{"id":109355,"date":"2026-09-24T00:56:43","date_gmt":"2026-09-23T16:56:43","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/109355.html"},"modified":"2026-09-24T00:56:43","modified_gmt":"2026-09-23T16:56:43","slug":"rag-%e6%96%87%e6%a1%a3%e5%88%87%e5%88%86%e7%ad%96%e7%95%a5%e5%85%a8%e6%99%af%e8%a7%a3%e6%9e%90%ef%bc%9a%e5%9b%ba%e5%ae%9a%e9%95%bf%e5%ba%a6-vs-%e8%af%ad%e4%b9%89%e5%88%87%e5%88%86%e7%9a%84%e6%8a%80","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/109355.html","title":{"rendered":"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5"},"content":{"rendered":"

RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790：\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5<\/h2>\n

\u6458\u8981<\/h3>\n

\u6587\u6863\u5207\u5206（Document Chunking）\u662f RAG（Retrieval-Augmented Generation）\u7cfb\u7edf\u4e2d\u6700\u57fa\u7840\u5374\u6700\u5bb9\u6613\u88ab\u4f4e\u4f30\u7684\u9884\u5904\u7406\u73af\u8282\u3002\u5207\u5206\u8d28\u91cf\u76f4\u63a5\u51b3\u5b9a\u68c0\u7d22\u53ec\u56de\u7387\u7684\u4e0a\u9650，\u8fdb\u800c\u5f71\u54cd LLM \u751f\u6210\u7b54\u6848\u7684\u51c6\u786e\u6027\u548c\u5e7b\u89c9\u7387\u3002\u672c\u6587\u7cfb\u7edf\u68b3\u7406\u4ece\u89c4\u5219\u5207\u5206\u5230\u8bed\u4e49\u5207\u5206\u3001\u4ece\u5ef6\u8fdf\u5207\u5206\u5230\u77e5\u8bc6\u56fe\u8c31\u7684\u5b8c\u6574\u6280\u672f\u6f14\u8fdb\u8def\u5f84，\u7ed3\u5408\u4ee3\u7801\u5b9e\u73b0\u3001\u53c2\u6570\u8c03\u4f18\u5efa\u8bae\u548c\u8bc4\u4f30\u65b9\u6cd5\u8bba，\u4e3a\u5de5\u7a0b\u5b9e\u8df5\u63d0\u4f9b\u53ef\u843d\u5730\u7684\u51b3\u7b56\u6846\u67b6\u3002<\/p>\n

\u5173\u952e\u8bcd：RAG\u3001\u6587\u6863\u5207\u5206\u3001Chunking\u3001\u8bed\u4e49\u5207\u5206\u3001Late Chunking\u3001Contextual Retrieval\u3001GraphRAG<\/p>\n

\n

1. \u5f15\u8a00：\u4e3a\u4ec0\u4e48\u5207\u5206\u662f RAG \u7684\u6838\u5fc3\u74f6\u9888<\/h3>\n

\u5728 RAG \u6807\u51c6\u5de5\u4f5c\u6d41\u4e2d，\u5207\u5206\u5904\u4e8e"\u6587\u6863\u89e3\u6790 \u2192 \u5207\u5206 \u2192 Embedding \u2192 \u68c0\u7d22 \u2192 \u751f\u6210"\u94fe\u8def\u7684\u6700\u524d\u7aef[citation:7]\u3002Amir Teymoori \u4e0e Chroma 2025 \u7684\u7814\u7a76\u90fd\u5c06\u5207\u5206\u5217\u4e3a"RAG \u6210\u8d25\u7684\u7b2c\u4e00\u53d8\u91cf"[citation:1]\u3002<\/p>\n

\u5207\u5206\u7684\u6839\u672c\u52a8\u56e0\u6709\u56db\u4e2a：<\/p>\n\n\u52a8\u56e0\u8bf4\u660e<\/tr>\n\n\n\n\n\n
\u4e0a\u4e0b\u6587\u7a97\u53e3\u9650\u5236<\/td>\n\u5408\u540c\u6570\u5341\u4e07\u5b57，\u65e0\u6cd5\u4e00\u6b21\u6027\u585e\u5165\u6a21\u578b；\u5373\u4f7f\u80fd\u585e，\u6210\u672c\u4e5f\u4e0d\u53ef\u63a5\u53d7<\/td>\n<\/tr>\n
\u68c0\u7d22\u7cbe\u5ea6<\/td>\n\u5411\u91cf\u8868\u793a\u5b58\u5728"\u5e73\u5747\u5316"\u503e\u5411\u2014\u2014\u6587\u672c\u8d8a\u957f，\u8bed\u4e49\u4fe1\u53f7\u8d8a\u88ab\u7a00\u91ca<\/td>\n<\/tr>\n
\u68c0\u7d22\u6548\u7387<\/td>\n\u5c0f\u5757\u624d\u80fd\u7cbe\u51c6\u547d\u4e2d"\u8fdd\u7ea6\u8d23\u4efb\u5728\u7b2c\u51e0\u6761"\u8fd9\u7c7b\u7ec6\u7c92\u5ea6\u67e5\u8be2<\/td>\n<\/tr>\n
\u751f\u6210\u8d28\u91cf<\/td>\n\u7247\u6bb5\u662f\u5426\u5b8c\u6574\u3001\u4e0a\u4e0b\u6587\u662f\u5426\u5145\u8db3，\u76f4\u63a5\u51b3\u5b9a\u5e7b\u89c9\u7387<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

1.1 \u6838\u5fc3\u77db\u76fe：\u7cbe\u51c6 vs \u5b8c\u6574<\/h4>\n

\u5757\u592a\u5c0f \u2192 \u68c0\u7d22\u7cbe\u51c6，\u4f46\u8131\u79bb\u4e0a\u4e0b\u6587（\u201c\u8be5\u65b9\u6848\u201d"\u4e0a\u8ff0\u65b9\u6cd5"\u60ac\u7a7a）
\n\u5757\u592a\u5927 \u2192 \u4e0a\u4e0b\u6587\u5b8c\u6574，\u4f46\u5411\u91cf\u88ab\u65e0\u5173\u5185\u5bb9\u7a00\u91ca，\u53ec\u56de\u7387\u4e0b\u964d<\/p>\n

\u8fd9\u4e00\u77db\u76fe\u8d2f\u7a7f\u6240\u6709\u5207\u5206\u65b9\u6cd5\u7684\u6f14\u8fdb\u3002\u7406\u89e3\u8fd9\u4e00\u70b9，\u5c31\u7406\u89e3\u4e86\u5207\u5206\u7b56\u7565\u7684\u672c\u8d28：\u5728\u7cbe\u51c6\u4e0e\u5b8c\u6574\u4e4b\u95f4\u5bfb\u627e\u6700\u4f18\u5e73\u8861\u70b9\u3002<\/p>\n

\n

2. \u5207\u5206\u7b56\u7565\u5168\u666f：\u4e09\u4ee3\u6280\u672f\u6f14\u8fdb<\/h3>\n

2.1 \u7b2c\u4e00\u4ee3：\u89c4\u5219\u5207\u5206（Structure-Agnostic）<\/h4>\n

2.1.1 \u56fa\u5b9a\u957f\u5ea6\u5207\u5206（Fixed-Size）<\/h5>\n

\u6309\u5b57\u7b26\u6570 \/ Token \u6570\u5747\u7b49\u5207\u5272，\u662f\u6700\u6734\u7d20\u7684\u65b9\u6cd5\u3002<\/p>\n

from<\/span> langchain.<\/span>text_splitter import<\/span> CharacterTextSplitter<\/p>\n

splitter =<\/span> CharacterTextSplitter(<\/span>
\n chunk_size=<\/span>512<\/span>,<\/span>
\n chunk_overlap=<\/span>50<\/span>,<\/span>
\n separator=<\/span>"\\\\n"<\/span>
\n)<\/span>
\nchunks =<\/span> splitter.<\/span>split_documents(<\/span>docs)<\/span><\/p>\n

\u4f18\u70b9：\u5b9e\u73b0\u6700\u7b80\u5355\u3001\u901f\u5ea6\u6700\u5feb\u3001\u5bf9\u4efb\u610f\u6587\u672c\u901a\u7528[citation:7]\u3002
\n\u7f3a\u70b9：\u5bb9\u6613\u5728\u53e5\u5b50\u4e2d\u95f4\u622a\u65ad，\u7834\u574f\u8bed\u4e49\u5b8c\u6574\u6027[citation:1]\u3002
\n\u9002\u7528\u573a\u666f：\u9ad8\u5ea6\u7edf\u4e00\u7684\u77ed\u6587\u672c（\u793e\u4ea4\u5e16\u5b50\u3001\u77ed\u6d88\u606f），\u6216\u4f5c\u4e3a\u57fa\u7ebf\u65b9\u6848[citation:1]\u3002<\/p>\n

2.1.2 \u9012\u5f52\u5b57\u7b26\u5207\u5206（Recursive Character）\u2014\u2014 \u5de5\u4e1a\u754c\u9996\u9009\u57fa\u7ebf<\/h5>\n

\u6309\u5206\u9694\u7b26\u4f18\u5148\u7ea7 \\\\n\\\\n \u2192 \\\\n \u2192 \u7a7a\u683c \u2192 \u5b57\u7b26 \u9010\u7ea7\u56de\u9000，\u4f18\u5148\u5728\u8bed\u4e49\u5b8c\u6574\u5904\u4e0b\u5200[citation:1][citation:10]\u3002<\/p>\n

from<\/span> langchain.<\/span>text_splitter import<\/span> RecursiveCharacterTextSplitter<\/p>\n

splitter =<\/span> RecursiveCharacterTextSplitter(<\/span>
\n chunk_size=<\/span>512<\/span>,<\/span>
\n chunk_overlap=<\/span>100<\/span>,<\/span>
\n length_function=<\/span>len<\/span> # \u5efa\u8bae\u7528 tiktoken \u7cbe\u786e\u8ba1\u6570<\/span>
\n)<\/span>
\nchunks =<\/span> splitter.<\/span>split_documents(<\/span>docs)<\/span><\/p>\n

\u5173\u952e\u53c2\u6570（\u7ecf\u9a8c\u503c）[citation:1]：<\/p>\n\n\u53c2\u6570\u9ed8\u8ba4\u503c\u8bf4\u660e<\/tr>\n\n\n\n\n
chunk_size<\/td>\n400\u2013512 token<\/td>\n\u77ed\u6587 200\u2013400；\u957f\u6587\/\u8bba\u6587 512\u2013800<\/td>\n<\/tr>\n
chunk_overlap<\/td>\n10\u201320%<\/td>\n\u7ed3\u6784\u5316\u6587\u6863 5\u201310%；\u975e\u7ed3\u6784\u5316 15\u201320%<\/td>\n<\/tr>\n
length_function<\/td>\ntiktoken<\/td>\n\u5fc5\u987b\u786e\u4fdd\u4e0e Embedding \u6a21\u578b\u7684 token \u8ba1\u6570\u4e00\u81f4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u26a0\ufe0f \u5e38\u89c1\u8bef\u533a：chunk_overlap \u5fc5\u987b\u4e0e chunk_size \u4f7f\u7528\u540c\u4e00\u5355\u4f4d（\u90fd\u662f token \u6216\u90fd\u662f char），\u5426\u5219\u884c\u4e3a\u4f1a"\u5bf9\u4e0d\u4e0a"[citation:1]\u3002<\/p>\n

\u9012\u5f52\u5207\u5206\u662f LangChain \u9ed8\u8ba4\u7b56\u7565，\u88ab\u79f0\u4e3a"\u751f\u4ea7\u73af\u5883 80% \u7684\u9009\u62e9"[citation:1]，\u9002\u5408\u4f5c\u4e3a\u9879\u76ee\u8d77\u70b9\u3002<\/p>\n

2.1.3 \u7ed3\u6784\u611f\u77e5\u5207\u5206（Structure-Aware）<\/h5>\n

\u76f4\u63a5\u5229\u7528\u6587\u6863\u56fa\u6709\u7ed3\u6784：\u6807\u9898\u5c42\u7ea7\u3001\u6bb5\u843d\u3001\u8868\u683c\u3001\u4ee3\u7801\u5757\u7b49[citation:13]\u3002<\/p>\n

from<\/span> langchain.<\/span>text_splitter import<\/span> MarkdownHeaderTextSplitter<\/p>\n

headers_to_split_on =<\/span> [<\/span>
\n (<\/span>"#"<\/span>,<\/span> "h1"<\/span>)<\/span>,<\/span>
\n (<\/span>"##"<\/span>,<\/span> "h2"<\/span>)<\/span>,<\/span>
\n (<\/span>"###"<\/span>,<\/span> "h3"<\/span>)<\/span>,<\/span>
\n]<\/span>
\nsplitter =<\/span> MarkdownHeaderTextSplitter(<\/span>headers_to_split_on=<\/span>headers_to_split_on)<\/span>
\nchunks =<\/span> splitter.<\/span>split_text(<\/span>markdown_text)<\/span><\/p>\n

\u4e24\u9636\u6bb5\u4e32\u8054\u7b56\u7565（\u63a8\u8350）：\u5148\u7528 MarkdownHeaderTextSplitter \u6cbf\u6807\u9898\u5206\u5927\u5757，\u518d\u5c06\u5927\u5757\u5582\u7ed9 RecursiveCharacterTextSplitter \u63a7\u5927\u5c0f\u2014\u2014\u65e2\u4fdd\u7ed3\u6784\u53c8\u4fdd\u957f\u5ea6[citation:1]\u3002<\/p>\n

PDF \u9875\u9762\u7ea7\u5207\u5206\u5728\u5b9e\u8df5\u4e2d\u4e5f"\u76f8\u5f53\u80fd\u6253"[citation:1]，\u5c24\u5176\u9002\u5408\u62a5\u544a\u3001\u8bba\u6587\u7b49\u683c\u5f0f\u7a33\u5b9a\u7684\u6587\u6863\u3002<\/p>\n

2.2 \u7b2c\u4e8c\u4ee3：\u8bed\u4e49\u5207\u5206（Semantic Chunking）<\/h4>\n

\u6838\u5fc3\u601d\u60f3：\u4e0d\u4f9d\u8d56\u7269\u7406\u8fb9\u754c，\u800c\u662f\u901a\u8fc7\u8ba1\u7b97\u76f8\u90bb\u53e5\u5b50\u7684 Embedding \u76f8\u4f3c\u5ea6，\u5728"\u8bed\u4e49\u7a81\u53d8\u70b9"\u5207\u5206[citation:17]\u3002<\/p>\n

from<\/span> langchain_experimental.<\/span>text_splitter import<\/span> SemanticChunker
\nfrom<\/span> langchain_openai import<\/span> OpenAIEmbeddings<\/p>\n

splitter =<\/span> SemanticChunker(<\/span>
\n OpenAIEmbeddings(<\/span>)<\/span>,<\/span>
\n breakpoint_threshold_type=<\/span>"percentile"<\/span>,<\/span>
\n breakpoint_threshold_amount=<\/span>50<\/span> # \u76f8\u4f3c\u5ea6\u4e0b\u964d\u8d85\u8fc7 50 \u767e\u5206\u4f4d\u65f6\u5207\u5206<\/span>
\n)<\/span>
\nchunks =<\/span> splitter.<\/span>split_text(<\/span>document)<\/span><\/p>\n

\u7b97\u6cd5\u6d41\u7a0b：<\/p>\n

  • \u5c06\u6587\u672c\u5207\u5206\u4e3a\u53e5\u5b50\u5e8f\u5217 [s\u2081, s\u2082, …, s\u2099]<\/li>\n
  • \u8ba1\u7b97\u6bcf\u4e2a\u53e5\u5b50\u7684 Embedding \u5411\u91cf<\/li>\n
  • \u8ba1\u7b97\u76f8\u90bb\u53e5\u5bf9\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6 sim(e\u1d62, e\u1d62\u208a\u2081)<\/li>\n
  • \u5f53\u76f8\u4f3c\u5ea6\u663e\u8457\u4e0b\u964d（\u4f4e\u4e8e\u9608\u503c）\u65f6，\u6807\u8bb0\u5207\u5206\u70b9<\/li>\n
  • \u5c06\u8fb9\u754c\u4e4b\u95f4\u7684\u53e5\u5b50\u5408\u5e76\u4e3a\u8bed\u4e49\u4e00\u81f4\u7684\u5757<\/li>\n

    \u6548\u679c：Chroma \u7814\u7a76\u663e\u793a，LLM \u8bed\u4e49\u5207\u5206\u53ef\u8fbe 91.9% recall（\u6240\u6709\u7b56\u7565\u6700\u9ad8）[citation:1]\u3002
    \n\u4ee3\u4ef7：\u6bcf\u53e5\u90fd\u9700\u8981 Embedding，\u8ba1\u7b97\u6210\u672c\u7ebf\u6027\u589e\u52a0[citation:1]\u3002<\/p>\n

    2.3 \u7b2c\u4e09\u4ee3：\u4e0a\u4e0b\u6587\u589e\u5f3a\u5207\u5206<\/h4>\n

    2.3.1 Late Chunking（\u5ef6\u8fdf\u5207\u5206）<\/h5>\n

    \u6838\u5fc3\u6d1e\u5bdf：\u4f20\u7edf"\u5148\u5207\u540e\u5d4c"\u5bfc\u81f4\u6bcf\u4e2a\u5757\u5728\u8131\u79bb\u5168\u6587\u8bed\u5883\u4e0b\u7f16\u7801，\u201cits"\u548c"the city"\u65e0\u6cd5\u5173\u8054\u5230"Berlin\u201d[citation:2][citation:5]\u3002<\/p>\n

    \u5de5\u4f5c\u6d41\u7a0b[citation:5]：<\/p>\n

  • \u6574\u4f53\u7f16\u7801：\u5c06\u5b8c\u6574\u6587\u6863（\u6700\u957f 8192 token）\u4e00\u6b21\u6027\u8f93\u5165\u957f\u4e0a\u4e0b\u6587 Embedding \u6a21\u578b，\u83b7\u53d6\u6bcf\u4e2a Token \u7684\u5411\u91cf\u8868\u793a<\/li>\n
  • \u8fb9\u754c\u6620\u5c04：\u5c06\u9884\u5b9a\u4e49\u7684\u5757\u8fb9\u754c\u6620\u5c04\u5230 Token Embedding \u5e8f\u5217<\/li>\n
  • \u5747\u503c\u6c60\u5316：\u5bf9\u6bcf\u4e2a\u5757\u7684 Token Embedding \u505a Mean Pooling，\u5f97\u5230\u5e26\u5168\u5c40\u4e0a\u4e0b\u6587\u7684\u5757\u5411\u91cf<\/li>\n

    import<\/span> requests
    \nimport<\/span> numpy as<\/span> np<\/p>\n

    def<\/span> late_chunk<\/span>(<\/span>document_text,<\/span> chunk_boundaries)<\/span>:<\/span>
    \n """\u5ef6\u8fdf\u5207\u5206：\u5148\u5d4c\u5165\u5168\u6587，\u518d\u63d0\u53d6\u5757\u5411\u91cf"""<\/span>
    \n # Step 1: \u83b7\u53d6\u5168\u6587\u7684 Token \u7ea7 Embedding<\/span>
    \n response =<\/span> requests.<\/span>post(<\/span>
    \n EMBEDDING_API_ENDPOINT,<\/span>
    \n headers=<\/span>{<\/span>"Authorization"<\/span>:<\/span> f"Bearer <\/span>{<\/span>JINA_API_KEY}<\/span><\/span>"<\/span><\/span>}<\/span>,<\/span>
    \n json=<\/span>{<\/span>
    \n "model"<\/span>:<\/span> "jina-embeddings-v3"<\/span>,<\/span>
    \n "input"<\/span>:<\/span> [<\/span>document_text]<\/span>,<\/span>
    \n "task"<\/span>:<\/span> "retrieval.passage"<\/span>,<\/span>
    \n "late_chunking"<\/span>:<\/span> True<\/span>
    \n }<\/span>
    \n )<\/span>
    \n token_embeddings =<\/span> response.<\/span>json(<\/span>)<\/span>[<\/span>"data"<\/span>]<\/span>[<\/span>0<\/span>]<\/span>[<\/span>"embeddings"<\/span>]<\/span><\/p>\n

    # Step 2: \u5bf9\u6bcf\u4e2a\u5757\u505a Mean Pooling<\/span>
    \n chunk_embeddings =<\/span> [<\/span>]<\/span>
    \n for<\/span> start,<\/span> end in<\/span> chunk_boundaries:<\/span>
    \n chunk_emb =<\/span> np.<\/span>mean(<\/span>token_embeddings[<\/span>start:<\/span>end]<\/span>,<\/span> axis=<\/span>0<\/span>)<\/span>
    \n chunk_embeddings.<\/span>append(<\/span>chunk_emb)<\/span><\/p>\n

    return<\/span> chunk_embeddings<\/p>\n

    \u5173\u952e\u7ed3\u679c[citation:5]：<\/p>\n

      \n
    • \u4f7f\u7528\u53e5\u5b50\u8fb9\u754c\u65f6，\u76f8\u5bf9\u63d0\u5347 3.63%（\u7edd\u5bf9 1.9%）<\/li>\n
    • \u4f7f\u7528\u56fa\u5b9a\u7a97\u53e3\u65f6，\u76f8\u5bf9\u63d0\u5347 3.46%（\u7edd\u5bf9 1.8%）<\/li>\n
    • \u4f7f\u7528\u8bed\u4e49\u5207\u5206\u65f6，\u76f8\u5bf9\u63d0\u5347 2.70%（\u7edd\u5bf9 1.5%）<\/li>\n<\/ul>\n

      \u641c\u7d22 \u201cBerlin\u201d \u65f6，\u6734\u7d20\u5207\u5206\u5bf9\u542b \u201cits\u201d \u548c \u201cthe city\u201d \u7684\u5757\u7ed9\u51fa 0.70\u20130.75 \u7684\u76f8\u4f3c\u5ea6；Late Chunking \u5c06\u5176\u63d0\u5347\u81f3 0.82\u20130.85[citation:5]\u3002<\/p>\n

      \u9650\u5236：\u53d7\u9650\u4e8e Embedding \u6a21\u578b\u7684\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6；\u5bf9\u4e8e\u8d85\u957f\u6587\u6863，Jina \u63d0\u51fa\u4e86 \u201cLong Late Chunking\u201d（\u91cd\u53e0\u5b8f\u5757\u5904\u7406）[citation:5]\u3002<\/p>\n

      2.3.2 Contextual Retrieval（\u4e0a\u4e0b\u6587\u68c0\u7d22）<\/h5>\n

      \u7531 Anthropic \u4e8e 2024 \u5e74 9 \u6708\u63d0\u51fa[citation:8]，\u601d\u8def\u662f\u8ba9 LLM \u4e3a\u6bcf\u4e2a\u5757\u751f\u6210\u4e00\u6bb5"\u51fa\u8eab\u8bf4\u660e"，\u518d\u4e0e\u539f\u6587\u62fc\u63a5\u540e Embedding\u3002<\/p>\n

      Prompt \u6a21\u677f：
      \n{{WHOLE_DOCUMENT}}
      \nHere is the chunk we want to situate within the whole document:
      \n{{CHUNK_CONTENT}}
      \nPlease give a short succinct context to situate this chunk…
      \nAnswer only with the succinct context and nothing else.<\/p>\n

      \u751f\u6210\u7684 50\u2013100 token \u4e0a\u4e0b\u6587\u88ab\u524d\u7f6e\u5230\u539f\u59cb\u5757，\u518d\u4e00\u8d77 Embedding[citation:8]\u3002<\/p>\n

      \u4e0e Late Chunking \u7684\u5bf9\u6bd4[citation:8]：<\/p>\n\n\u7ef4\u5ea6Late ChunkingContextual Retrieval<\/tr>\n\n\n\n\n\n\n
      \u4e0a\u4e0b\u6587\u6765\u6e90<\/td>\nEncoder \u7684 Attention \u77e9\u9635<\/td>\nLLM \u751f\u6210\u7684\u6587\u672c\u63cf\u8ff0<\/td>\n<\/tr>\n
      \u662f\u5426\u9700\u8981 LLM<\/td>\n\u274c \u4ec5\u9700 Embedding \u6a21\u578b<\/td>\n\u2705 \u6bcf\u4e2a\u5757\u90fd\u9700\u8981 LLM \u8c03\u7528<\/td>\n<\/tr>\n
      \u7d22\u5f15\u8017\u65f6（\u6bcf\u6587\u6863）<\/td>\n15.81 ms<\/td>\n1890.94 ms（\u7ea6 120\u00d7）<\/td>\n<\/tr>\n
      nDCG@10<\/td>\n61.0<\/td>\n72.4<\/td>\n<\/tr>\n
      \u9002\u7528\u89c4\u6a21<\/td>\n\u5927\u89c4\u6a21\u79bb\u7ebf\u7d22\u5f15<\/td>\n\u9ad8\u4ef7\u503c\u3001\u4f4e\u6570\u91cf\u6587\u6863<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      \u7ed3\u8bba：Late Chunking \u662f"\u514d\u8d39\u7684\u5348\u9910"\u2014\u2014\u51e0\u4e4e\u4e0d\u589e\u52a0\u989d\u5916\u6210\u672c\u5373\u53ef\u83b7\u5f97\u4e0a\u4e0b\u6587\u611f\u77e5；Contextual Retrieval \u8d28\u91cf\u66f4\u9ad8\u4f46\u4ee3\u4ef7\u6602\u8d35，\u9002\u5408\u5173\u952e\u6587\u6863[citation:8]\u3002<\/p>\n

      \n

      3. \u8fdb\u9636\u67b6\u6784：\u4ece\u6241\u5e73\u5207\u5206\u5230\u5c42\u6b21\u5316\u7d22\u5f15<\/h3>\n

      3.1 Small-to-Big（\u7236\u5b50\u6587\u6863\u7b56\u7565）<\/h4>\n

      \u601d\u8def\u6781\u5176\u6734\u7d20：\u68c0\u7d22\u7528\u5c0f\u5757，\u751f\u6210\u7528\u5927\u5757[citation:1]\u3002<\/p>\n

      \u4e09\u6b65\u6d41\u7a0b：<\/p>\n

    • \u7528\u7ec6\u7c92\u5ea6\u5b50\u5757（256\u2013512 token）\u505a\u5411\u91cf\u68c0\u7d22，\u4fdd\u8bc1\u53ec\u56de\u7cbe\u51c6<\/li>\n
    • \u547d\u4e2d\u540e，\u901a\u8fc7\u5143\u6570\u636e\u56de\u6eaf\u5230\u7236\u5757（1024\u20132048 token）<\/li>\n
    • \u5c06\u5b8c\u6574\u7684\u7236\u5757\u5582\u7ed9 LLM \u751f\u6210\u7b54\u6848<\/li>\n

      # \u4f2a\u4ee3\u7801\u793a\u610f<\/span>
      \nchild_chunks =<\/span> split_into_small(<\/span>document,<\/span> size=<\/span>256<\/span>)<\/span> # \u5b50\u5757\u505a\u68c0\u7d22<\/span>
      \nparent_map =<\/span> build_parent_mapping(<\/span>child_chunks,<\/span> document,<\/span> parent_size=<\/span>1024<\/span>)<\/span><\/p>\n

      retrieved_children =<\/span> vector_store.<\/span>search(<\/span>query,<\/span> top_k=<\/span>5<\/span>)<\/span>
      \nretrieved_parents =<\/span> [<\/span>parent_map[<\/span>c]<\/span> for<\/span> c in<\/span> retrieved_children]<\/span> # \u56de\u6eaf\u7236\u5757<\/span>
      \nanswer =<\/span> llm.<\/span>generate(<\/span>query,<\/span> context=<\/span>retrieved_parents)<\/span><\/p>\n

      \u9002\u7528\u573a\u666f：\u5927\u591a\u6570\u95ee\u7b54\u578b RAG \u7cfb\u7edf，\u6027\u4ef7\u6bd4\u6700\u9ad8\u7684\u7ec4\u5408\u62f3[citation:1]\u3002<\/p>\n

      3.2 RAPTOR：\u9012\u5f52\u6458\u8981\u6811<\/h4>\n

      \u7531\u65af\u5766\u798f\u5927\u5b66\u63d0\u51fa，ICLR 2024 \u53d1\u8868[citation:9]\u3002\u6838\u5fc3\u601d\u60f3：\u5c06\u5207\u597d\u7684\u5757\u4e00\u5c42\u4e00\u5c42\u505a\u805a\u7c7b+\u6458\u8981，\u957f\u6210\u4e00\u68f5"\u6458\u8981\u6811"\u3002<\/p>\n

      \u6784\u5efa\u6d41\u7a0b：<\/p>\n

    • \u53f6\u5b50\u5c42：\u539f\u59cb\u6587\u6863\u5757（\u7ec6\u7c92\u5ea6）<\/li>\n
    • \u805a\u7c7b：\u7528 K-Means \/ \u5c42\u6b21\u805a\u7c7b\u5c06\u8bed\u4e49\u76f8\u8fd1\u7684\u5757\u805a\u5408<\/li>\n
    • \u6458\u8981：LLM \u4e3a\u6bcf\u4e2a\u805a\u7c7b\u751f\u6210\u9ad8\u5c42\u6458\u8981（\u7236\u8282\u70b9）<\/li>\n
    • \u9012\u5f52：\u5c06\u6458\u8981\u518d\u6b21\u805a\u7c7b\u3001\u6458\u8981，\u76f4\u5230\u6839\u8282\u70b9<\/li>\n

      \u68c0\u7d22\u65b9\u5f0f：\u4ece\u6839\u8282\u70b9\u5f00\u59cb\u81ea\u4e0a\u800c\u4e0b\u904d\u5386，\u5b8f\u89c2\u95ee\u9898\u547d\u4e2d\u4e0a\u5c42\u6458\u8981，\u7ec6\u8282\u95ee\u9898\u4e00\u8def\u5411\u4e0b\u5230\u53f6\u5b50[citation:3][citation:12]\u3002<\/p>\n

      \u6548\u679c：\u5728 QuALITY \u57fa\u51c6\u4e0a，\u914d\u5408 GPT-4 \u6700\u4f73\u6027\u80fd\u63d0\u5347 20%[citation:9]\u3002<\/p>\n

      3.3 GraphRAG：\u77e5\u8bc6\u56fe\u8c31 + \u793e\u533a\u805a\u7c7b<\/h4>\n

      \u7531\u5fae\u8f6f\u63d0\u51fa，\u601d\u8def\u66f4\u5f7b\u5e95：\u5c06\u6587\u6863\u4e2d\u7684\u5b9e\u4f53\u548c\u5173\u7cfb\u62bd\u53d6\u4e3a\u77e5\u8bc6\u56fe\u8c31，\u6309\u793e\u533a（Leiden \u7b97\u6cd5）\u805a\u7c7b\u5e76\u751f\u6210\u6458\u8981[citation:6]\u3002<\/p>\n

      \u9002\u7528\u573a\u666f：\u8de8\u6587\u6863\u3001\u591a\u8df3\u63a8\u7406\u7684\u5168\u5c40\u95ee\u7b54（\u5982"\u8fd9\u4e9b\u6750\u6599\u6574\u4f53\u5728\u8bb2\u4ec0\u4e48"）[citation:3]\u3002
      \n\u4ee3\u4ef7：\u6784\u5efa\u6210\u672c\u9ad8，\u9002\u5408\u79bb\u7ebf\u6162\u6162\u5efa[citation:9]\u3002<\/p>\n

      \n

      4. \u4e2d\u6587\u5207\u5206\u7684\u7279\u6b8a\u8003\u91cf<\/h3>\n

      \u4e2d\u6587\u5207\u5206\u76f8\u6bd4\u82f1\u6587\u9762\u4e34\u989d\u5916\u6311\u6218[citation:13][citation:23]：<\/p>\n\n\u6311\u6218\u8bf4\u660e\u89e3\u51b3\u65b9\u6848<\/tr>\n\n\n\n\n\n
      \u65e0\u5929\u7136\u7a7a\u683c\u8fb9\u754c<\/td>\n\u4e0d\u80fd\u7528\u7a7a\u683c\u5206\u8bcd<\/td>\n\u4f7f\u7528 jieba \/ HanLP \u7b49\u5206\u8bcd\u5de5\u5177<\/td>\n<\/tr>\n
      \u6807\u70b9\u7528\u6cd5\u7075\u6d3b<\/td>\n\u201c\u3002\u201d \u4e4b\u5916\u8fd8\u6709 \u201c！\u201d \u201c？\u201d \u7b49<\/td>\n\u5b8c\u5584\u7684\u6b63\u5219\u5206\u53e5\u89c4\u5219<\/td>\n<\/tr>\n
      \u957f\u53e5\u65e0\u6807\u70b9<\/td>\n\u4e00\u5927\u957f\u53e5\u4e0d\u5e26\u6807\u70b9<\/td>\n\u7ed3\u5408\u8bed\u4e49\u5206\u6790\u505a\u4e8c\u6b21\u5207\u5206<\/td>\n<\/tr>\n
      \u56db\u5b57\u6210\u8bed\/\u53e4\u8bd7\u8bcd<\/td>\n\u5fc5\u987b\u4fdd\u6301\u5b8c\u6574<\/td>\n\u81ea\u5b9a\u4e49\u8bcd\u5178 + \u89c4\u5219\u4fdd\u62a4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      import<\/span> jieba
      \nimport<\/span> re<\/p>\n

      def<\/span> chinese_sentence_split<\/span>(<\/span>text)<\/span>:<\/span>
      \n """\u4e2d\u6587\u5206\u53e5：\u7ed3\u5408\u6807\u70b9\u548c\u6b63\u5219"""<\/span>
      \n sentences =<\/span> re.<\/span>split(<\/span>r'(?<=[\u3002！？；\\\\.\\\\!\\\\?\\\\;])'<\/span>,<\/span> text)<\/span>
      \n return<\/span> [<\/span>s.<\/span>strip(<\/span>)<\/span> for<\/span> s in<\/span> sentences if<\/span> s.<\/span>strip(<\/span>)<\/span>]<\/span><\/p>\n

      def<\/span> chinese_chunk_with_jieba<\/span>(<\/span>text,<\/span> chunk_size=<\/span>512<\/span>)<\/span>:<\/span>
      \n """\u7ed3\u5408 jieba \u5206\u8bcd\u7684\u4e2d\u6587\u5207\u5206"""<\/span>
      \n words =<\/span> jieba.<\/span>lcut(<\/span>text,<\/span> cut_all=<\/span>False<\/span>)<\/span>
      \n chunks =<\/span> [<\/span>]<\/span>
      \n current =<\/span> [<\/span>]<\/span>
      \n current_len =<\/span> 0<\/span>
      \n for<\/span> word in<\/span> words:<\/span>
      \n if<\/span> current_len +<\/span> len<\/span>(<\/span>word)<\/span> ><\/span> chunk_size and<\/span> current:<\/span>
      \n chunks.<\/span>append(<\/span>""<\/span>.<\/span>join(<\/span>current)<\/span>)<\/span>
      \n current =<\/span> [<\/span>word]<\/span>
      \n current_len =<\/span> len<\/span>(<\/span>word)<\/span>
      \n else<\/span>:<\/span>
      \n current.<\/span>append(<\/span>word)<\/span>
      \n current_len +=<\/span> len<\/span>(<\/span>word)<\/span>
      \n if<\/span> current:<\/span>
      \n chunks.<\/span>append(<\/span>""<\/span>.<\/span>join(<\/span>current)<\/span>)<\/span>
      \n return<\/span> chunks<\/p>\n

      \u5b9e\u8df5\u5efa\u8bae：\u5148\u7528\u7ed3\u6784\u8fb9\u754c\u505a\u7b2c\u4e00\u6b21\u5207\u5206，\u518d\u7528\u53e5\u7ea7\/\u9012\u5f52\u7b56\u7565\u505a\u4e8c\u6b21\u7ec6\u5206；\u4f18\u5148\u4f7f\u7528"\u7ed3\u6784\u91cd\u53e0"（\u7236\u6807\u9898\u8def\u5f84\u3001\u4e0a\u6bb5\u6807\u9898+\u6458\u8981）\u66ff\u4ee3\u5927\u6bd4\u4f8b\u5b57\u7b26\u91cd\u53e0[citation:27]\u3002<\/p>\n

      \n

      5. \u5de5\u7a0b\u9009\u578b\u51b3\u7b56\u6811<\/h3>\n

      \u7efc\u5408\u4ee5\u4e0a\u5206\u6790，\u7ed9\u51fa\u4ee5\u4e0b\u9009\u578b\u5efa\u8bae[citation:1][citation:13]：<\/p>\n

      \u6587\u6863\u662f\u5426\u6709\u6e05\u6670\u7ed3\u6784？
      \n\u251c\u2500\u2500 \u662f（Markdown \/ HTML \/ PDF \/ \u4ee3\u7801）
      \n\u2502 \u2192 \u7ed3\u6784\u5207\u5206（\u6807\u9898\u5c42\u7ea7）+ \u9012\u5f52\u5b57\u7b26\u5207\u5206（\u4e24\u9636\u6bb5）
      \n\u2502
      \n\u2514\u2500\u2500 \u5426（\u7eaf\u6587\u672c \/ \u6df7\u5408\u5185\u5bb9）
      \n \u2502
      \n \u251c\u2500\u2500 \u9700\u8981\u6781\u5feb\u901f\u5ea6 \/ \u539f\u578b\u9a8c\u8bc1？
      \n \u2502 \u2192 \u56fa\u5b9a\u957f\u5ea6\u5207\u5206（512 token + 10% overlap）
      \n \u2502
      \n \u251c\u2500\u2500 \u9700\u8981\u8bed\u4e49\u5b8c\u6574\u6027（\u9ad8\u8d28\u91cf\u95ee\u7b54）？
      \n \u2502 \u2192 \u8bed\u4e49\u5207\u5206（SemanticChunker）
      \n \u2502
      \n \u251c\u2500\u2500 \u6587\u6863\u6709\u9891\u7e41\u8de8\u5757\u5f15\u7528（"\u5982\u4e0a\u6240\u8ff0""\u8be6\u89c1\u7b2c\u4e8c\u8282"）？
      \n \u2502 \u2192 Late Chunking \/ Contextual Retrieval
      \n \u2502
      \n \u2514\u2500\u2500 \u95ee\u7b54\u578b RAG，\u517c\u987e\u7cbe\u5ea6\u548c\u4e0a\u4e0b\u6587？
      \n \u2192 Small-to-Big（\u7236\u5b50\u6587\u6863\u7b56\u7565）<\/p>\n

      \u67e5\u8be2\u662f\u5426\u8de8\u6587\u6863 \/ \u591a\u8df3？
      \n\u251c\u2500\u2500 \u662f \u2192 RAPTOR（\u5c42\u7ea7\u603b\u7ed3）\/ GraphRAG（\u77e5\u8bc6\u56fe\u8c31）
      \n\u2514\u2500\u2500 \u5426 \u2192 \u4e0a\u8ff0\u7b56\u7565\u5df2\u8db3\u591f<\/p>\n

      5.1 \u901f\u67e5\u8868<\/h4>\n\n\u573a\u666f\u63a8\u8350\u7b56\u7565\u53c2\u6570\u8d77\u70b9\u9884\u671f\u6548\u679c<\/tr>\n\n\n\n\n\n\n\n
      \u901a\u7528\u57fa\u7ebf<\/td>\n\u9012\u5f52\u5b57\u7b26\u5207\u5206<\/td>\n512 token, 10% overlap<\/td>\n\u7a33\u5b9a\u53ef\u8dd1\u901a<\/td>\n<\/tr>\n
      Markdown\/PDF\/\u4ee3\u7801<\/td>\n\u7ed3\u6784\u5207\u5206 + \u9012\u5f52<\/td>\n\u6309\u6807\u9898\u5206\u5927\u5757，\u518d\u63a7 512<\/td>\n\u7ed3\u6784\u5b8c\u6574<\/td>\n<\/tr>\n
      FAQ\/\u77ed\u95ee\u7b54<\/td>\n\u53e5\u5b50\u5207\u5206 + \u805a\u5408<\/td>\n200\u2013400 token<\/td>\n\u7cbe\u786e\u5339\u914d<\/td>\n<\/tr>\n
      \u957f\u6587\u6863\/\u8bba\u6587<\/td>\n\u8bed\u4e49\u5207\u5206 \/ Late Chunking<\/td>\n512\u2013800 token<\/td>\n\u8bed\u4e49\u8fde\u8d2f<\/td>\n<\/tr>\n
      \u95ee\u7b54\u578b RAG<\/td>\nSmall-to-Big<\/td>\n\u5b50 256 \/ \u7236 1024<\/td>\n\u7cbe\u5ea6+\u4e0a\u4e0b\u6587<\/td>\n<\/tr>\n
      \u8de8\u6587\u6863\/\u591a\u8df3<\/td>\nRAPTOR \/ GraphRAG<\/td>\n\u2014<\/td>\n\u5168\u5c40\u89c6\u91ce<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
      \n

      6. \u5e38\u89c1\u9677\u9631\u4e0e\u89c4\u907f<\/h3>\n

      6.1 \u8868\u683c\u548c\u7248\u9762\u7ed3\u6784\u4e22\u5931<\/h4>\n

      \u8868\u683c\u88ab\u786c\u5207\u6210\u7eaf\u6587\u672c\u540e，\u884c\u5217\u5173\u7cfb\u5168\u4e22\u3002\u89e3\u51b3\u65b9\u6848：\u5207\u5206\u524d\u5148\u505a\u7ed3\u6784\u89e3\u6790，\u5c06\u8868\u683c\u8f6c\u4e3a"\u8868\u5934:\u5355\u5143\u683c\u503c"\u7684\u683c\u5f0f[citation:13]\u3002<\/p>\n

      6.2 Overlap \u8fc7\u5927<\/h4>\n

      \u8d85\u8fc7 30% \u7684 overlap \u901a\u5e38\u5bfc\u81f4\u7d22\u5f15\u4f53\u79ef\u4e0e\u68c0\u7d22\u5f00\u9500\u663e\u8457\u4e0a\u5347，\u4f46\u5bf9\u6548\u679c\u7684\u589e\u76ca\u8fb9\u9645\u9012\u51cf[citation:13]\u3002\u5efa\u8bae\u4ece 10\u201320% \u8d77\u6b65\u3002<\/p>\n

      6.3 \u7b54\u6848\u6a2a\u8de8\u4e24\u4e2a\u76f8\u90bb\u5757<\/h4>\n

      \u7ed3\u8bba\u5728\u5757 A，\u8bba\u636e\u5728\u5757 B\u3002Reranker \u53ea\u80fd\u8c03\u987a\u5e8f，\u65e0\u6cd5"\u53d8\u51fa"\u672a\u53ec\u56de\u7684\u5757[citation:1]\u3002\u89e3\u51b3\u65b9\u6848：\u9002\u5f53\u589e\u52a0 overlap \u6216\u91c7\u7528 Small-to-Big \u56de\u6eaf\u7236\u5757\u3002<\/p>\n

      6.4 \u5207\u5206\u7b56\u7565\u7684\u5f71\u54cd\u88ab\u4f4e\u4f30<\/h4>\n

      \u5207\u5206\u7b56\u7565\u5bf9\u68c0\u7d22\u8d28\u91cf\u7684\u5f71\u54cd，\u53ef\u80fd\u6bd4\u6362\u4e00\u4e2a\u66f4\u5f3a\u7684 Embedding \u6a21\u578b\u8fd8\u8981\u5927[citation:1]\u3002<\/p>\n

      \u4e0e\u5176\u7ea0\u7ed3\u6362\u6a21\u578b，\u4e0d\u5982\u5148\u628a\u5207\u5206\u53c2\u6570\u8c03\u597d\u2014\u2014\u6536\u76ca\u5f80\u5f80\u66f4\u5b9e\u5728\u3002<\/p>\n

      6.5 Metadata \u7f3a\u5931<\/h4>\n

      \u6bcf\u4e2a\u5757\u5e94\u4fdd\u7559\u6765\u6e90\u3001\u7ae0\u8282\u6807\u9898\u3001\u9875\u7801\u7b49\u5143\u6570\u636e[citation:32]，\u8fd9\u5bf9\u540e\u7eed Reranking \u548c\u7b54\u6848\u6eaf\u6e90\u81f3\u5173\u91cd\u8981\u3002<\/p>\n

      \n

      7. \u8bc4\u4f30\u9a71\u52a8：\u5982\u4f55\u79d1\u5b66\u8c03\u4f18\u5207\u5206\u7b56\u7565<\/h3>\n

      \u6838\u5fc3\u539f\u5219：\u5207\u5206\u53c2\u6570\u4e0d\u80fd\u62cd\u8111\u888b，\u5fc5\u987b\u7528\u6570\u636e\u8bf4\u8bdd[citation:26][citation:30]\u3002<\/p>\n

      7.1 \u56db\u6b65\u8bc4\u4f30\u6d41\u7a0b<\/h4>\n

      Step 1：\u89e3\u6790\u5f52\u4e00\u5316<\/h5>\n

      \u5c06\u4e0d\u540c\u683c\u5f0f\u6587\u6863\u7edf\u4e00\u4e3a\u5e26\u7ed3\u6784\u6807\u7b7e\u7684\u6587\u672c[citation:1]\u3002<\/p>\n

      Step 2：\u5efa\u7acb\u57fa\u7ebf<\/h5>\n

      \u7528\u9012\u5f52\u5207\u5206 + overlap \u8dd1\u901a\u5168\u94fe\u8def，\u83b7\u53d6 Recall@K \u57fa\u51c6\u503c[citation:36]\u3002<\/p>\n

      Step 3：\u6784\u5efa\u8bc4\u6d4b\u96c6<\/h5>\n

      \u51c6\u5907 30\u2013200 \u6761（\u95ee\u9898, \u671f\u671b\u539f\u6587\u6bb5\u843d）\u5bf9[citation:24][citation:30]\u3002<\/p>\n

      \u26a0\ufe0f \u91cd\u8981：\u7528\u771f\u5b9e\u7528\u6237\u67e5\u8be2\u6784\u5efa，\u800c\u975e LLM \u5408\u6210\u95ee\u9898\u3002\u5408\u6210\u95ee\u9898\u7684\u63aa\u8f9e\u4e0e\u6587\u6863\u9ad8\u5ea6\u4e00\u81f4，\u4f1a\u865a\u9ad8\u6240\u6709\u7b56\u7565\u7684\u53ec\u56de\u7387[citation:30]\u3002<\/p>\n

      Step 4：\u6d88\u878d\u5b9e\u9a8c<\/h5>\n

      \u5728\u540c\u4e00\u6570\u636e\u96c6\u4e0a\u5bf9\u6bd4\u4e0d\u540c\u7b56\u7565：<\/p>\n

      def<\/span> recall_at_k<\/span>(<\/span>retriever,<\/span> golden_set,<\/span> k=<\/span>5<\/span>)<\/span>:<\/span>
      \n """Recall@K \u8bc4\u4f30\u811a\u624b\u67b6"""<\/span>
      \n hits =<\/span> 0<\/span>
      \n for<\/span> query,<\/span> expected_passage in<\/span> golden_set:<\/span>
      \n results =<\/span> retriever.<\/span>retrieve(<\/span>query,<\/span> top_k=<\/span>k)<\/span>
      \n if<\/span> any<\/span>(<\/span>expected_passage in<\/span> r.<\/span>text for<\/span> r in<\/span> results)<\/span>:<\/span>
      \n hits +=<\/span> 1<\/span>
      \n return<\/span> hits \/<\/span> len<\/span>(<\/span>golden_set)<\/span><\/p>\n

      # \u5bf9\u6bd4\u4e0d\u540c\u7b56\u7565<\/span>
      \nstrategies =<\/span> [<\/span>
      \n (<\/span>"Fixed 512"<\/span>,<\/span> fixed_retriever)<\/span>,<\/span>
      \n (<\/span>"Recursive 512"<\/span>,<\/span> recursive_retriever)<\/span>,<\/span>
      \n (<\/span>"Semantic"<\/span>,<\/span> semantic_retriever)<\/span>,<\/span>
      \n (<\/span>"Small-to-Big"<\/span>,<\/span> parent_child_retriever)<\/span>,<\/span>
      \n]<\/span><\/p>\n

      for<\/span> name,<\/span> retriever in<\/span> strategies:<\/span>
      \n r5 =<\/span> recall_at_k(<\/span>retriever,<\/span> golden_set,<\/span> k=<\/span>5<\/span>)<\/span>
      \n r10 =<\/span> recall_at_k(<\/span>retriever,<\/span> golden_set,<\/span> k=<\/span>10<\/span>)<\/span>
      \n print<\/span>(<\/span>f"<\/span>{<\/span>name}<\/span><\/span>: Recall@5=<\/span>{<\/span>r5:<\/span>.2%<\/span>}<\/span><\/span>, Recall@10=<\/span>{<\/span>r10:<\/span>.2%<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

      7.2 \u5173\u952e\u8bc4\u4f30\u6307\u6807<\/h4>\n\n\u6307\u6807\u542b\u4e49\u8bf4\u660e<\/tr>\n\n\n\n\n\n\n\n\n
      Recall@K<\/td>\nTop-K \u662f\u5426\u5305\u542b\u76ee\u6807\u5757<\/td>\n\u6700\u6838\u5fc3\u6307\u6807，\u51b3\u5b9a\u7b54\u6848\u8d28\u91cf\u4e0a\u9650<\/td>\n<\/tr>\n
      Precision@K<\/td>\n\u53ec\u56de\u5757\u4e2d\u6709\u591a\u5c11\u76f8\u5173<\/td>\n\u63a7\u5236\u566a\u58f0<\/td>\n<\/tr>\n
      MRR<\/td>\n\u7b2c\u4e00\u4e2a\u76f8\u5173\u5757\u7684\u4f4d\u7f6e<\/td>\n\u53cd\u6620\u6392\u5e8f\u8d28\u91cf<\/td>\n<\/tr>\n
      nDCG@K<\/td>\n\u76f8\u5173\u5757\u7684\u6392\u5e8f\u8d28\u91cf（\u52a0\u6743）<\/td>\n\u591a\u76f8\u5173\u5757\u573a\u666f<\/td>\n<\/tr>\n
      Faithfulness<\/td>\n\u751f\u6210\u7b54\u6848\u662f\u5426\u5fe0\u4e8e\u4e0a\u4e0b\u6587<\/td>\n\u7aef\u5230\u7aef\u6307\u6807<\/td>\n<\/tr>\n
      Latency<\/td>\n\u68c0\u7d22\u5ef6\u8fdf<\/td>\n\u5de5\u7a0b\u7ea6\u675f<\/td>\n<\/tr>\n
      Index Size<\/td>\n\u7d22\u5f15\u4f53\u79ef<\/td>\n\u6210\u672c\u7ea6\u675f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      7.3 \u5b9e\u9a8c\u77e9\u9635\u5efa\u8bae<\/h4>\n

      \u53c2\u8003 StackAI \u7684\u5b9e\u8df5[citation:24]，\u5efa\u8bae\u8dd1\u4e00\u4e2a\u53c2\u6570\u7f51\u683c：<\/p>\n

        \n
      • chunk_size \u00d7 overlap：300\/500\/800 token \u00d7 0%\/10%\/20% overlap<\/li>\n
      • \u7b56\u7565\u5bf9\u6bd4：\u9012\u5f52 vs \u53e5\u5b50 vs \u8bed\u4e49 vs \u5c42\u6b21\u5316<\/li>\n
      • Reranker：\u6709\/\u65e0 \u5404\u8dd1\u4e00\u7ec4<\/li>\n<\/ul>\n

        \u5982\u679c\u4e24\u79cd\u7b56\u7565\u5dee\u8ddd\u5728 1\u20132% \u4ee5\u5185，\u4e0d\u503c\u5f97\u8fc7\u5ea6\u4f18\u5316；\u8d85\u8fc7 5%，\u5207\u5206\u5c31\u662f\u7b2c\u4e00\u4f18\u5148\u7ea7[citation:36]\u3002<\/p>\n

        \n

        8. \u524d\u6cbf\u8d8b\u52bf<\/h3>\n

        8.1 Agentic Chunking<\/h4>\n

        \u8ba9 Agent \u6309\u6587\u6863\u7c7b\u578b\u52a8\u6001\u9009\u62e9\u5207\u5206\u7b56\u7565（Markdown \u4e00\u79cd\u3001\u4ee3\u7801\u4e00\u79cd\u3001\u5bf9\u8bdd\u4e00\u79cd）\u3002Amir Teymoori \u9884\u6d4b：2025 \u4e0b\u534a\u5e74 Agentic Chunking \u5c06\u8fdb\u5165\u4e3b\u6d41\u751f\u4ea7\u7ea7 RAG[citation:1]\u3002<\/p>\n

        8.2 \u591a\u6a21\u6001\u5207\u5206<\/h4>\n

        \u56fe\u6587\u5173\u8054\u5207\u5206\u3001\u89c6\u9891\u6309\u573a\u666f\u5206\u5272\u3001\u97f3\u9891\u9759\u97f3\u68c0\u6d4b\u7b49[citation:4]，\u5c06\u5207\u5206\u4ece\u7eaf\u6587\u672c\u6269\u5c55\u5230\u591a\u6a21\u6001\u9886\u57df\u3002<\/p>\n

        8.3 Layout-Aware Segmentation<\/h4>\n

        \u6700\u65b0\u7814\u7a76（\u5982 DocETL\u3001BookRAG）\u5f3a\u8c03\u4fdd\u7559\u6587\u6863\u539f\u59cb\u7248\u9762\u4fe1\u606f（\u6bb5\u843d\u3001\u8868\u683c\u3001\u56fe\u3001\u516c\u5f0f），\u907f\u514d\u56fa\u5b9a chunk \u5bfc\u81f4\u7684\u4fe1\u606f\u788e\u7247\u5316[citation:6]\u3002<\/p>\n

        \n

        9. \u603b\u7ed3<\/h3>\n

        \u6587\u6863\u5207\u5206\u6ca1\u6709\u94f6\u5f39\u3002\u672c\u6587\u68b3\u7406\u7684\u6280\u672f\u5168\u666f\u53ef\u5f52\u7eb3\u4e3a\u4e09\u4e2a\u5c42\u6b21：<\/p>\n

      • \u89c4\u5219\u5c42（\u56fa\u5b9a\/\u9012\u5f52\/\u7ed3\u6784）：\u7b80\u5355\u9ad8\u6548，\u662f\u5de5\u7a0b\u8d77\u70b9<\/li>\n
      • \u8bed\u4e49\u5c42（\u8bed\u4e49\u5207\u5206\/Late Chunking\/Contextual Retrieval）：\u63d0\u5347\u8bed\u4e49\u8fde\u8d2f\u6027，\u4ee3\u4ef7\u662f\u8ba1\u7b97\u6210\u672c<\/li>\n
      • \u67b6\u6784\u5c42（Small-to-Big\/RAPTOR\/GraphRAG）：\u901a\u8fc7\u5c42\u6b21\u5316\u7d22\u5f15\u89e3\u51b3\u590d\u6742\u67e5\u8be2<\/li>\n

        \u5de5\u7a0b\u5b9e\u8df5\u7684\u6838\u5fc3\u5fc3\u6cd5：<\/p>\n

        \u5148\u8dd1\u901a\u57fa\u7ebf \u2192 \u6784\u5efa\u8bc4\u6d4b\u96c6 \u2192 \u6d88\u878d\u5b9e\u9a8c \u2192 \u6570\u636e\u9a71\u52a8\u9009\u578b \u2192 \u6301\u7eed\u76d1\u63a7\u8fed\u4ee3<\/p>\n

        \u56f4\u7ed5\u68c0\u7d22\u8d28\u91cf\u3001\u4e0a\u4e0b\u6587\u5b8c\u6574\u6027\u548c\u6210\u672c\u5ef6\u8fdf\u8fd9\u4e09\u4e2a\u7ef4\u5ea6，\u5728\u4f60\u7684\u771f\u5b9e\u6570\u636e\u4e0a\u4e0d\u65ad\u5b9e\u9a8c，\u624d\u80fd\u6536\u655b\u5230\u6700\u4f18\u89e3\u3002<\/p>\n

        \n

        \u53c2\u8003\u8d44\u6599<\/h3>\n
      • Chroma. (2025). RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u4e0e\u5b9e\u73b0\u7ec6\u8282\u2014\u2014\u5168\u7f51\u6280\u672f\u6c47\u603b<\/li>\n
      • Jina AI. (2024). Late Chunking: Contextual Chunk Embeddings Using Long-Context Embedding Models. arXiv:2409.04701<\/li>\n
      • Sarthi, P. et al. (2024). RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval. ICLR 2024<\/li>\n
      • Redis. (2025). Chunking for RAG: Strategies, tradeoffs & common mistakes<\/li>\n
      • Denser AI. (2025). RAG Chunking Strategies 2026: 8 Methods Compared<\/li>\n
      • BetterYeah. (2025). AI\u77e5\u8bc6\u5e93\u5982\u4f55\u8fdb\u884c\u6587\u672c\u5207\u5206？6\u79cd\u65b9\u6cd5<\/li>\n
      • \u963f\u91cc\u4e91\u5f00\u53d1\u8005\u793e\u533a. (2025). RAG\u5206\u5757\u6280\u672f\u5168\u666f\u56fe：5\u5927\u7b56\u7565\u89e3\u5256<\/li>\n
      • Particula Tech. (2025). Contextual Retrieval vs Late Chunking: Which One to Use<\/li>\n
      • \u5f97\u7269\u6280\u672f. (2025). RAG\u2014Chunking\u7b56\u7565\u5b9e\u6218<\/li>\n
      • Airbyte. (2025). RAG Document Chunking: 6 Best Practices<\/li>\n
      • StackAI. (2025). Chunking Strategies for RAG: How to Optimize Document Retrieval<\/li>\n
      • Beefed AI. (2025). Chunking Strategies for Reliable RAG<\/li>\n
      • CSDN. (2025). RAG \u7cfb\u7edf\u4e2d\u6587\u6863\u5207\u5206\u4e0e\u5411\u91cf\u5316\u7684\u6838\u5fc3\u6280\u672f\u4e0e\u5b9e\u8df5<\/li>\n","protected":false},"excerpt":{"rendered":"

        RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790：\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5
        \n\u6458\u8981
        \n\u6587\u6863\u5207\u5206（Document Chunking）\u662f RAG（Retrieval-Augmented Generation）\u7cfb\u7edf\u4e2d\u6700\u57fa\u7840\u5374\u6700\u5bb9\u6613\u88ab\u4f4e\u4f30\u7684\u9884\u5904\u7406\u73af\u8282\u3002\u5207\u5206\u8d28\u91cf\u76f4\u63a5\u51b3\u5b9a\u68c0\u7d22\u53ec\u56de\u7387\u7684\u4e0a\u9650，\u8fdb\u800c\u5f71\u54cd LLM \u751f\u6210\u7b54\u6848\u7684\u51c6\u786e\u6027\u548c\u5e7b\u89c9\u7387\u3002\u672c\u6587\u7cfb\u7edf\u68b3\u7406\u4ece\u89c4\u5219\u5207\u5206\u5230\u8bed\u4e49\u5207\u5206\u3001\u4ece\u5ef6\u8fdf\u5207\u5206\u5230\u77e5\u8bc6\u56fe\u8c31\u7684\u5b8c\u6574\u6280\u672f\u6f14\u8fdb\u8def\u5f84&#<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[1699,52,50,188,51],"topic":[],"class_list":["post-109355","post","type-post","status-publish","format-standard","hentry","category-server","tag-agi","tag-aigc","tag-50","tag-188","tag-51"],"yoast_head":"\nRAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/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:\/\/www.wsisp.com\/helps\/109355.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790：\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5 \u6458\u8981 \u6587\u6863\u5207\u5206（Document Chunking）\u662f RAG（Retrieval-Augmented Generation）\u7cfb\u7edf\u4e2d\u6700\u57fa\u7840\u5374\u6700\u5bb9\u6613\u88ab\u4f4e\u4f30\u7684\u9884\u5904\u7406\u73af\u8282\u3002\u5207\u5206\u8d28\u91cf\u76f4\u63a5\u51b3\u5b9a\u68c0\u7d22\u53ec\u56de\u7387\u7684\u4e0a\u9650，\u8fdb\u800c\u5f71\u54cd LLM \u751f\u6210\u7b54\u6848\u7684\u51c6\u786e\u6027\u548c\u5e7b\u89c9\u7387\u3002\u672c\u6587\u7cfb\u7edf\u68b3\u7406\u4ece\u89c4\u5219\u5207\u5206\u5230\u8bed\u4e49\u5207\u5206\u3001\u4ece\u5ef6\u8fdf\u5207\u5206\u5230\u77e5\u8bc6\u56fe\u8c31\u7684\u5b8c\u6574\u6280\u672f\u6f14\u8fdb\u8def\u5f84&#\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/109355.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-23T16:56:43+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/109355.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/109355.html\",\"name\":\"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-23T16:56:43+00:00\",\"dateModified\":\"2026-09-23T16:56:43+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/109355.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/109355.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/109355.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","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:\/\/www.wsisp.com\/helps\/109355.html","og_locale":"zh_CN","og_type":"article","og_title":"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790：\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5 \u6458\u8981 \u6587\u6863\u5207\u5206（Document Chunking）\u662f RAG（Retrieval-Augmented Generation）\u7cfb\u7edf\u4e2d\u6700\u57fa\u7840\u5374\u6700\u5bb9\u6613\u88ab\u4f4e\u4f30\u7684\u9884\u5904\u7406\u73af\u8282\u3002\u5207\u5206\u8d28\u91cf\u76f4\u63a5\u51b3\u5b9a\u68c0\u7d22\u53ec\u56de\u7387\u7684\u4e0a\u9650，\u8fdb\u800c\u5f71\u54cd LLM \u751f\u6210\u7b54\u6848\u7684\u51c6\u786e\u6027\u548c\u5e7b\u89c9\u7387\u3002\u672c\u6587\u7cfb\u7edf\u68b3\u7406\u4ece\u89c4\u5219\u5207\u5206\u5230\u8bed\u4e49\u5207\u5206\u3001\u4ece\u5ef6\u8fdf\u5207\u5206\u5230\u77e5\u8bc6\u56fe\u8c31\u7684\u5b8c\u6574\u6280\u672f\u6f14\u8fdb\u8def\u5f84&#","og_url":"https:\/\/www.wsisp.com\/helps\/109355.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-23T16:56:43+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/109355.html","url":"https:\/\/www.wsisp.com\/helps\/109355.html","name":"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-23T16:56:43+00:00","dateModified":"2026-09-23T16:56:43+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/109355.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/109355.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/109355.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"RAG \u6587\u6863\u5207\u5206\u7b56\u7565\u5168\u666f\u89e3\u6790\uff1a\u56fa\u5b9a\u957f\u5ea6 vs \u8bed\u4e49\u5207\u5206\u7684\u6280\u672f\u6743\u8861\u4e0e\u5de5\u7a0b\u5b9e\u8df5"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/109355","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=109355"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/109355\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=109355"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=109355"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=109355"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=109355"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}