\n\n| \n #<\/p>\n<\/td>\n | \n \u6a21\u5757<\/p>\n<\/td>\n | \n \u6838\u5fc3\u804c\u8d23<\/p>\n<\/td>\n | \n \u5e38\u89c1\u6280\u672f\u9009\u578b<\/p>\n<\/td>\n | \n \u505a\u7838\u4e86\u7684\u5178\u578b\u75c7\u72b6<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2460<\/p>\n<\/td>\n | \n \u91c7\u96c6\u4e0e\u540c\u6b65<\/p>\n<\/td>\n | \n \u628a\u5404\u5904\u6587\u6863\u5b9e\u65f6\/\u5b9a\u65f6\u62c9\u8fdb\u6765，\u68c0\u6d4b\u53d8\u66f4<\/p>\n<\/td>\n | \n Kafka \/ \u5b9a\u65f6\u4efb\u52a1 \/ Webhook<\/p>\n<\/td>\n | \n \u77e5\u8bc6\u5e93\u6c38\u8fdc\u662f\u65e7\u6570\u636e，\u7528\u6237\u95ee\u5230\u4e0a\u5468\u7684\u6587\u6863\u5b83\u4e0d\u77e5\u9053<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2461<\/p>\n<\/td>\n | \n \u7ed3\u6784\u5316\u5207\u5206<\/p>\n<\/td>\n | \n \u628a\u6587\u6863\u5207\u6210\u68c0\u7d22\u53cb\u597d\u7684\u5757，\u5e26\u4e0a\u5143\u6570\u636e\u548c\u6743\u9650<\/p>\n<\/td>\n | \n \u81ea\u7814\u5207\u5206\u5668 \/ LangChain Splitter<\/p>\n<\/td>\n | \n \u7b54\u975e\u6240\u95ee\u3001\u8868\u683c\u7c7b\u95ee\u9898\u5168\u519b\u8986\u6ca1<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2462<\/p>\n<\/td>\n | \n \u5d4c\u5165\u4e0e\u7d22\u5f15<\/p>\n<\/td>\n | \n \u6587\u672c\u8f6c\u5411\u91cf，\u5efa\u5411\u91cf+\u5173\u952e\u8bcd\u53cc\u8def\u7d22\u5f15<\/p>\n<\/td>\n | \n BGE-M3 \/ text-embedding-3 \/ ES<\/p>\n<\/td>\n | \n \u6362\u4e2a\u8bf4\u6cd5\u5c31\u641c\u4e0d\u5230；\u4e13\u6709\u540d\u8bcd\u641c\u4e0d\u5230<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2463<\/p>\n<\/td>\n | \n \u67e5\u8be2\u7406\u89e3<\/p>\n<\/td>\n | \n \u628a\u7528\u6237\u7684\u53e3\u8bed\u5316\u63d0\u95ee\u53d8\u6210\u53ef\u68c0\u7d22\u7684\u67e5\u8be2<\/p>\n<\/td>\n | \n \u5c0f\u6a21\u578b\u6539\u5199 \/ \u89c4\u5219<\/p>\n<\/td>\n | \n \u591a\u8f6e\u5bf9\u8bdd\u5168\u5d29（"\u90a3\u5b83\u591a\u5c11\u94b1？"\u641c\u4e0d\u5230\u4e1c\u897f）<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2464<\/p>\n<\/td>\n | \n \u6df7\u5408\u68c0\u7d22<\/p>\n<\/td>\n | \n \u5411\u91cf+BM25 \u53cc\u8def\u53ec\u56de，RRF \u878d\u5408<\/p>\n<\/td>\n | \n pgvector \/ Milvus \/ Qdrant + ES<\/p>\n<\/td>\n | \n \u8981\u4e48\u6982\u5ff5\u95ee\u9898\u641c\u4e0d\u5230，\u8981\u4e48\u7cbe\u786e\u8bcd\u641c\u4e0d\u51c6<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2465<\/p>\n<\/td>\n | \n \u91cd\u6392\u5e8f<\/p>\n<\/td>\n | \n \u7528\u91cd\u6a21\u578b\u7ed9\u5019\u9009\u7cbe\u6392<\/p>\n<\/td>\n | \n bge-reranker \/ Cohere Rerank<\/p>\n<\/td>\n | \n \u5173\u952e\u5757\u5728 Top5 \u4e4b\u5916，\u6a21\u578b"\u770b\u5230\u4e86\u4f46\u6ca1\u7528\u4e0a"<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2466<\/p>\n<\/td>\n | \n \u4e0a\u4e0b\u6587\u7ec4\u88c5<\/p>\n<\/td>\n | \n \u538b\u7f29\u3001\u53bb\u91cd\u3001\u6392\u5e8f\u3001\u6743\u9650\u6821\u9a8c<\/p>\n<\/td>\n | \n \u5c0f\u6a21\u578b\u62bd\u53d6 \/ LLMLingua<\/p>\n<\/td>\n | \n Token \u7206\u70b8\u6210\u672c\u9ad8；\u65e0\u5173\u5185\u5bb9\u6c61\u67d3\u7b54\u6848<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2467<\/p>\n<\/td>\n | \n \u751f\u6210\u4e0e\u6821\u9a8c<\/p>\n<\/td>\n | \n Prompt \u6a21\u677f\u3001\u5f15\u7528\u6807\u6ce8\u3001\u8f93\u51fa\u6821\u9a8c<\/p>\n<\/td>\n | \n GPT \/ Claude \/ Qwen \/ DeepSeek<\/p>\n<\/td>\n | \n \u5e7b\u89c9\u3001\u5570\u55e6\u3001\u683c\u5f0f\u4e0d\u7a33\u5b9a<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2468<\/p>\n<\/td>\n | \n \u8bc4\u6d4b\u4f53\u7cfb<\/p>\n<\/td>\n | \n \u79bb\u7ebf\u8bc4\u6d4b\u96c6 + \u5728\u7ebf\u53cd\u9988\u95ed\u73af<\/p>\n<\/td>\n | \n RAGAS \/ \u81ea\u7814<\/p>\n<\/td>\n | \n \u6539\u4ec0\u4e48\u90fd\u9760\u611f\u89c9，\u6c38\u8fdc\u4e0d\u77e5\u9053\u81ea\u5df1\u53d8\u597d\u8fd8\u662f\u53d8\u5dee<\/p>\n<\/td>\n<\/tr>\n |
\n| \n \u2469<\/p>\n<\/td>\n | \n \u53ef\u89c2\u6d4b\u6027<\/p>\n<\/td>\n | \n \u5168\u94fe\u8def Trace\u3001\u6210\u672c\u4e0e\u8d28\u91cf\u76d1\u63a7<\/p>\n<\/td>\n | \n Langfuse \/ \u81ea\u5efa<\/p>\n<\/td>\n | \n \u51fa\u4e86\u95ee\u9898\u5f00\u76f2\u76d2\u6392\u67e5<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u4e13\u680f\u540e\u7eed\u6bcf\u4e00\u671f，\u5c31\u6309\u8fd9\u5f20\u8868\u9010\u4e2a\u6a21\u5757\u6df1\u6316（\u9884\u544a\u89c1 0.7 \u8282）\u3002\u672c\u671f\u5148\u628a \u2460~\u2467 \u7684\u5173\u952e\u8bbe\u8ba1\u51b3\u7b56\u8bb2\u6e05\u695a，\u8ba9\u4f60\u6709\u4e00\u5f20\u80fd\u6307\u5bfc\u5b9e\u6218\u7684\u5730\u56fe\u3002<\/p>\n 0.4 \u6570\u636e\u94fe\u8def\u8be6\u89e3：\u4ece\u539f\u59cb\u6587\u6863\u5230\u53ef\u68c0\u7d22\u7684\u77e5\u8bc6<\/span><\/span><\/h3>\n\u79bb\u7ebf\u94fe\u8def\u7684\u76ee\u6807\u53ea\u6709\u4e00\u53e5\u8bdd：\u628a"\u4eba\u5199\u7684\u6587\u6863"\u8f6c\u6362\u6210"\u673a\u5668\u53ef\u68c0\u7d22\u7684\u77e5\u8bc6\u5355\u5143"，\u5e76\u4e14\u6bcf\u4e2a\u77e5\u8bc6\u5355\u5143\u90fd\u5b8c\u6574\u3001\u53ef\u6eaf\u6e90\u3001\u5e26\u6743\u9650\u3002<\/p>\n 0.4.1 \u91c7\u96c6\u4e0e\u540c\u6b65：\u88ab\u4f4e\u4f30\u7684\u6a21\u5757<\/span><\/span><\/h4>\nDemo \u91cc\u6ca1\u6709\u8fd9\u4e2a\u6a21\u5757\u2014\u2014\u4f60\u624b\u52a8\u628a\u4e00\u4efd PDF \u5582\u8fdb\u53bb\u3002\u751f\u4ea7\u91cc\u8fd9\u662f\u7b2c\u4e00\u4e2a\u5751：\u77e5\u8bc6\u662f\u6d3b\u7684\u3002\u6587\u6863\u4f1a\u66f4\u65b0\u3001\u4f1a\u4f5c\u5e9f\u3001\u4f1a\u8fc1\u79fb；\u4eca\u5929\u5efa\u597d\u7684\u7d22\u5f15，\u4e24\u5468\u540e\u5c31\u662f\u8bef\u5bfc\u7528\u6237\u7684\u8fc7\u671f\u77e5\u8bc6\u3002<\/p>\n \u540c\u6b65\u673a\u5236\u7684\u6838\u5fc3\u8bbe\u8ba1\u662f\u57fa\u4e8e\u5185\u5bb9\u54c8\u5e0c\u7684\u589e\u91cf\u68c0\u6d4b：<\/p>\n \u6587\u6863\u66f4\u65b0\u4e8b\u4ef6\u6d41：<\/span><\/p>\n\u00a0\u00a0\u6587\u6863\u7cfb\u7edf \u2500\u2500(Webhook\/\u8f6e\u8be2)\u2500\u2500\u2192 \u540c\u6b65\u670d\u52a1<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u251c\u2500 \u8ba1\u7b97\u65b0\u5185\u5bb9 SHA-256<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u251c\u2500 \u4e0e\u65e7\u54c8\u5e0c\u5bf9\u6bd4<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u25bc \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u25bc \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u25bc<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u65e0\u53d8\u5316 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u5185\u5bb9\u53d8\u4e86 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u6587\u6863\u5220\u9664\u4e86<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u7ed3\u675f \u00a0\u00a0\u00a0\u00a0\u00a0\u5220\u65e7\u5757 \u2192 \u91cd\u65b0 \u00a0\u00a0\u00a0\u00a0\u00a0\u5220\u6240\u6709\u5757<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u5207\u5757\/\u5d4c\u5165\/\u5165\u5e93 \u00a0\u00a0\u00a0\u00a0+ \u6807\u8bb0\u6e90\u6587\u6863\u5220\u9664<\/span><\/p>\n\u4e09\u4e2a\u5fc5\u987b\u505a\u5bf9\u7684\u5730\u65b9：<\/p>\n \u5220\u9664\u8981\u5148\u4e8e\u91cd\u5efa\u3002\u6587\u6863\u66f4\u65b0\u65f6，\u5148\u5220\u6389\u5b83\u7684\u6240\u6709\u65e7\u5757\u518d\u5199\u65b0\u5757，\u5426\u5219\u65b0\u65e7\u7248\u672c\u5e76\u5b58，\u68c0\u7d22\u7ed3\u679c\u91cc\u540c\u4e00\u4e2a\u95ee\u9898\u6709\u4e24\u4e2a\u77db\u76fe\u7684"\u6807\u51c6\u7b54\u6848"\u3002<\/li>\n\u5757 ID \u8981\u7a33\u5b9a\u4e14\u5e26\u7248\u672c\u4fe1\u606f\u3002\u63a8\u8350\u683c\u5f0f {doc_id}:{chunk_seq}<\/span>，doc_id \u7528\u5185\u5bb9\u54c8\u5e0c\u6216\u6587\u6863\u7cfb\u7edf ID，\u4fdd\u8bc1\u5e42\u7b49\u91cd\u8dd1\u3002<\/li>\n\u540c\u6b65\u8981\u5f02\u6b65\u5316\u3002\u5927\u6587\u6863\u7684\u89e3\u6790\u548c\u5d4c\u5165\u5f88\u8017\u65f6（\u4e00\u4efd 100 \u9875 PDF \u5d4c\u5165\u53ef\u80fd\u8981\u51e0\u5341\u79d2），\u540c\u6b65\u63a5\u53e3\u4f1a\u88ab\u62d6\u6b7b\u3002\u751f\u4ea7\u4e0a\u7528\u6d88\u606f\u961f\u5217（Kafka topic: doc-events<\/span>），\u540c\u6b65\u670d\u52a1\u53ea\u8d1f\u8d23"\u58f0\u660e\u53d8\u66f4"，\u6d88\u8d39\u8005\u8d1f\u8d23\u91cd\u5efa\u7d22\u5f15\u3002<\/li>\n0.4.2 \u683c\u5f0f\u89e3\u6790：RAG \u8d28\u91cf\u7684\u7b2c\u4e00\u9053\u5206\u6c34\u5cad<\/span><\/span><\/h4>\n"\u5783\u573e\u8fdb，\u5783\u573e\u51fa"\u5728 RAG \u91cc\u6700\u5178\u578b\u7684\u4f53\u73b0\u5c31\u662f\u89e3\u6790\u73af\u8282\u3002\u4e0d\u540c\u683c\u5f0f\u7684\u89e3\u6790\u8d28\u91cf\u5929\u5dee\u5730\u522b：<\/p>\n \n\n\n| \n \u683c\u5f0f<\/p>\n<\/td>\n | \n \u89e3\u6790\u96be\u5ea6<\/p>\n<\/td>\n | \n \u5e38\u89c1\u5751<\/p>\n<\/td>\n | \n \u63a8\u8350\u65b9\u6848<\/p>\n<\/td>\n<\/tr>\n | \n| \n Markdown<\/p>\n<\/td>\n | \n \u2605<\/p>\n<\/td>\n | \n \u57fa\u672c\u6ca1\u6709<\/p>\n<\/td>\n | \n \u76f4\u63a5\u6309\u6807\u9898\u7ed3\u6784\u5207<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u7eaf\u6587\u672c<\/p>\n<\/td>\n | \n \u2605<\/p>\n<\/td>\n | \n \u7f16\u7801\u3001\u6362\u884c\u6df7\u4e71<\/p>\n<\/td>\n | \n \u6e05\u6d17 + \u6bb5\u843d\u91cd\u7ec4<\/p>\n<\/td>\n<\/tr>\n | \n| \n HTML<\/p>\n<\/td>\n | \n \u2605\u2605<\/p>\n<\/td>\n | \n \u5bfc\u822a\u680f\/\u9875\u7709\u9875\u811a\u566a\u97f3<\/p>\n<\/td>\n | \n trafilatura \/ readability<\/p>\n<\/td>\n<\/tr>\n | \n| \n Word\/DOCX<\/p>\n<\/td>\n | \n \u2605\u2605<\/p>\n<\/td>\n | \n \u8868\u683c\u53d8\u4e71\u7801\u3001\u6279\u6ce8\u6df7\u5165<\/p>\n<\/td>\n | \n python-docx \/ Unstructured<\/p>\n<\/td>\n<\/tr>\n | \n| \n PDF<\/p>\n<\/td>\n | \n \u2605\u2605\u2605\u2605<\/p>\n<\/td>\n | \n \u53cc\u680f\u6392\u7248\u4e71\u5e8f\u3001\u8868\u683c\u788e\u7247\u5316\u3001\u626b\u63cf\u4ef6\u65e0\u6587\u672c<\/p>\n<\/td>\n | \n \u7248\u9762\u5206\u6790\u6a21\u578b（\u5982 MinerU \/ PaddleOCR）<\/p>\n<\/td>\n<\/tr>\n | \n| \n Excel<\/p>\n<\/td>\n | \n \u2605\u2605\u2605<\/p>\n<\/td>\n | \n \u591a\u7ea7\u8868\u5934\u3001\u5408\u5e76\u5355\u5143\u683c<\/p>\n<\/td>\n | \n \u6574\u8868\u8f6c\u4e3a Markdown \u8868\u683c\u4fdd\u7559<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u56fe\u7247<\/p>\n<\/td>\n | \n \u2605\u2605\u2605\u2605\u2605<\/p>\n<\/td>\n | \n \u65e0\u6587\u672c，\u7eaf\u89c6\u89c9\u4fe1\u606f<\/p>\n<\/td>\n | \n OCR + \u591a\u6a21\u6001\u6a21\u578b\u751f\u6210\u63cf\u8ff0<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n PDF \u662f\u91cd\u707e\u533a，\u503c\u5f97\u5355\u72ec\u5f3a\u8c03\u3002\u00a0\u4f20\u7edf PDF \u89e3\u6790\u5e93（PyPDF2 \u4e4b\u7c7b）\u6309"\u6587\u672c\u6d41"\u62bd\u53d6，\u9047\u5230\u53cc\u680f\u8bba\u6587\u4f1a\u628a\u5de6\u53f3\u680f\u4ea4\u53c9\u8bfb\u53d6，\u51fa\u6765\u7684\u6587\u672c\u5b8c\u5168\u662f\u4e71\u7684\u2014\u2014\u800c\u8fd9\u79cd\u4e71\u6587\u672c\u88ab\u5207\u5757\u3001\u5d4c\u5165\u540e，\u68c0\u7d22\u8d28\u91cf\u5fc5\u7136\u5d29\u574f\u3002\u6b63\u786e\u59ff\u52bf\u662f\u7528\u7248\u9762\u5206\u6790（Layout Analysis）：\u5148\u8bc6\u522b\u51fa\u6807\u9898\u3001\u6b63\u6587\u3001\u8868\u683c\u3001\u56fe\u7247\u7684\u533a\u57df\u548c\u9605\u8bfb\u987a\u5e8f，\u518d\u6309\u533a\u57df\u62bd\u53d6\u3002\u5f00\u6e90\u65b9\u6848\u5982 MinerU\u3001PaddleOCR \u90fd\u53ef\u4ee5\u505a\u5230\u3002<\/p>\n \u8868\u683c\u8981\u5355\u72ec\u5904\u7406：\u8868\u683c\u6c38\u8fdc\u6574\u5757\u4fdd\u7559，\u4e0d\u53c2\u4e0e\u666e\u901a\u5206\u5757，\u5e76\u628a\u8868\u5934\u62fc\u8fdb\u6bcf\u4e2a\u8868\u683c\u5757（\u56e0\u4e3a\u5d4c\u5165\u6a21\u578b\u770b\u5230 120 980 340<\/span>\u00a0\u8fd9\u4e32\u6570\u5b57\u6beb\u65e0\u8bed\u4e49，\u770b\u5230 \u4ea7\u54c1A | Q1\u9500\u552e\u989d: 120<\/span>\u00a0\u624d\u6709\u610f\u4e49）\u3002<\/p>\n0.4.3 \u7ed3\u6784\u5316\u5207\u5206：\u51b3\u5b9a\u68c0\u7d22\u4e0a\u9650\u7684\u4e00\u6b65<\/span><\/span><\/h4>\n\u5206\u5757（Chunking）\u662f\u6574\u4e2a RAG \u91cc\u6027\u4ef7\u6bd4\u6700\u9ad8\u7684\u4f18\u5316\u70b9\u2014\u2014\u4e0d\u82b1\u4e00\u5206\u94b1\u6a21\u578b\u8c03\u7528，\u7eaf\u7b56\u7565\u8c03\u6574，\u5c31\u80fd\u5e26\u6765\u4e24\u4f4d\u6570\u7684\u8d28\u91cf\u63d0\u5347\u3002\u53cd\u4e4b，\u5206\u5757\u505a\u7838\u4e86，\u540e\u9762\u6240\u6709\u73af\u8282\u90fd\u5728\u7ed9\u4e0a\u6e38\u64e6\u5c41\u80a1\u3002<\/p>\n \u5206\u5757\u7684\u6838\u5fc3\u77db\u76fe\u662f\u4e00\u5bf9\u5f20\u529b：<\/p>\n \u5757\u592a\u5927 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u5757\u592a\u5c0f<\/span>
\n\u00a0\u00a0\u2502 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502<\/span>
\n\u00a0\u00a0\u2502 \u00a0\u2713 \u5355\u5757\u4fe1\u606f\u5b8c\u6574 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502 \u00a0\u2713 \u8bed\u4e49\u805a\u7126，\u5d4c\u5165\u5411\u91cf\u66f4\u7cbe\u51c6<\/span>
\n\u00a0\u00a0\u2502 \u00a0\u2713 \u5c11\u51e0\u5757\u5c31\u591f\u56de\u7b54 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502 \u00a0\u2713 \u68c0\u7d22\u547d\u4e2d\u66f4\u7cbe\u786e<\/span>
\n\u00a0\u00a0\u2502 \u00a0\u2717 \u5d4c\u5165\u5411\u91cf\u8bed\u4e49\u88ab\u7a00\u91ca \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502 \u00a0\u2717 \u5355\u5757\u4fe1\u606f\u4e0d\u5b8c\u6574<\/span>
\n\u00a0\u00a0\u2502 \u00a0\u2717 \u65e0\u5173\u5185\u5bb9\u5360\u636e\u4e0a\u4e0b\u6587\u7a97\u53e3 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502 \u00a0\u2717 \u9700\u8981\u53ec\u56de\u5f88\u591a\u5757\u624d\u80fd\u62fc\u51fa\u7b54\u6848<\/span>
\n\u00a0\u00a0\u2502 \u00a0\u2717 \u68c0\u7d22\u533a\u5206\u5ea6\u5dee（\u4ec0\u4e48\u90fd\u80fd\u5339\u914d\u4e0a） \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502 \u00a0\u2717 \u7b54\u6848\u788e\u7247\u5316<\/span><\/p>\n\u6211\u7684\u5b9e\u6218\u57fa\u51c6\u914d\u7f6e（\u540e\u7eed\u7b2c 3 \u671f\u4f1a\u5c55\u5f00）：<\/p>\n \n- \u666e\u901a\u6b63\u6587：300~800 token，overlap 10%~15%<\/li>\n
- \u6309\u7ed3\u6784\u5207，\u4e0d\u6309\u5b57\u6570\u76f2\u5207：Markdown \u6309\u6807\u9898\u5c42\u7ea7，\u4ee3\u7801\u6309\u51fd\u6570，FAQ \u6309\u95ee\u7b54\u5bf9<\/li>\n
- \u6bcf\u5757\u5f3a\u5236\u5e26\u4e0a\u4e0b\u6587\u6807\u9898（Contextual Header）：\u6bd4\u5982 "\u5458\u5de5\u624b\u518c > \u8003\u52e4\u5236\u5ea6 > \u52a0\u73ed\u8c03\u4f11"<\/span>，\u8fd9\u6bb5\u524d\u7f00\u53c2\u4e0e\u5d4c\u5165<\/li>\n
- \u8868\u683c\u6574\u5757\u4fdd\u7559；\u4ee3\u7801\u5757\u4e0d\u5207\u65ad<\/li>\n<\/ul>\n
\u4e3a\u4ec0\u4e48"\u4e0a\u4e0b\u6587\u6807\u9898\u53c2\u4e0e\u5d4c\u5165"\u8fd9\u4e48\u91cd\u8981？\u770b\u4e00\u4e2a\u4f8b\u5b50：\u67d0\u6587\u6863\u91cc\u6709\u4e00\u6bb5"\u6709\u6548\u671f 30 \u4e2a\u81ea\u7136\u65e5，\u81ea\u7b7e\u6536\u4e4b\u65e5\u8d77\u8ba1\u7b97"\u3002\u8fd9\u53e5\u8bdd\u672c\u8eab\u6ca1\u6709\u4efb\u4f55\u4e3b\u9898\u4fe1\u53f7\u2014\u2014\u5d4c\u5165\u540e\u5b83\u548c\u4efb\u4f55\u5173\u4e8e"\u671f\u9650"\u7684\u67e5\u8be2\u90fd\u76f8\u4f3c\u5ea6\u5e73\u5e73\u3002\u4f46\u52a0\u4e0a\u524d\u7f00"\u9000\u8d27\u653f\u7b56 > 7\u5929\u65e0\u7406\u7531\u9000\u8d27 > \u4f8b\u5916\u60c5\u5f62"\u4e4b\u540e，\u8fd9\u6bb5\u6587\u5b57\u7684\u8bed\u4e49\u77ac\u95f4\u951a\u5b9a\u3002\u77ed\u6587\u672c\u5757\u7684\u68c0\u7d22\u8d28\u91cf，\u5f88\u5927\u7a0b\u5ea6\u53d6\u51b3\u4e8e\u5b83\u7684"\u5750\u6807"\u4fe1\u606f\u3002<\/p>\n \u8fd8\u6709\u4e00\u4e2a\u8fdb\u9636\u65b9\u5411\u63d0\u524d\u9884\u544a：LLM \u751f\u6210\u4e0a\u4e0b\u6587（Contextual Retrieval，Anthropic \u63d0\u51fa）\u2014\u2014\u7528 LLM \u7ed9\u6bcf\u4e2a\u5757\u751f\u6210\u4e00\u53e5"\u8fd9\u4e2a\u5757\u5728\u8bb2\u4ec0\u4e48\u3001\u4f4d\u4e8e\u5168\u6587\u4ec0\u4e48\u4f4d\u7f6e"\u7684\u63cf\u8ff0，\u62fc\u5728\u5757\u524d\u9762\u518d\u5d4c\u5165\u3002\u6210\u672c\u4e0a\u5347（\u6bcf\u5757\u4e00\u6b21 LLM \u8c03\u7528），\u4f46\u53ec\u56de\u7387\u63d0\u5347\u663e\u8457，\u9002\u5408\u6838\u5fc3\u77e5\u8bc6\u5e93\u3002\u7b2c 3 \u671f\u8be6\u7ec6\u7b97\u8fd9\u7b14\u8d26\u3002<\/p>\n 0.4.4 \u5143\u6570\u636e\u4e0e\u6743\u9650：\u51b3\u5b9a\u7cfb\u7edf\u80fd\u4e0d\u80fd\u4e0a\u751f\u4ea7<\/span><\/span><\/h4>\nDemo \u548c\u751f\u4ea7\u7684\u5206\u754c\u7ebf，\u5f80\u5f80\u5c31\u5728\u8fd9\u91cc\u3002\u6bcf\u4e2a\u77e5\u8bc6\u5355\u5143\u9664\u4e86\u6587\u672c\u548c\u5411\u91cf，\u8fd8\u5fc5\u987b\u643a\u5e26：<\/p>\n \n\n\n| \n \u5b57\u6bb5<\/p>\n<\/td>\n | \n \u7528\u9014<\/p>\n<\/td>\n | \n \u7f3a\u5931\u540e\u679c<\/p>\n<\/td>\n<\/tr>\n | \n| \n source_id<\/span>\u00a0\/ doc_id<\/span><\/p>\n<\/td>\n | \n \u6eaf\u6e90\u3001\u5f15\u7528\u6807\u6ce8\u3001\u5220\u9664\u91cd\u5efa<\/p>\n<\/td>\n | \n \u65e0\u6cd5\u5f15\u7528\u51fa\u5904；\u6587\u6863\u66f4\u65b0\u65f6\u65e7\u5757\u5220\u4e0d\u5e72\u51c0<\/p>\n<\/td>\n<\/tr>\n | \n| \n chunk_seq<\/span><\/p>\n<\/td>\n | \n \u5757\u6392\u5e8f\u3001\u90bb\u5757\u6269\u5c55<\/p>\n<\/td>\n | \n \u65e0\u6cd5\u7ed9\u6a21\u578b"\u524d\u56e0\u540e\u679c"<\/p>\n<\/td>\n<\/tr>\n | \n| \n acl<\/span>（\u90e8\u95e8\/\u89d2\u8272）<\/p>\n<\/td>\n | \n \u68c0\u7d22\u9636\u6bb5\u6743\u9650\u8fc7\u6ee4<\/p>\n<\/td>\n | \n \u85aa\u916c\u6587\u6863\u88ab\u5168\u516c\u53f8\u95ee\u51fa\u6765，\u5b89\u5168\u4e8b\u6545<\/p>\n<\/td>\n<\/tr>\n | \n| \n tenant_id<\/span><\/p>\n<\/td>\n | \n \u591a\u79df\u6237\u9694\u79bb<\/p>\n<\/td>\n | \n \u79df\u6237 A \u641c\u5230\u79df\u6237 B \u7684\u6570\u636e<\/p>\n<\/td>\n<\/tr>\n | \n| \n doc_type<\/span>\u00a0\/ tags<\/span><\/p>\n<\/td>\n | \n \u5143\u6570\u636e\u8fc7\u6ee4（"\u53ea\u5728\u5236\u5ea6\u7c7b\u6587\u6863\u91cc\u641c"）<\/p>\n<\/td>\n | \n \u68c0\u7d22\u8303\u56f4\u5931\u63a7<\/p>\n<\/td>\n<\/tr>\n | \n| \n version<\/span>\u00a0\/ effective_date<\/span><\/p>\n<\/td>\n | \n \u65b0\u65e7\u7248\u672c\u7ba1\u7406<\/p>\n<\/td>\n | \n \u7528\u6237\u62ff\u5230\u5df2\u4f5c\u5e9f\u7684\u653f\u7b56<\/p>\n<\/td>\n<\/tr>\n | \n| \n embed_model<\/span><\/p>\n<\/td>\n | \n \u6a21\u578b\u5347\u7ea7\u8fc1\u79fb<\/p>\n<\/td>\n | \n \u6362\u5d4c\u5165\u6a21\u578b\u65f6\u65b0\u65e7\u5411\u91cf\u7a7a\u95f4\u6df7\u7528，\u68c0\u7d22\u5168\u4e71<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u5176\u4e2d\u6743\u9650\u8fc7\u6ee4\u7684\u65f6\u673a\u662f\u6700\u5bb9\u6613\u72af\u7684\u67b6\u6784\u9519\u8bef\u3002\u9519\u8bef\u505a\u6cd5：\u5411\u91cf\u68c0\u7d22\u53d6 Top-8，\u7136\u540e\u5728\u4e1a\u52a1\u5c42\u8fc7\u6ee4\u6743\u9650\u3002\u95ee\u9898\u5728\u4e8e\u5411\u91cf\u68c0\u7d22\u8fd4\u56de\u7684 Top-8 \u53ef\u80fd\u5927\u90e8\u5206\u662f\u7528\u6237\u65e0\u6743\u770b\u7684\u5757，\u8fc7\u6ee4\u540e\u53ea\u5269 1~2 \u5757，\u53ec\u56de\u7387\u584c\u65b9\u3002\u6b63\u786e\u505a\u6cd5：\u628a\u6743\u9650\u6761\u4ef6\u5199\u8fdb\u68c0\u7d22\u67e5\u8be2\u672c\u8eab（\u7b2c 5 \u671f\u7ed9\u51fa pgvector\/Milvus \u7684\u5177\u4f53\u5199\u6cd5），\u8ba9 ANN \u641c\u7d22\u53ea\u5728"\u6709\u6743\u770b\u7684\u5b50\u7a7a\u95f4"\u91cc\u8fdb\u884c\u3002<\/p>\n 0.5 \u67e5\u8be2\u94fe\u8def\u8be6\u89e3：\u4ece\u7528\u6237\u95ee\u9898\u5230\u6700\u7ec8\u7b54\u6848<\/span><\/span><\/h3>\n\u79bb\u7ebf\u94fe\u8def\u5efa\u597d\u4e86"\u56fe\u4e66\u9986"，\u5728\u7ebf\u94fe\u8def\u662f"\u501f\u9605\u670d\u52a1"\u3002\u8fd9\u6761\u94fe\u8def\u6709\u56db\u4e2a\u6a21\u5757：\u67e5\u8be2\u7406\u89e3 \u2192 \u6df7\u5408\u68c0\u7d22 \u2192 \u91cd\u6392\u5e8f \u2192 \u4e0a\u4e0b\u6587\u7ec4\u88c5\u4e0e\u751f\u6210\u3002<\/p>\n 0.5.1 \u67e5\u8be2\u7406\u89e3：\u591a\u8f6e\u5bf9\u8bdd\u7684\u751f\u6b7b\u7ebf<\/span><\/span><\/h4>\n\u7528\u6237\u7684\u95ee\u9898\u4ece\u6765\u4e0d\u662f\u7406\u60f3\u7684\u68c0\u7d22\u8bcd\u3002\u4e09\u4e2a\u771f\u5b9e\u95ee\u9898：<\/p>\n \u95ee\u9898\u4e00：\u53e3\u8bed\u5316\u3001\u6709\u7701\u7565\u3002\u00a0"\u90a3\u4e2a\u62a5\u9500\u6d41\u7a0b\u9ebb\u70e6\u5417？"\u2014\u2014\u68c0\u7d22"\u62a5\u9500\u6d41\u7a0b"\u6ca1\u95ee\u9898，\u4f46"\u90a3\u4e2a"\u6307\u7684\u662f\u4ec0\u4e48？\u7eaf\u5411\u91cf\u68c0\u7d22\u628a\u6574\u53e5\u8bdd\u5d4c\u5165，\u4fe1\u53f7\u88ab\u7a00\u91ca\u3002<\/p>\n \u95ee\u9898\u4e8c：\u591a\u8f6e\u6307\u4ee3\u3002\u00a0\u7b2c\u4e00\u8f6e\u95ee"\u5e74\u5047\u6709\u51e0\u5929？"，\u7b2c\u4e8c\u8f6e\u95ee"\u90a3\u8981\u63d0\u524d\u591a\u4e45\u7533\u8bf7？"\u2014\u2014"\u90a3"\u6307\u5e74\u5047\u3002\u5982\u679c\u62ff\u7b2c\u4e8c\u8f6e\u539f\u6587\u53bb\u68c0\u7d22，\u4ec0\u4e48\u90fd\u641c\u4e0d\u5230\u3002\u8fd9\u662f\u591a\u8f6e RAG \u5d29\u6e83\u7684\u5934\u53f7\u539f\u56e0\u3002<\/p>\n \u95ee\u9898\u4e09：\u4e00\u53e5\u8bdd\u6df7\u6742\u591a\u79cd\u610f\u56fe\u3002\u00a0"\u5bf9\u6bd4\u4e00\u4e0b A \u548c B \u4ea7\u54c1\u7684\u4ef7\u683c，\u987a\u4fbf\u8bf4\u4e0b A \u7684\u9000\u6b3e\u653f\u7b56"\u2014\u2014\u8fd9\u5176\u5b9e\u662f\u4e09\u6b21\u68c0\u7d22\u3002<\/p>\n \u67e5\u8be2\u7406\u89e3\u6a21\u5757\u7684\u804c\u8d23\u5c31\u662f\u628a\u8fd9\u4e9b\u95ee\u9898\u89c4\u6574\u6210\u53ef\u68c0\u7d22\u7684\u5f62\u6001\u3002\u7528\u4e00\u4e2a\u4fbf\u5b9c\u7684\u5c0f\u6a21\u578b（tier 0 \u7ea7\u522b）\u505a\u4e09\u4ef6\u4e8b：\u6307\u4ee3\u6d88\u89e3\u3001\u67e5\u8be2\u6539\u5199\u3001\u5173\u952e\u8bcd\u62bd\u53d6：<\/p>\n # query_understand.py \u2014\u2014 \u67e5\u8be2\u7406\u89e3（\u751f\u4ea7\u7248\u8282\u9009）<\/span>
\nREWRITE_PROMPT = """\u4f60\u662f\u4e00\u4e2a\u68c0\u7d22\u67e5\u8be2\u6539\u5199\u5668\u3002\u6839\u636e\u5bf9\u8bdd\u5386\u53f2，\u628a\u7528\u6237\u7684\u6700\u65b0\u95ee\u9898\u6539\u5199\u6210<\/span>
\n1~3 \u4e2a\u72ec\u7acb\u7684\u3001\u4e0d\u542b\u6307\u4ee3\u8bcd\u7684\u68c0\u7d22\u67e5\u8be2\u3002<\/span><\/p>\n\u89c4\u5219：<\/span>
\n– \u6d88\u89e3\u6240\u6709\u4ee3\u8bcd\u548c\u7701\u7565（"\u5b83\/\u90a3\u4e2a\/\u4e0a\u9762\u7684"\u2192\u5177\u4f53\u5bf9\u8c61）<\/span>
\n– \u6bcf\u4e2a\u67e5\u8be2\u5fc5\u987b\u8131\u79bb\u5bf9\u8bdd\u5386\u53f2\u4e5f\u80fd\u88ab\u72ec\u7acb\u7406\u89e3<\/span>
\n– \u5982\u679c\u95ee\u9898\u5305\u542b\u591a\u4e2a\u5b50\u610f\u56fe，\u62c6\u6210\u591a\u4e2a\u67e5\u8be2，\u6bcf\u884c\u4e00\u4e2a<\/span>
\n– \u53ea\u8f93\u51fa\u67e5\u8be2，\u4e0d\u8981\u4efb\u4f55\u89e3\u91ca<\/span><\/p>\n\u5bf9\u8bdd\u5386\u53f2：<\/span>
\n{history}<\/span><\/p>\n\u7528\u6237\u6700\u65b0\u95ee\u9898：{question}<\/span><\/p>\n\u6539\u5199\u540e\u7684\u68c0\u7d22\u67e5\u8be2："""<\/span><\/p>\nasync def rewrite_query(history: list[dict], question: str) -> list[str]:<\/span>
\n\u00a0\u00a0\u00a0\u00a0resp = await cheap_client.chat.completions.create( \u00a0\u00a0# \u4fbf\u5b9c\u7684\u5c0f\u6a21\u578b<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0model="gpt-4o-mini",<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0messages=[{"role": "user",<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0"content": REWRITE_PROMPT.format(<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0history=render_history(history[-6:]), \u00a0# \u53ea\u5e26\u6700\u8fd1\u51e0\u8f6e<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0question=question)}],<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0temperature=0,<\/span>
\n\u00a0\u00a0\u00a0\u00a0)<\/span>
\n\u00a0\u00a0\u00a0\u00a0return [l.strip("-\u2022 ").strip()<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0for l in resp.choices[0].message.content.splitlines() if l.strip()]<\/span><\/p>\n\u5de5\u7a0b\u7ec6\u8282：\u6539\u5199\u7528\u6700\u4fbf\u5b9c\u7684\u5c0f\u6a21\u578b（\u8fd9\u6b65\u4e0d\u9700\u8981\u667a\u529b，\u53ea\u9700\u8981\u89c4\u6574\u80fd\u529b），temperature \u56fa\u5b9a 0 \u4fdd\u8bc1\u7a33\u5b9a，\u5386\u53f2\u53ea\u5e26\u6700\u8fd1 3~6 \u8f6e（\u66f4\u65e9\u7684\u8f6e\u6b21\u5bf9\u6307\u4ee3\u6d88\u89e3\u5e2e\u52a9\u6781\u5c0f，\u5374\u663e\u8457\u589e\u52a0\u5ef6\u8fdf\u548c\u6210\u672c）\u3002\u591a\u67e5\u8be2\u60c5\u5f62，\u6bcf\u6761\u67e5\u8be2\u72ec\u7acb\u8d70\u4e00\u904d\u68c0\u7d22，\u7ed3\u679c\u5408\u5e76\u53bb\u91cd\u540e\u518d\u8fdb\u91cd\u6392\u5e8f\u3002<\/p>\n \u53e6\u5916，\u6539\u5199\u7ed3\u679c\u6700\u597d\u7f13\u5b58：\u540c\u4e00\u4e2a\u4f1a\u8bdd\u91cc\u76f8\u4f3c\u95ee\u9898\u7684\u6539\u5199\u7ed3\u679c\u9ad8\u5ea6\u91cd\u590d，Redis \u7f13\u5b58\u6309 (conversation_id, question_hash)<\/span>\u00a0\u5b58 10 \u5206\u949f，\u80fd\u7701\u4e0b\u53ef\u89c2\u7684\u8c03\u7528\u3002<\/p>\n0.5.2 \u6df7\u5408\u68c0\u7d22：\u4e3a\u4ec0\u4e48\u7eaf\u5411\u91cf\u68c0\u7d22\u6ce8\u5b9a\u4e0d\u591f<\/span><\/span><\/h4>\n\u5411\u91cf\u68c0\u7d22（\u7528\u5d4c\u5165\u6a21\u578b\u628a\u6587\u672c\u6620\u5c04\u6210\u9ad8\u7ef4\u5411\u91cf，\u6309\u4f59\u5f26\u76f8\u4f3c\u5ea6\u627e\u8fd1\u90bb）\u64c5\u957f\u8bed\u4e49\u5339\u914d\u2014\u2014"\u9000\u8d27\u8981\u591a\u4e45\u5230\u8d26"\u80fd\u53ec\u56de"\u9000\u6b3e\u65f6\u6548\u8bf4\u660e"\u3002\u4f46\u5b83\u6709\u4e00\u4e2a\u7ed3\u6784\u6027\u5f31\u70b9：\u5bf9\u7cbe\u786e\u7b26\u53f7\u7684\u533a\u5206\u5ea6\u6781\u4f4e\u3002<\/p>\n \u5d4c\u5165\u6a21\u578b\u7684\u5de5\u4f5c\u539f\u7406\u51b3\u5b9a\u4e86：\u8bed\u4e49\u76f8\u8fd1\u7684\u6587\u672c，\u5411\u91cf\u5c31\u76f8\u8fd1\u3002\u800c"\u9519\u8bef\u7801 E-4021"\u548c"\u9519\u8bef\u7801 E-4022"\u8bed\u4e49\u51e0\u4e4e\u4e00\u6837（\u90fd\u662f\u9519\u8bef\u7801），\u5411\u91cf\u51e0\u4e4e\u91cd\u5408\u2014\u2014\u4f46\u5bf9\u7528\u6237\u6765\u8bf4\u5b83\u4eec\u5929\u5dee\u5730\u522b\u3002\u540c\u7c7b\u95ee\u9898\u8fd8\u5305\u62ec：\u4ea7\u54c1\u578b\u53f7（X200 vs X200 Pro）\u3001\u4eba\u540d\u3001\u6cd5\u89c4\u7f16\u53f7\u3001\u7248\u672c\u53f7（v2.3 vs v2.3.1）\u3002<\/p>\n BM25（\u7ecf\u5178\u5173\u952e\u8bcd\u68c0\u7d22\u7b97\u6cd5，\u57fa\u4e8e\u8bcd\u9891\u548c\u9006\u6587\u6863\u9891\u7387）\u6070\u597d\u76f8\u53cd：\u5b83\u5bf9\u7cbe\u786e\u8bcd\u6781\u5176\u654f\u9510，\u4f46\u5bf9\u540c\u4e49\u6539\u5199\u65e0\u80fd\u4e3a\u529b\u2014\u2014\u641c"\u9000\u8d27"\u6c38\u8fdc\u53ec\u4e0d\u56de\u53ea\u8bf4"\u9000\u6b3e"\u7684\u6587\u6863\u3002<\/p>\n \u4e24\u6761\u8def\u7684\u80fd\u529b\u4e92\u8865，\u4e0d\u662f\u9526\u4e0a\u6dfb\u82b1，\u662f\u5fc5\u9700\u54c1：<\/p>\n \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u67e5\u8be2\u8bed\u4e49\u5339\u914d\u80fd\u529b<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u25b2<\/span>
\n\u00a0\u00a0\u00a0"\u9000\u8d27\u591a\u4e45\u5230\u8d26" \u2500\u2500\u2500\u2500\u2192\u2502\u25cf\u2190\u2500\u2500 \u5411\u91cf\u8def\u547d\u4e2d（\u8bed\u4e49\u8fd1\u4f3c）<\/span>
\n\u00a0\u00a0\u00a0"\u9000\u6b3e\u65f6\u6548" \u00a0\u00a0\u00a0\u00a0\u2500\u2500\u2500\u2500\u2192\u2502<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u25cf\u2190\u2500\u2500 \u5173\u952e\u8bcd\u8def\u547d\u4e2d（\u7cbe\u786e\u5339\u914d）<\/span>
\n\u00a0\u00a0\u00a0"\u9519\u8bef\u7801 E-4021" \u2500\u2500\u2192\u2502 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0（\u5411\u91cf\u8def\u6b64\u65f6\u4e24\u773c\u4e00\u62b9\u9ed1）<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2502<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2192 \u67e5\u8be2\u7cbe\u786e\u5339\u914d\u80fd\u529b<\/span><\/p>\n\u6240\u4ee5\u751f\u4ea7\u7ea7 RAG \u7684\u6807\u51c6\u5f62\u6001\u662f\u53cc\u8def\u53ec\u56de + \u878d\u5408：<\/p>\n # hybrid_retrieval.py \u2014\u2014 \u53cc\u8def\u53ec\u56de\u4e0e RRF \u878d\u5408（\u539f\u7406\u7248）<\/span>
\nimport asyncio<\/span><\/p>\nclass HybridRetriever:<\/span>
\n\u00a0\u00a0\u00a0\u00a0def __init__(self, vector_store, keyword_store, embed_fn):<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0self.vs, self.kw, self.embed = vector_store, keyword_store, embed_fn<\/span><\/p>\n\u00a0\u00a0\u00a0\u00a0async def retrieve(self, queries: list[str], top_k: int = 8) -> list[dict]:<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0# \u591a\u6761\u6539\u5199\u67e5\u8be2 \u00d7 \u53cc\u8def\u68c0\u7d22，\u5168\u90e8\u5e76\u884c<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0tasks = []<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0for q in queries:<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0tasks.append(self.vs.search(await self.embed(q), top_k=50))<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0tasks.append(self.kw.search(q, top_k=50))<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0results = await asyncio.gather(*tasks)<\/span><\/p>\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0fused = self._rrf(results, k=60) \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0# RRF \u878d\u5408，\u89c1\u4e0b<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0return fused[:top_k]<\/span><\/p>\n\u00a0\u00a0\u00a0\u00a0@staticmethod<\/span>
\n\u00a0\u00a0\u00a0\u00a0def _rrf(result_lists, k: int = 60) -> list[dict]:<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0"""Reciprocal Rank Fusion：\u6309\u6392\u540d\u5012\u6570\u878d\u5408，\u7ed5\u5f00\u5206\u6570\u5f52\u4e00\u5316\u96be\u9898"""<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0scores, items = {}, {}<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0for results in result_lists:<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0for rank, item in enumerate(results):<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0cid = item["chunk_id"]<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0scores[cid] = scores.get(cid, 0) + 1.0 \/ (k + rank + 1)<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0items[cid] = item<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0return [{**items[cid], "rrf": s}<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0for cid, s in sorted(scores.items(), key=lambda x: -x[1])]<\/span><\/p>\n\u4e3a\u4ec0\u4e48\u7528 RRF \u800c\u4e0d\u662f"\u5206\u6570\u52a0\u6743"？\u56e0\u4e3a\u4f59\u5f26\u76f8\u4f3c\u5ea6\u548c BM25 \u5206\u6570\u5728\u5b8c\u5168\u4e0d\u540c\u7684\u91cf\u7eb2\u4e0a\u2014\u2014\u4f59\u5f26\u5728 0~1，BM25 \u7406\u8bba\u4e0a\u65e0\u4e0a\u9650\u4e14\u5206\u5e03\u968f\u8bed\u6599\u6f02\u79fb\u3002\u76f4\u63a5\u52a0\u6743\u7b49\u4e8e\u62ff\u7c73\u548c\u516c\u65a4\u76f8\u52a0\u3002RRF \u53ea\u770b"\u6bcf\u8def\u91cc\u6392\u7b2c\u51e0\u540d"，\u5f7b\u5e95\u7ed5\u5f00\u5f52\u4e00\u5316\u95ee\u9898，\u4e14\u6548\u679c\u7a33\u5065，\u662f\u5b66\u672f\u754c\u548c\u5de5\u4e1a\u754c\u5171\u540c\u7684\u6807\u51c6\u7b54\u6848（k=60 \u662f\u539f\u8bba\u6587\u63a8\u8350\u503c）\u3002<\/p>\n 0.5.3 \u91cd\u6392\u5e8f：\u5168\u94fe\u8def\u6027\u4ef7\u6bd4\u6700\u9ad8\u7684\u4e00\u6b65<\/span><\/span><\/h4>\n\u53cc\u8def\u53ec\u56de\u89e3\u51b3\u4e86"\u627e\u5f97\u5230"，\u4f46\u878d\u5408\u7ed3\u679c\u7684\u8d28\u91cf\u53c2\u5dee\u2014\u2014\u5411\u91cf\u8def\u548c\u5173\u952e\u8bcd\u8def\u540450\u6761\u5019\u9009，\u878d\u5408\u6392\u5e8f\u53ea\u662f"\u7c97\u6392"\u3002\u6700\u540e\u4e00\u6b65\u7cbe\u6392，\u7528 Cross-Encoder \u91cd\u6392\u6a21\u578b\u3002<\/p>\n \u539f\u7406\u5dee\u5f02\u4e00\u5f20\u56fe\u770b\u61c2：<\/p>\n \u53cc\u5854（Bi-Encoder，\u7528\u4e8e\u53ec\u56de）： \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u4ea4\u53c9\u7f16\u7801（Cross-Encoder，\u7528\u4e8e\u7cbe\u6392）：<\/span><\/p>\n\u00a0\u00a0query \u2500\u2500\u2192 [\u7f16\u7801\u5668] \u2500\u2500\u2192 \u5411\u91cfq \u2500\u2510 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0query \u2500\u2510<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u251c\u2500\u2500\u2192 \u4f59\u5f26\u76f8\u4f3c\u5ea6 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u251c\u2500\u2500\u2192 [\u540c\u4e00\u4e2a\u7f16\u7801\u5668<\/span>
\n\u00a0\u00a0doc \u00a0\u00a0\u2500\u2500\u2192 [\u7f16\u7801\u5668] \u2500\u2500\u2192 \u5411\u91cfd \u2500\u2518 \u00a0\u00a0\u00a0(q\u548cd\u4ece\u672a"\u89c1\u9762") \u00a0doc \u2500\u2518 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u8054\u5408\u7f16\u7801\u4ea4\u4e92] \u2500\u2500\u2192 \u76f8\u5173\u6027\u5206\u6570<\/span><\/p>\n\u00a0\u00a0\u2713 \u53ef\u79bb\u7ebf\u9884\u8ba1\u7b97，\u6beb\u79d2\u7ea7\u68c0\u7d22 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2713 \u7cbe\u5ea6\u9ad8（\u8bcd\u4e0e\u8bcd\u6df1\u5ea6\u4ea4\u4e92）<\/span>
\n\u00a0\u00a0\u2717 \u7cbe\u5ea6\u6709\u5929\u82b1\u677f \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2717 \u6bcf\u5bf9(query,doc)\u90fd\u8981\u8dd1\u4e00\u904d\u6a21\u578b<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u2192 \u53ea\u80fd\u7528\u4e8e\u5c0f\u5019\u9009\u96c6（\u51e0\u5341\u6761）<\/span><\/p>\n\u6b63\u56e0\u4e3a Cross-Encoder \u6162，\u624d\u9700\u8981"\u7c97\u53ec\u56de \u2192 \u7cbe\u6392\u5e8f"\u7684\u4e24\u7ea7\u6d41\u6c34\u7ebf：\u4fbf\u5b9c\u7684\u53cc\u5854\u6a21\u578b\u4ece\u767e\u4e07\u7ea7\u6587\u6863\u91cc\u635e\u51fa 50 \u6761\u5019\u9009，\u8d35\u7684\u91cd\u6392\u6a21\u578b\u53ea\u5bf9\u8fd9 50 \u6761\u7cbe\u7b97\u76f8\u5173\u6027，\u53d6 Top 5~8 \u7ed9\u6a21\u578b\u3002\u8fd9\u662f\u4fe1\u606f\u68c0\u7d22\u9886\u57df\u51e0\u5341\u5e74\u9a8c\u8bc1\u8fc7\u7684\u7ecf\u5178\u67b6\u6784\u3002<\/p>\n # rerank.py \u2014\u2014 \u91cd\u6392\u5e8f（\u751f\u4ea7\u7248\u8282\u9009）<\/span>
\nfrom sentence_transformers import CrossEncoder<\/span><\/p>\nclass Reranker:<\/span>
\n\u00a0\u00a0\u00a0\u00a0def __init__(self, model_name="BAAI\/bge-reranker-v2-m3"):<\/span>
\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0self.model = CrossEncoder(model_name, max_length=1024)<\/span><\/p>\n\u00a0\u00a0\u00a0\u00a0def rerank(self, query: str, candidates: list[dict], top_k: int = 6) -> list[dict]:<\/span> | | |