{"id":26952,"date":"2025-02-25T17:17:21","date_gmt":"2025-02-25T09:17:21","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=26952"},"modified":"2025-02-25T17:17:21","modified_gmt":"2025-02-25T09:17:21","slug":"yizhangtujieshiqingben","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/yizhangtujieshiqingben\/","title":{"rendered":"\u4e00\u5f20\u56fe\u89e3\u91ca\u6e05\u695a\u6784\u5efaRAG\u7cfb\u7edf\u5168\u8c8c"},"content":{"rendered":"

\u8fd9\u5f20\u56fe\u6e05\u6670\u5730\u63cf\u7ed8\u4e86\u4e00\u4e2a\u73b0\u4ee3\u5316\u7684\u3001\u590d\u6742\u7684\u95ee\u9898\u89e3\u7b54\uff08Question Answering, QA\uff09\u6216\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08Retrieval-Augmented Generation, RAG\uff09\u7cfb\u7edf\u7684\u67b6\u6784\u84dd\u56fe\u3002\u5b83\u4ece\u7528\u6237\u63d0\u51fa\u95ee\u9898\u5f00\u59cb\uff0c\u4e00\u76f4\u5230\u6700\u7ec8\u751f\u6210\u7b54\u6848\uff0c\u8be6\u7ec6\u5c55\u793a\u4e86\u4e2d\u95f4\u7684\u5404\u4e2a\u5173\u952e\u73af\u8282\u548c\u6280\u672f\u9009\u62e9\u3002\u6211\u4eec\u53ef\u4ee5\u5c06\u6574\u4e2a\u6d41\u7a0b\u5206\u89e3\u4e3a\u4ee5\u4e0b\u51e0\u4e2a\u6838\u5fc3\u9636\u6bb5\uff1a<\/p>\n

\"\u4e00\u5f20\u56fe\u89e3\u91ca\u6e05\u695aRAG-1\"<\/p>\n

 <\/p>\n

1\u3001\u67e5\u8be2\u6784\u5efa (Query Construction)<\/h3>\n

\u8fd9\u662f\u7528\u6237\u4e0e\u7cfb\u7edf\u4ea4\u4e92\u7684\u7b2c\u4e00\u6b65\uff0c\u4e5f\u662f\u7cfb\u7edf\u7406\u89e3\u7528\u6237\u610f\u56fe\u7684\u8d77\u70b9\u3002\u56fe\u50cf\u4e2d\u5c55\u793a\u4e86\u9488\u5bf9\u4e0d\u540c\u7c7b\u578b\u6570\u636e\u5e93\u7684\u67e5\u8be2\u6784\u5efa\u65b9\u5f0f\uff1a
\na. \u5173\u7cfb\u578b\u6570\u636e\u5e93 (Relational DBs): \u00a0\u5bf9\u4e8e\u5173\u7cfb\u578b\u6570\u636e\u5e93\uff0c\u5e38\u89c1\u7684\u67e5\u8be2\u6784\u5efa\u65b9\u5f0f\u662f Text-to-SQL\u3002\u8fd9\u610f\u5473\u7740\u7cfb\u7edf\u9700\u8981\u5c06\u7528\u6237\u7684\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u8f6c\u5316\u4e3a\u7ed3\u6784\u5316\u7684SQL\u67e5\u8be2\u8bed\u53e5\u3002\u8fd9\u901a\u5e38\u6d89\u53ca\u5230\u81ea\u7136\u8bed\u8a00\u7406\u89e3 (NLU) \u6280\u672f\uff0c\u4ee5\u53ca\u5c06\u81ea\u7136\u8bed\u8a00\u8bed\u4e49\u6620\u5c04\u5230SQL\u8bed\u6cd5\u548c\u6570\u636e\u5e93Schema\u7684\u80fd\u529b\u3002\u56fe\u4e2d\u8fd8\u63d0\u5230\u4e86 SQL w\/ PGVector\uff0c\u53ef\u4ee5\u7ed3\u5408\u5411\u91cf\u6570\u636e\u5e93 (PGVector\u662fPostgreSQL\u7684\u5411\u91cf\u6269\u5c55) \u6765\u589e\u5f3aSQL\u67e5\u8be2\u7684\u80fd\u529b\uff0c\u4f8b\u5982\u8fdb\u884c\u8bed\u4e49\u76f8\u4f3c\u5ea6\u641c\u7d22\uff0c\u4ece\u800c\u66f4\u7075\u6d3b\u5730\u5904\u7406\u7528\u6237\u7684\u6a21\u7cca\u6216\u8bed\u4e49\u5316\u7684\u67e5\u8be2\u3002
\nb. \u56fe\u6570\u636e\u5e93 (Graph DBs): \u5bf9\u4e8e\u56fe\u6570\u636e\u5e93\uff0c\u5bf9\u5e94\u7684\u67e5\u8be2\u6784\u5efa\u65b9\u5f0f\u662f Text-to-Cypher\u3002Cypher\u662f\u4e00\u79cd\u7528\u4e8e\u56fe\u6570\u636e\u5e93Neo4j\u7684\u67e5\u8be2\u8bed\u8a00\uff0c\u7c7b\u4f3c\u4e8eSQL\u4f46\u66f4\u9002\u5408\u56fe\u7ed3\u6784\u7684\u67e5\u8be2\u3002Text-to-Cypher \u9700\u8981\u5c06\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u8f6c\u5316\u4e3aCypher\u67e5\u8be2\u8bed\u53e5\uff0c\u8fd9\u9700\u8981\u7406\u89e3\u56fe\u6570\u636e\u5e93\u7684\u7ed3\u6784\u548c\u56fe\u67e5\u8be2\u8bed\u8a00\u7684\u7279\u6027\u3002
\nc. \u5411\u91cf\u6570\u636e\u5e93<\/a> (Vector DBs): \u5bf9\u4e8e\u5411\u91cf\u6570\u636e\u5e93\uff0c\u56fe\u4e2d\u5c55\u793a\u4e86 Self-query retriever\u3002\u8fd9\u610f\u5473\u7740\u7cfb\u7edf\u53ef\u4ee5\u6839\u636e\u7528\u6237\u7684\u95ee\u9898\uff0c\u81ea\u52a8\u751f\u6210\u5143\u6570\u636e\u8fc7\u6ee4\u5668\uff0c\u5e76\u76f4\u63a5\u67e5\u8be2\u5411\u91cf\u6570\u636e\u5e93\u3002\u5411\u91cf\u6570\u636e\u5e93\u901a\u5e38\u5b58\u50a8\u6587\u672c\u6216\u6570\u636e\u7684\u5411\u91cf\u8868\u793a\uff0c\u901a\u8fc7\u76f8\u4f3c\u5ea6\u641c\u7d22\u8fdb\u884c\u68c0\u7d22\u3002Self-query retriever \u7684\u5173\u952e\u5728\u4e8e\u80fd\u591f\u4ece\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u4e2d\u63d0\u53d6\u51fa\u7528\u4e8e\u8fc7\u6ee4\u7684\u7ed3\u6784\u5316\u4fe1\u606f\uff0c\u5e76\u7ed3\u5408\u5411\u91cf\u76f8\u4f3c\u5ea6\u641c\u7d22\uff0c\u5b9e\u73b0\u66f4\u7cbe\u786e\u7684\u68c0\u7d22\u3002<\/p>\n

\u603b\u7ed3\u67e5\u8be2\u6784\u5efa\u9636\u6bb5\uff1a\u8fd9\u4e2a\u9636\u6bb5\u7684\u6838\u5fc3\u76ee\u6807\u662f\u5c06\u7528\u6237\u7528\u81ea\u7136\u8bed\u8a00\u63d0\u51fa\u7684\u95ee\u9898\u8f6c\u5316\u4e3a\u7cfb\u7edf\u53ef\u4ee5\u7406\u89e3\u548c\u6267\u884c\u7684\u67e5\u8be2\u8bed\u53e5\uff0c\u9488\u5bf9\u4e0d\u540c\u7c7b\u578b\u7684\u6570\u636e\u5b58\u50a8\uff08\u5173\u7cfb\u578b\u3001\u56fe\u3001\u5411\u91cf\uff09\uff0c\u91c7\u7528\u4e86\u4e0d\u540c\u7684\u67e5\u8be2\u8bed\u8a00\u548c\u6280\u672f\u3002\u4f53\u73b0\u4e86\u7cfb\u7edf\u5bf9\u591a\u6a21\u6001\u6570\u636e\u548c\u67e5\u8be2\u65b9\u5f0f\u7684\u652f\u6301\u3002<\/p>\n

\n2\u3001\u67e5\u8be2\u7ffb\u8bd1 (Query Translation)<\/h3>\n

\u5728\u67e5\u8be2\u6784\u5efa\u4e4b\u540e\uff0c\u6709\u65f6\u9700\u8981\u5bf9\u7528\u6237\u7684\u539f\u59cb\u67e5\u8be2\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u52a0\u5de5\u548c\u4f18\u5316\uff0c\u4ee5\u4fbf\u66f4\u6709\u6548\u5730\u68c0\u7d22\u548c\u7406\u89e3\u7528\u6237\u7684\u610f\u56fe\u3002\u56fe\u4e2d\u5c55\u793a\u4e86\u4e24\u79cd\u4e3b\u8981\u7684\u67e5\u8be2\u7ffb\u8bd1\u7b56\u7565\uff1a<\/p>\n

a. \u67e5\u8be2\u5206\u89e3 (Query Decomposition): \u00a0\u5bf9\u4e8e\u590d\u6742\u7684\u95ee\u9898\uff0c\u53ef\u4ee5\u5c06\u5176\u5206\u89e3\u4e3a\u66f4\u5c0f\u7684\u3001\u66f4\u6613\u4e8e\u5904\u7406\u7684\u5b50\u95ee\u9898 (Sub\/Step-back question(s))\u3002 \u8fd9\u53ef\u4ee5\u4f7f\u7528 Multi-query, Step-back, RAG-Fusion \u7b49\u6280\u672f\u3002 – Multi-query \u53ef\u80fd\u6307\u751f\u6210\u591a\u4e2a\u4e0d\u540c\u7684\u67e5\u8be2\uff0c\u4ece\u4e0d\u540c\u89d2\u5ea6\u63a2\u7d22\u95ee\u9898\u3002 – Step-back \u53ef\u80fd\u6307\u5148\u56de\u7b54\u4e00\u4e9b\u66f4\u57fa\u7840\u7684\u3001\u524d\u63d0\u6027\u7684\u95ee\u9898\uff0c\u7136\u540e\u518d\u9010\u6b65\u89e3\u51b3\u6700\u7ec8\u7684\u590d\u6742\u95ee\u9898\u3002 – RAG-Fusion \u53ef\u80fd\u662f\u6307\u5c06\u68c0\u7d22\u589e\u5f3a\u751f\u6210\u4e0e\u67e5\u8be2\u878d\u5408\u6280\u672f\u7ed3\u5408\uff0c\u901a\u8fc7\u591a\u6b21\u68c0\u7d22\u548c\u878d\u5408\uff0c\u66f4\u5168\u9762\u5730\u7406\u89e3\u7528\u6237\u610f\u56fe\u3002 – \u6838\u5fc3\u601d\u60f3\u662f Decompose or re-phrase the input question\uff0c\u5373\u5206\u89e3\u6216\u91cd\u8ff0\u8f93\u5165\u95ee\u9898\uff0c\u964d\u4f4e\u590d\u6742\u95ee\u9898\u7684\u5904\u7406\u96be\u5ea6\u3002
\nb. \u4f2a\u6587\u6863\u751f\u6210 (Pseudo-documents): \u00a0HyDE (Hypothetical Document Embeddings) \u662f\u4e00\u4e2a\u5178\u578b\u7684\u4f2a\u6587\u6863\u751f\u6210\u65b9\u6cd5\u3002\u5176\u601d\u60f3\u662f\uff0c\u5148\u8ba9\u6a21\u578b\u6839\u636e\u95ee\u9898\u751f\u6210\u4e00\u4e2a\u201c\u5047\u8bbe\u7684\u6587\u6863\u201d\uff08\u4f2a\u6587\u6863\uff09\uff0c\u8fd9\u4e2a\u4f2a\u6587\u6863\u5e76\u4e0d\u9700\u8981\u771f\u5b9e\u5b58\u5728\uff0c\u4f46\u5b83\u5e94\u8be5\u5305\u542b\u6a21\u578b\u5bf9\u95ee\u9898\u7b54\u6848\u7684\u521d\u6b65\u7406\u89e3\u548c\u9884\u6d4b\u3002\u7136\u540e\uff0c\u5c06\u4f2a\u6587\u6863\u548c\u771f\u5b9e\u6587\u6863\u4e00\u8d77\u8fdb\u884c\u5411\u91cf\u8868\u793a\uff0c\u5e76\u8fdb\u884c\u76f8\u4f3c\u5ea6\u68c0\u7d22\u3002HyDE \u65e8\u5728\u901a\u8fc7\u5f15\u5165\u6a21\u578b\u7684\u5148\u9a8c\u77e5\u8bc6\uff0c\u5e2e\u52a9\u5411\u91cf\u68c0\u7d22\u5668\u66f4\u597d\u5730\u627e\u5230\u76f8\u5173\u7684\u771f\u5b9e\u6587\u6863\u3002<\/p>\n

\u603b\u7ed3\u67e5\u8be2\u7ffb\u8bd1\u9636\u6bb5\uff1a\u8fd9\u4e2a\u9636\u6bb5\u65e8\u5728\u4f18\u5316\u7528\u6237\u67e5\u8be2\uff0c\u4f7f\u5176\u66f4\u9002\u5408\u540e\u7eed\u7684\u68c0\u7d22\u8fc7\u7a0b\u3002\u901a\u8fc7\u67e5\u8be2\u5206\u89e3\u53ef\u4ee5\u5904\u7406\u590d\u6742\u95ee\u9898\uff0c\u901a\u8fc7\u4f2a\u6587\u6863\u751f\u6210\u53ef\u4ee5\u63d0\u5347\u5411\u91cf\u68c0\u7d22\u7684\u51c6\u786e\u6027\uff0c\u4f53\u73b0\u4e86\u7cfb\u7edf\u5728\u7406\u89e3\u548c\u5904\u7406\u7528\u6237\u610f\u56fe\u4e0a\u7684\u7075\u6d3b\u6027\u548c\u667a\u80fd\u5316\u3002<\/p>\n

\n3\u3001\u8def\u7531 (Routing)<\/h3>\n

\u5f53\u7cfb\u7edf\u63a5\u6536\u5230\u7ffb\u8bd1\u540e\u7684\u67e5\u8be2\u540e\uff0c\u9700\u8981\u51b3\u5b9a\u5c06\u67e5\u8be2\u8def\u7531\u5230\u54ea\u4e2a\u6216\u54ea\u4e9b\u6570\u636e\u6e90\u8fdb\u884c\u68c0\u7d22\u3002\u56fe\u4e2d\u5c55\u793a\u4e86\u4e24\u79cd\u8def\u7531\u7b56\u7565\uff1a<\/p>\n

a. \u903b\u8f91\u8def\u7531 (Logical routing): \u00a0Let LLM choose DB based on the question\u3002\u8fd9\u610f\u5473\u7740\u4f7f\u7528\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u6765\u6839\u636e\u95ee\u9898\u7684\u5185\u5bb9\u548c\u7279\u70b9\uff0c\u5224\u65ad\u5e94\u8be5\u67e5\u8be2\u54ea\u4e2a\u6570\u636e\u5e93\u3002\u4f8b\u5982\uff0c\u5982\u679c\u95ee\u9898\u6d89\u53ca\u5230\u77e5\u8bc6\u56fe\u8c31\u76f8\u5173\u7684\u5b9e\u4f53\u548c\u5173\u7cfb\uff0c\u5219\u8def\u7531\u5230\u56fe\u6570\u636e\u5e93\uff1b\u5982\u679c\u95ee\u9898\u6d89\u53ca\u5230\u7ed3\u6784\u5316\u6570\u636e\u67e5\u8be2\uff0c\u5219\u8def\u7531\u5230\u5173\u7cfb\u578b\u6570\u636e\u5e93\uff1b\u5982\u679c\u95ee\u9898\u66f4\u504f\u5411\u4e8e\u8bed\u4e49\u641c\u7d22\uff0c\u5219\u8def\u7531\u5230\u5411\u91cf\u6570\u636e\u5e93\u3002
\nb. \u8bed\u4e49\u8def\u7531 (Semantic routing): \u00a0Embed question and choose prompt based on similarity\u3002\u8fd9\u79cd\u65b9\u5f0f\u9996\u5148\u5c06\u95ee\u9898\u8fdb\u884c\u5411\u91cf\u5d4c\u5165 (Embed)\uff0c\u7136\u540e\u57fa\u4e8e\u95ee\u9898\u5411\u91cf\u7684\u76f8\u4f3c\u5ea6\uff0c\u9009\u62e9\u4e0d\u540c\u7684Prompt (Prompt #1, Prompt #2)\u3002 \u8fd9\u610f\u5473\u7740\u9488\u5bf9\u4e0d\u540c\u7c7b\u578b\u7684\u95ee\u9898\u6216\u610f\u56fe\uff0c\u7cfb\u7edf\u9884\u8bbe\u4e86\u4e0d\u540c\u7684Prompt\u7b56\u7565\uff0c\u901a\u8fc7\u8bed\u4e49\u76f8\u4f3c\u5ea6\u6765\u81ea\u52a8\u9009\u62e9\u6700\u5408\u9002\u7684Prompt\uff0c\u4ee5\u5f15\u5bfc\u540e\u7eed\u7684\u68c0\u7d22\u6216\u751f\u6210\u8fc7\u7a0b\u3002<\/p>\n

\u603b\u7ed3\u8def\u7531\u9636\u6bb5\uff1a\u8def\u7531\u9636\u6bb5\u662f\u7cfb\u7edf\u667a\u80fd\u51b3\u7b56\u7684\u5173\u952e\u6b65\u9aa4\uff0c\u5b83\u6839\u636e\u95ee\u9898\u7684\u5185\u5bb9\u548c\u7279\u70b9\uff0c\u9009\u62e9\u6700\u5408\u9002\u7684\u6570\u636e\u6e90\u548c\u5904\u7406\u7b56\u7565\uff0c\u4f53\u73b0\u4e86\u7cfb\u7edf\u5728\u8d44\u6e90\u7ba1\u7406\u548c\u4efb\u52a1\u8c03\u5ea6\u4e0a\u7684\u667a\u80fd\u5316\u3002<\/p>\n

\n4\u3001\u7d22\u5f15 (Indexing)<\/h3>\n

\u4e3a\u4e86\u9ad8\u6548\u5730\u8fdb\u884c\u68c0\u7d22\uff0c\u6570\u636e\u9700\u8981\u9884\u5148\u8fdb\u884c\u7d22\u5f15\u3002\u56fe\u50cf\u7684\u84dd\u8272\u533a\u57df\u5c55\u793a\u4e86\u591a\u79cd\u7d22\u5f15\u4f18\u5316\u7684\u7b56\u7565\uff1a<\/p>\n

a. Chunk Optimization (\u5206\u5757\u4f18\u5316): \u00a0\u5728\u5904\u7406\u957f\u6587\u6863\u65f6\uff0c\u901a\u5e38\u9700\u8981\u5c06\u6587\u6863\u5206\u5272\u6210\u5757 (Chunk)\uff0c\u7136\u540e\u5bf9\u5757\u8fdb\u884c\u7d22\u5f15\u3002Chunk Optimization \u5173\u6ce8\u5982\u4f55\u66f4\u6709\u6548\u5730\u8fdb\u884c\u5206\u5757\u3002
\n– Split by Characters, Sections, Semantic Delimiters: \u4e0d\u540c\u7684\u5206\u5757\u7b56\u7565\uff0c\u4f8b\u5982\u6309\u5b57\u7b26\u6570\u3001\u7ae0\u8282\u3001\u6216\u8bed\u4e49\u5206\u9694\u7b26\u8fdb\u884c\u5206\u5272\u3002
\n– Semantic Splitter: \u00a0\u5f3a\u8c03\u8bed\u4e49\u5206\u5757\u7684\u91cd\u8981\u6027\uff0c\u4f18\u5316\u7528\u4e8eembedding\u7684chunk size\uff0c\u4f7f\u5f97\u6bcf\u4e2achunk\u5728\u8bed\u4e49\u4e0a\u66f4\u52a0\u5b8c\u6574\u548c\u72ec\u7acb\uff0c\u4ece\u800c\u63d0\u9ad8embedding\u7684\u8d28\u91cf\u548c\u68c0\u7d22\u6548\u679c\u3002<\/p>\n

b. Multi-representation indexing (\u591a\u91cd\u8868\u793a\u7d22\u5f15): \u00a0Summary -> {} -> Relational DB \/ Vectorstore\u3002\u8fd9\u8868\u793a\u53ef\u4ee5\u4e3a\u6587\u6863\u521b\u5efa\u591a\u79cd\u8868\u793a\u5f62\u5f0f\u8fdb\u884c\u7d22\u5f15\uff0c\u4f8b\u5982\uff0c\u9664\u4e86\u6587\u6863\u7684\u539f\u59cb\u6587\u672c\u5757\u5916\uff0c\u8fd8\u53ef\u4ee5\u751f\u6210\u6587\u6863\u7684\u6458\u8981 (Summary)\uff0c\u5e76\u5c06\u6458\u8981\u4e5f\u8fdb\u884c\u7d22\u5f15\u3002\u8fd9\u6837\u53ef\u4ee5\u4f7f\u7528\u4e0d\u540c\u7684\u8868\u793a\u5f62\u5f0f\u6765\u6ee1\u8db3\u4e0d\u540c\u7684\u67e5\u8be2\u9700\u6c42\u3002\u56fe\u4e2d\u6697\u793a\u6458\u8981\u53ef\u4ee5\u5b58\u50a8\u5728\u5173\u7cfb\u578b\u6570\u636e\u5e93\u6216\u5411\u91cf\u6570\u636e\u5e93\u4e2d\u3002
\n– Parent Document, Dense X: \u00a0\u53ef\u80fd\u6307\u5c06\u6587\u6863\u53ca\u5176\u7236\u6587\u6863\u4fe1\u606f\u4e00\u8d77\u7d22\u5f15\uff0c\u4ee5\u53ca\u4f7f\u7528Dense Representation (Dense X) \u6765\u8868\u793a\u6587\u6863\uff0cDense X \u53ef\u80fd\u6307\u7a20\u5bc6\u5411\u91cf\u8868\u793a\u3002
\n– Convert documents into compact retrieval<\/a> units (e.g., a summary): \u00a0\u5f3a\u8c03\u5c06\u6587\u6863\u8f6c\u5316\u4e3a\u66f4\u7d27\u51d1\u7684\u68c0\u7d22\u5355\u5143\uff0c\u4f8b\u5982\u6458\u8981\uff0c\u4ee5\u63d0\u9ad8\u68c0\u7d22\u6548\u7387\u3002<\/p>\n

c. Specialized Embeddings (\u4e13\u7528Embedding): \u00a0Fine-tuning, CoLBERT, [0, 1, … ] -> Vectorstore\u3002\u8fd9\u6307\u7684\u662f\u53ef\u4ee5\u4f7f\u7528\u4e13\u95e8\u8bad\u7ec3\u7684\u6216\u5fae\u8c03\u8fc7\u7684Embedding\u6a21\u578b\uff0c\u4f8b\u5982 CoLBERT\uff0c\u6765\u751f\u6210\u6587\u6863\u7684\u5411\u91cf\u8868\u793a\uff0c\u5e76\u5c06\u8fd9\u4e9b\u5411\u91cf\u5b58\u50a8\u5728\u5411\u91cf\u6570\u636e\u5e93\u4e2d\u3002
\n– Domain-specific and \/ or advanced embedding models: \u00a0\u5f3a\u8c03\u53ef\u4ee5\u4f7f\u7528\u9886\u57df\u7279\u5b9a\u7684\u6216\u66f4\u5148\u8fdb\u7684Embedding\u6a21\u578b\uff0c\u4ee5\u83b7\u5f97\u66f4\u7cbe\u51c6\u7684\u8bed\u4e49\u8868\u793a\uff0c\u63d0\u5347\u68c0\u7d22\u6548\u679c\u3002<\/p>\n

d. Heirarchical Indexing Summaries (\u5206\u5c42\u7d22\u5f15\u6458\u8981): \u00a0Splits -> cluser -> cluser -> … -> RAPTOR -> Graph DB\u3002 RAPTOR (\u53ef\u80fd\u662f\u6307\u4e00\u79cd\u5206\u5c42\u6587\u6863\u6458\u8981\u548c\u7d22\u5f15\u65b9\u6cd5) \u00a0\u901a\u8fc7\u591a\u5c42\u805a\u7c7b (cluser) \u7684\u65b9\u5f0f\uff0c\u6784\u5efa\u6587\u6863\u6458\u8981\u7684\u5c42\u6b21\u7ed3\u6784\u3002
\n– Tree of document summarization at various abstraction levels: \u00a0\u5f3a\u8c03 RAPTOR \u6784\u5efa\u7684\u662f\u4e00\u4e2a\u591a\u5c42\u62bd\u8c61\u7ea7\u522b\u7684\u6587\u6863\u6458\u8981\u6811\u3002
\n– \u5b58\u50a8\u5728\u56fe\u6570\u636e\u5e93 (Graph DB) \u4e2d\uff0c\u5229\u7528\u56fe\u6570\u636e\u5e93\u6765\u5b58\u50a8\u548c\u7ba1\u7406\u8fd9\u79cd\u5206\u5c42\u7d22\u5f15\u7ed3\u6784\uff0c\u65b9\u4fbf\u8fdb\u884c\u591a\u5c42\u6b21\u7684\u68c0\u7d22\u548c\u5bfc\u822a\u3002<\/p>\n

\u603b\u7ed3\u7d22\u5f15\u9636\u6bb5\uff1a\u7d22\u5f15\u9636\u6bb5\u5173\u6ce8\u5982\u4f55\u9ad8\u6548\u3001\u6709\u6548\u5730\u7ec4\u7ec7\u548c\u8868\u793a\u6570\u636e\uff0c\u4ee5\u4fbf\u8fdb\u884c\u5feb\u901f\u548c\u51c6\u786e\u7684\u68c0\u7d22\u3002\u4ece\u5206\u5757\u4f18\u5316\u3001\u591a\u91cd\u8868\u793a\u3001\u4e13\u7528Embedding\u5230\u5206\u5c42\u7d22\u5f15\u6458\u8981\uff0c\u4f53\u73b0\u4e86\u5728\u7d22\u5f15\u7b56\u7565\u4e0a\u7684\u591a\u6837\u6027\u548c\u5148\u8fdb\u6027\u3002<\/p>\n

\n5\u3001\u68c0\u7d22 (Retrieval)<\/h3>\n

\u57fa\u4e8e\u8def\u7531\u9009\u62e9\u7684\u6570\u636e\u6e90\u548c\u7d22\u5f15\uff0c\u7cfb\u7edf\u8fdb\u884c\u5b9e\u9645\u7684\u68c0\u7d22\u8fc7\u7a0b\u3002\u56fe\u50cf\u7684\u7eff\u8272\u533a\u57df\u5c55\u793a\u4e86\u68c0\u7d22\u7684\u4e24\u4e2a\u4e3b\u8981\u65b9\u9762\uff1a<\/p>\n

a. \u6392\u5e8f (Ranking): \u00a0Question -> {} -> Relevance<\/a> -> Filter\u3002\u68c0\u7d22\u5230\u7684\u6587\u6863\u9700\u8981\u6839\u636e\u5176\u4e0e\u67e5\u8be2\u7684\u76f8\u5173\u6027\u8fdb\u884c\u6392\u5e8f (Ranking)\u3002
\n– Re-Rank, RankGPT, RAG-Fusion: \u00a0\u63d0\u53ca\u4e86\u4e00\u4e9b\u9ad8\u7ea7\u7684\u6392\u5e8f\u6280\u672f\uff0c\u4f8b\u5982 Re-Rank (\u91cd\u6392\u5e8f\uff0c\u5728\u521d\u59cb\u68c0\u7d22\u7ed3\u679c\u57fa\u7840\u4e0a\u8fdb\u884c\u66f4\u7cbe\u7ec6\u7684\u6392\u5e8f)\u3001RankGPT (\u7528GPT\u7b49\u5927\u578b\u6a21\u578b\u8fdb\u884c\u6392\u5e8f)\u3001RAG-Fusion (\u5c06\u6392\u5e8f\u4e0e\u68c0\u7d22\u589e\u5f3a\u751f\u6210\u878d\u5408)\u3002
\n– Rank or filter \/ compress documents based on relevance: \u00a0\u6392\u5e8f\u7684\u76ee\u7684\u53ef\u4ee5\u662f\u76f4\u63a5\u6392\u5e8f\u8fd4\u56de\u6700\u76f8\u5173\u7684\u6587\u6863\uff0c\u4e5f\u53ef\u4ee5\u662f\u57fa\u4e8e\u76f8\u5173\u6027\u8fdb\u884c\u8fc7\u6ee4 (Filter) \u6216\u538b\u7f29 (compress) \u6587\u6863\uff0c\u4ee5\u4fbf\u540e\u7eed\u5904\u7406\u3002 – CRAG (\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u68c0\u7d22\u589e\u5f3a\u751f\u6210\u65b9\u6cd5\uff0cContext-Relevant Answer Generation) \u5728\u6392\u5e8f\u73af\u8282\u4e5f\u51fa\u73b0\uff0c\u6392\u5e8f\u8fc7\u7a0b\u4e5f\u9700\u8981\u8003\u8651\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002
\nb. \u4e3b\u52a8\u68c0\u7d22 (Active retrieval): \u00a0{} -> CRAG -> Answer\u3002Re-retrieve and \/ or retrieve from new data sources (e.g., web) if retrieved documents are not relevant. \u00a0\u4e3b\u52a8\u68c0\u7d22\u6307\u7684\u662f\uff0c\u5982\u679c\u521d\u59cb\u68c0\u7d22\u7ed3\u679c\u4e0d\u7406\u60f3\uff0c\u7cfb\u7edf\u53ef\u4ee5\u4e3b\u52a8\u8fdb\u884c\u91cd\u65b0\u68c0\u7d22 (Re-retrieve) \u6216\u8005\u4ece\u65b0\u7684\u6570\u636e\u6e90 (\u4f8b\u5982 Web) \u8fdb\u884c\u68c0\u7d22\u3002
\n– CRAG \u5728\u4e3b\u52a8\u68c0\u7d22\u4e2d\u4e5f\u51fa\u73b0\uff0c\u8fdb\u4e00\u6b65\u5f3a\u8c03\u4e86\u4e0a\u4e0b\u6587\u76f8\u5173\u6027\u548c\u8fed\u4ee3\u68c0\u7d22\u7684\u91cd\u8981\u6027\u3002
\n– Self-RAG, RRR (Retrieval-Rewrite-Read) \u7b49\u6280\u672f\u53ef\u80fd\u4e5f\u4e0e\u4e3b\u52a8\u68c0\u7d22\u76f8\u5173\uff0c\u65e8\u5728\u901a\u8fc7\u8fed\u4ee3\u7684\u68c0\u7d22\u548c\u751f\u6210\u8fc7\u7a0b\uff0c\u4e0d\u65ad\u4f18\u5316\u68c0\u7d22\u7ed3\u679c\u548c\u7b54\u6848\u8d28\u91cf\u3002<\/p>\n

\u603b\u7ed3\u68c0\u7d22\u9636\u6bb5\uff1a\u68c0\u7d22\u9636\u6bb5\u7684\u6838\u5fc3\u76ee\u6807\u662f\u627e\u5230\u4e0e\u7528\u6237\u67e5\u8be2\u6700\u76f8\u5173\u7684\u6587\u6863\u6216\u4fe1\u606f\u3002\u4ece\u6392\u5e8f\u5230\u4e3b\u52a8\u68c0\u7d22\uff0c\u4f53\u73b0\u4e86\u7cfb\u7edf\u5728\u68c0\u7d22\u7b56\u7565\u4e0a\u7684\u7cbe\u7ec6\u5316\u548c\u667a\u80fd\u5316\uff0c\u529b\u6c42\u63d0\u4f9b\u9ad8\u8d28\u91cf\u7684\u68c0\u7d22\u7ed3\u679c\u3002<\/p>\n

\n6\u3001\u751f\u6210 (Generation)<\/h3>\n

\u6700\u7ec8\uff0c\u7cfb\u7edf\u9700\u8981\u6839\u636e\u68c0\u7d22\u5230\u7684\u6587\u6863\u751f\u6210\u7b54\u6848\uff0c\u5e76\u5448\u73b0\u7ed9\u7528\u6237\u3002\u56fe\u50cf\u7684\u7d2b\u8272\u533a\u57df\u5c55\u793a\u4e86\u751f\u6210\u9636\u6bb5\u7684\u6838\u5fc3\u6280\u672f\uff1a<\/p>\n

a. \u4e3b\u52a8\u68c0\u7d22 (Active retrieval) (\u518d\u6b21\u51fa\u73b0): \u00a0{} -> Answer -> Self-RAG, RRR -> Question re-writing and \/ or re-retrieval of documents\u3002\u4e3b\u52a8\u68c0\u7d22\u5728\u751f\u6210\u9636\u6bb5\u4e5f\u626e\u6f14\u91cd\u8981\u89d2\u8272\u3002
\n– Self-RAG (Self-Retrieval Augmented Generation) \u662f\u4e00\u79cd\u81ea\u68c0\u7d22\u589e\u5f3a\u751f\u6210\u65b9\u6cd5\uff0c\u5b83\u5141\u8bb8\u751f\u6210\u6a21\u578b\u5728\u751f\u6210\u7b54\u6848\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6839\u636e\u9700\u8981\u4e3b\u52a8\u8fdb\u884c\u68c0\u7d22\uff0c\u5e76\u6839\u636e\u68c0\u7d22\u7ed3\u679c\u8c03\u6574\u751f\u6210\u7b56\u7565\u3002 – RRR (Retrieval-Rewrite-Read) \u662f\u4e00\u79cd\u8fed\u4ee3\u7684\u751f\u6210\u6d41\u7a0b\uff0c\u53ef\u80fd\u5305\u62ec\u68c0\u7d22 (Retrieval)\u3001\u91cd\u5199\u95ee\u9898 (Rewrite) \u548c\u9605\u8bfb\u6587\u6863 (Read) \u7b49\u6b65\u9aa4\uff0c\u901a\u8fc7\u591a\u6b21\u8fed\u4ee3\u4f18\u5316\u7b54\u6848\u8d28\u91cf\u3002
\n– Use generation quality to inform question re-writing and \/ or re-retrieval of documents: \u00a0\u5f3a\u8c03\u53ef\u4ee5\u5229\u7528\u751f\u6210\u7b54\u6848\u7684\u8d28\u91cf\u6765\u6307\u5bfc\u95ee\u9898\u91cd\u5199\u548c\u6587\u6863\u91cd\u68c0\u7d22\uff0c\u5f62\u6210\u4e00\u4e2a\u95ed\u73af\u7684\u4f18\u5316\u8fc7\u7a0b\u3002<\/p>\n

\u603b\u7ed3\u751f\u6210\u9636\u6bb5\uff1a\u751f\u6210\u9636\u6bb5\u662f\u6700\u7ec8\u8f93\u51fa\u7b54\u6848\u7684\u5173\u952e\u6b65\u9aa4\u3002\u4e3b\u52a8\u68c0\u7d22\u548c\u81ea\u68c0\u7d22\u589e\u5f3a\u751f\u6210\u6280\u672f (Self-RAG, RRR) \u00a0\u4f7f\u5f97\u751f\u6210\u8fc7\u7a0b\u66f4\u52a0\u667a\u80fd\u5316\u548c\u53ef\u63a7\uff0c\u80fd\u591f\u751f\u6210\u66f4\u51c6\u786e\u3001\u66f4\u7b26\u5408\u7528\u6237\u9700\u6c42\u7b54\u6848\u3002<\/p>\n

\u6574\u4f53\u603b\u7ed3: \u8fd9\u5f20\u56fe\u6e05\u6670\u5730\u5c55\u793a\u4e86\u4e00\u4e2a\u73b0\u4ee3RAG\u7cfb\u7edf\u7684\u590d\u6742\u6027\u548c\u7cbe\u7ec6\u7a0b\u5ea6\u3002\u5b83\u6db5\u76d6\u4e86\u4ece\u67e5\u8be2\u7406\u89e3\u3001\u6570\u636e\u8def\u7531\u3001\u7d22\u5f15\u4f18\u5316\u3001\u9ad8\u6548\u68c0\u7d22\u5230\u6700\u7ec8\u7b54\u6848\u751f\u6210\u7684\u5b8c\u6574\u6d41\u7a0b\uff0c\u5e76\u5c55\u793a\u4e86\u5728\u6bcf\u4e2a\u73af\u8282\u53ef\u4ee5\u91c7\u7528\u7684\u591a\u79cd\u5148\u8fdb\u6280\u672f\u548c\u7b56\u7565\u3002<\/p>\n

\u5173\u952e\u4eae\u70b9\u548c\u8d8b\u52bf:<\/p>\n

    \n
  • \u591a\u6570\u636e\u5e93\u652f\u6301: \u7cfb\u7edf\u652f\u6301\u5173\u7cfb\u578b\u6570\u636e\u5e93\u3001\u56fe\u6570\u636e\u5e93\u548c\u5411\u91cf\u6570\u636e\u5e93\uff0c\u80fd\u591f\u5904\u7406\u4e0d\u540c\u7c7b\u578b\u7684\u6570\u636e\u548c\u67e5\u8be2\u9700\u6c42\u3002<\/li>\n
  • \u67e5\u8be2\u4f18\u5316\u548c\u7ffb\u8bd1: \u00a0\u901a\u8fc7\u67e5\u8be2\u5206\u89e3\u548c\u4f2a\u6587\u6863\u751f\u6210\u7b49\u6280\u672f\uff0c\u63d0\u5347\u7cfb\u7edf\u5bf9\u590d\u6742\u548c\u8bed\u4e49\u5316\u67e5\u8be2\u7684\u5904\u7406\u80fd\u529b\u3002<\/li>\n
  • \u667a\u80fd\u8def\u7531: \u00a0\u5229\u7528LLM\u548c\u8bed\u4e49\u76f8\u4f3c\u5ea6\u8fdb\u884c\u8def\u7531\u51b3\u7b56\uff0c\u5b9e\u73b0\u6570\u636e\u6e90\u7684\u667a\u80fd\u9009\u62e9\u548c\u4efb\u52a1\u8c03\u5ea6\u3002<\/li>\n
  • \u7d22\u5f15\u4f18\u5316\u591a\u6837\u6027: \u00a0\u4ece\u5206\u5757\u3001\u591a\u91cd\u8868\u793a\u3001\u4e13\u7528Embedding\u5230\u5206\u5c42\u7d22\u5f15\u6458\u8981\uff0c\u4f53\u73b0\u4e86\u7d22\u5f15\u7b56\u7565\u7684\u591a\u6837\u6027\u548c\u6df1\u5ea6\u4f18\u5316\u3002<\/li>\n
  • \u68c0\u7d22\u7684\u7cbe\u7ec6\u5316\u548c\u4e3b\u52a8\u6027: \u00a0\u4ece\u6392\u5e8f\u7b97\u6cd5\u5230\u4e3b\u52a8\u68c0\u7d22\uff0c\u7cfb\u7edf\u529b\u6c42\u63d0\u4f9b\u9ad8\u8d28\u91cf\u3001\u76f8\u5173\u7684\u68c0\u7d22\u7ed3\u679c\u3002<\/li>\n
  • \u751f\u6210\u4e0e\u68c0\u7d22\u7684\u6df1\u5ea6\u878d\u5408: \u00a0Self-RAG, RRR \u7b49\u6280\u672f\u8868\u660e\uff0c\u751f\u6210\u9636\u6bb5\u4e0d\u518d\u662f\u7b80\u5355\u7684\u4fe1\u606f\u62fc\u63a5\uff0c\u800c\u662f\u4e0e\u68c0\u7d22\u8fc7\u7a0b\u6df1\u5ea6\u878d\u5408\uff0c\u5f62\u6210\u8fed\u4ee3\u4f18\u5316\u7684\u95ed\u73af\u3002<\/li>\n<\/ul>\n

    \u8fd9\u5f20\u56fe\u4ee3\u8868\u4e86\u5f53\u524dRAG\u7cfb\u7edf\u53d1\u5c55\u7684\u4e00\u4e2a\u91cd\u8981\u8d8b\u52bf\uff0c\u5373\u66f4\u52a0\u6ce8\u91cd\u7cfb\u7edf\u7684\u667a\u80fd\u5316\u3001\u6a21\u5757\u5316\u548c\u53ef\u6269\u5c55\u6027\u3002\u672a\u6765\u7684RAG\u7cfb\u7edf\u5c06\u4e0d\u4ec5\u4ec5\u662f\u7b80\u5355\u7684\u201c\u68c0\u7d22+\u751f\u6210\u201d\uff0c\u800c\u662f\u4f1a\u671d\u7740\u66f4\u52a0\u667a\u80fd\u5316\u7684\u65b9\u5411\u53d1\u5c55\uff0c\u80fd\u591f\u66f4\u597d\u5730\u7406\u89e3\u7528\u6237\u610f\u56fe\uff0c\u66f4\u6709\u6548\u5730\u5229\u7528\u591a\u6a21\u6001\u6570\u636e\uff0c\u66f4\u7cbe\u51c6\u5730\u8fdb\u884c\u68c0\u7d22\u548c\u751f\u6210\uff0c\u5e76\u6700\u7ec8\u63d0\u4f9b\u66f4\u4f18\u8d28\u3001\u66f4\u4e2a\u6027\u5316\u7684\u7528\u6237\u4f53\u9a8c\u3002\u8fd9\u5f20\u56fe\u4e3a\u6211\u4eec\u7406\u89e3\u548c\u6784\u5efa\u4e0b\u4e00\u4ee3RAG\u7cfb\u7edf\u63d0\u4f9b\u4e86\u975e\u5e38\u6709\u4ef7\u503c\u7684\u53c2\u8003\u6846\u67b6\u3002<\/p>\n

    \u53c2\u8003\u6587\u732e\uff1a
    \n[1] GitHub\uff1ahttps:\/\/github.com\/bRAGAI\/bRAG-langchain\/
    \n[2] https:\/\/bragai.dev\/<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u8fd9\u5f20\u56fe\u6e05\u6670\u5730\u63cf\u7ed8\u4e86\u4e00\u4e2a\u73b0\u4ee3\u5316\u7684\u3001\u590d\u6742\u7684\u95ee\u9898\u89e3\u7b54\uff08Question Answering, QA\uff09\u6216\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08Retrieval-Augmented Generation, RAG\uff09\u7cfb\u7edf\u7684\u67b6\u6784\u84dd\u56fe\u3002\u5b83\u4ece\u7528\u6237\u63d0\u51fa\u95ee\u9898\u5f00\u59cb\uff0c\u4e00\u76f4\u5230\u6700\u7ec8\u751f\u6210\u7b54\u6848\uff0c…<\/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-26952","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/26952","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=26952"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/26952\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=26952"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=26952"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=26952"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}