp;<\/p>\n<p><strong>04.<\/strong><strong>\u603b\u7ed3<\/strong><\/p>\n<p>\u6211\u4eec\u4ecb\u7ecd\u4e86Late Chunking\u4ea7\u751f\u7684\u80cc\u666f\uff0c\u57fa\u672c\u6982\u5ff5\u4ee5\u53ca\u57fa\u7840\u5b9e\u73b0\uff0c\u7136\u540e\u901a\u8fc7\u5728Mivlus\u6d4b\u8bd5\u53d1\u73b0\uff0cLate Chunking\u6548\u679c\u4e0d\u9519\u3002\u603b\u4f53\u6765\u770b\uff0cLate Chunking\u5728\u51c6\u786e\u6027\u3001\u6548\u7387\u548c\u6613\u4e8e\u5b9e\u65bd\u65b9\u9762\u7684\u7ed3\u5408\uff0c\u4f7f\u5176\u6210\u4e3aRAG\u5e94\u7528\u7684\u4e00\u4e2a\u6709\u6548\u7684\u65b9\u6cd5\u3002<\/p>\n<p><strong>\u53c2\u8003\u6587\u6863:<\/strong><\/p>\n<ul>\n<li>https:\/\/stackoverflow.blog\/2024\/06\/06\/breaking-up-is-hard-to-do-chunking-in-rag-applications<\/li>\n<li>https:\/\/jina.ai\/news\/late-chunking-in-long-context-embedding-models\/<\/li>\n<li>https:\/\/jina.ai\/news\/what-late-chunking-really-is-and-what-its-not-part-ii\/<\/li>\n<\/ul>\n<p><strong>\u793a\u4f8b\u4ee3\u7801\uff1a<\/strong><\/p>\n<p>\u94fe\u63a5: https:\/\/pan.baidu.com\/s\/1cYNfZTTXd7RwjnjPFylReg?pwd=1234 \u63d0\u53d6\u7801: 1234\u4ee3\u7801\u5728 aws g4dn.xlarge \u673a\u5668\u4e0a\u8fd0\u884c<\/p>\n","protected":false},"excerpt":{"rendered":"<p>01.\u80cc\u666f \u5728RAG\u5e94\u7528\u5f00\u53d1\u4e2d\uff0c\u7b2c\u4e00\u6b65\u5c31\u662f\u5bf9\u4e8e\u6587\u6863\u8fdb\u884cchunking\uff08\u5206\u5757\uff09\uff0c\u9ad8\u6548\u7684\u6587\u6863\u5206\u5757\uff0c\u53ef\u4ee5\u6709\u6548\u7684\u63d0\u9ad8\u540e\u7eed\u7684\u53ec\u56de\u5185\u5bb9\u7684\u51c6\u786e\u6027\u3002\u800c\u5bf9\u4e8e\u5982\u4f55\u9ad8\u6548\u7684\u5206\u5757\u662f\u4e2a\u8ba8\u8bba\u7684\u70ed\u70b9\uff0c\u6709\u8bf8\u5982\u56fa\u5b9a\u5927\u5c0f\u5206\u5757\uff0c\u968f\u673a\u5927\u5c0f\u5206\u5757\uff0c\u6ed1\u52a8\u7a97\u53e3\u91cd\u65b0\u91c7\u6837\uff0c\u9012\u5f52\u5206\u5757\uff0c\u57fa\u4e8e\u5185\u5bb9&#8230;<\/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-16331","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts\/16331","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/comments?post=16331"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/posts\/16331\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/media?parent=16331"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/categories?post=16331"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/pt\/wp-json\/wp\/v2\/tags?post=16331"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}