{"id":112755,"date":"2026-10-04T20:17:20","date_gmt":"2026-10-04T12:17:20","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/112755.html"},"modified":"2026-10-04T20:17:20","modified_gmt":"2026-10-04T12:17:20","slug":"%e4%bc%81%e4%b8%9a-rag-%e7%9a%84%e4%b8%89%e4%bb%a3%e6%bc%94%e8%bf%9b%ef%bc%9a%e4%bb%8e%e6%b7%b7%e5%90%88%e6%a3%80%e7%b4%a2%e5%88%b0-graphrag-%e5%86%8d%e5%88%b0-agentic","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/112755.html","title":{"rendered":"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic"},"content":{"rendered":"

\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb：\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic<\/h2>\n

\u4f01\u4e1a\u91cc\u7684\u6734\u7d20 RAG \u5931\u8d25\u4ece\u6765\u4e0d\u662f\u56e0\u4e3a\u6a21\u578b\u4e0d\u591f\u5f3a，\u800c\u662f\u56e0\u4e3a\u56db\u4e2a\u7ed3\u6784\u6027\u95ee\u9898：\u7f29\u5199\u591a\u4e49（CRM \u5230\u5e95\u662f\u5ba2\u6237\u5173\u7cfb\u7cfb\u7edf\u8fd8\u662f\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406）\u3001\u672f\u8bed\u4e0d\u4e00\u81f4（\u540c\u4e00\u4e2a\u4e1c\u897f\u5728\u4e09\u4e2a\u90e8\u95e8\u6709\u4e09\u4e2a\u540d\u5b57）\u3001\u65e0\u6cd5\u56de\u7b54\u300c\u8fd9\u53e5\u8bdd\u662f\u4e0d\u662f grounded（\u6709\u636e\u53ef\u4f9d）\u300d\u3001\u4ee5\u53ca\u5355\u8df3\u68c0\u7d22\u5929\u751f\u7b54\u4e0d\u4e86\u6bd4\u8f83\u3001\u89c4\u5212\u3001\u6392\u969c\u8fd9\u7c7b\u591a\u8df3\u95ee\u9898\u3002\u672c\u6587\u6309\u4e09\u4ee3\u67b6\u6784\u62c6\u5f00\u8bb2：\u7b2c\u4e00\u4ee3\u6df7\u5408\u68c0\u7d22\u7528 dense \u5411\u91cf\u7ba1\u540c\u4e49\u3001BM25 \u7ba1\u7cbe\u786e\u5339\u914d\u4e0e\u4ea7\u54c1\u7801，\u4e24\u8def\u5e76\u53d1\u6267\u884c\u628a\u68c0\u7d22\u5ef6\u8fdf\u538b\u6389 40% \u4ee5\u4e0a，\u518d\u7528 RRF（Reciprocal Rank Fusion，Cormack \/ Clarke \/ Buettcher 2009）\u505a\u514d\u5f52\u4e00\u5316\u7684\u878d\u5408\u6392\u5e8f，\u914d\u4e09\u5c42\u53bb\u91cd\u4e0e cross-encoder \u91cd\u6392；\u7b2c\u4e8c\u4ee3 GraphRAG \u628a chunk \u4e4b\u95f4\u7684\u5b9e\u4f53\u3001\u5173\u7cfb\u3001\u672c\u4f53\u5c42\u5efa\u8d77\u6765，\u5e76\u7ed9\u51fa\u4e00\u4e2a\u5173\u952e\u5de5\u7a0b\u5224\u65ad\u2014\u2014\u751f\u4ea7\u73af\u5883\u7528\u89c4\u5219\u5316\u591a\u8d9f\u5b9e\u4f53\u62bd\u53d6\u4f18\u4e8e LLM NER，\u56e0\u4e3a\u5b83\u662f\u5fae\u79d2\u7ea7\u3001\u96f6\u8fb9\u9645\u6210\u672c\u3001\u786e\u5b9a\u6027\u7684；\u7b2c\u4e09\u4ee3 Agentic \u8865\u4e0a\u5de5\u5177\u4f7f\u7528\u3001\u8bb0\u5fc6\u7ba1\u7406\u3001\u89c4\u5212\u3001\u534f\u8c03\u3001\u8bc4\u4f30\u4e94\u6761\u7eaa\u5f8b\u4e0e\u5b89\u5168\u8fb9\u754c，\u7528\u4e58\u6cd5\u800c\u975e\u5e73\u5747\u6765\u7b97\u7f6e\u4fe1\u5ea6（0.9 x 0.9 = 0.81，0.9 x 0.1 = 0.09，\u800c\u4e0d\u662f\u8bef\u5bfc\u6027\u7684 0.5），\u4f4e\u4e8e\u9608\u503c\u89e6\u53d1\u6709\u754c\u91cd\u8bd5，\u5e76\u7ed9\u6bcf\u4e2a\u7ec4\u4ef6\u5206\u914d\u663e\u5f0f\u5ef6\u8fdf\u9884\u7b97\u3002\u6587\u672b\u9644\u4e0a\u6458\u8981\u6c61\u67d3\u5411\u91cf\u5e93\u3001\u65e0\u9650\u585e\u4e0a\u4e0b\u6587\u3001AI \u540c\u65f6\u8bfb\u5199\u540c\u4e00\u5e93\u8fd9\u4e09\u4e2a\u53cd\u6a21\u5f0f\u7684\u4fee\u6cd5，\u4ee5\u53ca 8 \u6761\u8e29\u5751\u6392\u67e5\u624b\u518c\u3002<\/p>\n

\u5173\u952e\u8bcd：\u4f01\u4e1a RAG\u3001\u6df7\u5408\u68c0\u7d22\u3001RRF\u3001GraphRAG\u3001Agentic RAG\u3001cross-encoder\u3001\u7f6e\u4fe1\u5ea6\u3001\u5ef6\u8fdf\u9884\u7b97<\/p>\n

\u4e00\u3001\u5199\u5728\u524d\u9762：\u4e3a\u4ec0\u4e48\u73b0\u5728\u5fc5\u987b\u804a\u8fd9\u4ef6\u4e8b<\/h3>\n

2026 \u5e74\u8fd8\u5728\u8ba8\u8bba RAG（Retrieval-Augmented Generation，\u68c0\u7d22\u589e\u5f3a\u751f\u6210）\u67b6\u6784\u9009\u578b，\u542c\u8d77\u6765\u50cf\u662f\u7092\u51b7\u996d\u3002\u4f46\u6211\u5728\u8fc7\u53bb\u534a\u5e74\u91cc\u89c1\u8fc7\u7684\u4f01\u4e1a RAG \u9879\u76ee，\u5931\u8d25\u7387\u4f9d\u7136\u5f88\u9ad8，\u800c\u4e14\u5931\u8d25\u65b9\u5f0f\u9ad8\u5ea6\u96f7\u540c：\u68c0\u7d22\u63a5\u4e0a\u4e86，\u5411\u91cf\u5e93\u5efa\u8d77\u6765\u4e86，demo \u95ee\u7b54\u770b\u7740\u633a\u50cf\u56de\u4e8b，\u4e00\u653e\u5230\u771f\u5b9e\u4e1a\u52a1\u95ee\u7b54\u91cc\u5c31\u88ab\u7528\u6237\u6295\u8bc9\u300c\u7b54\u975e\u6240\u95ee\u300d\u6216\u8005\u300c\u8bf4\u7684\u662f\u6211\u4eec\u516c\u53f8\u7684\u4e1c\u897f\u4f46\u6570\u5b57\u662f\u7f16\u7684\u300d\u3002<\/p>\n

\u590d\u76d8\u4e0b\u6765，\u6839\u56e0\u51e0\u4e4e\u4e0d\u5728\u6a21\u578b\u4fa7\u30022026-09-08 KDnuggets \u90a3\u7bc7\u4f01\u4e1a RAG \u4e0e Agentic AI \u7cfb\u7edf\u7684\u68b3\u7406\u91cc\u628a\u6f14\u8fdb\u8def\u5f84\u5206\u6210\u4e09\u4ee3，\u6211\u8ba4\u540c\u8fd9\u4e2a\u5206\u6cd5，\u4f46\u60f3\u8865\u4e00\u53e5\u66f4\u76f4\u767d\u7684\u8bdd：**\u8fd9\u4e09\u4ee3\u4e0d\u662f\u300c\u6a21\u578b\u8d8a\u6765\u8d8a\u5f3a\u300d\u7684\u6f14\u8fdb，\u800c\u662f\u300c\u5de5\u7a0b\u7ea6\u675f\u8d8a\u6765\u8d8a\u663e\u5f0f\u300d\u7684\u6f14\u8fdb\u3002**\u6bcf\u4e00\u4ee3\u89e3\u51b3\u7684\u662f\u4e0a\u4e00\u4ee3\u5728\u5de5\u7a0b\u4e0a\u8bf4\u4e0d\u51fa\u53e3\u7684\u90a3\u4e2a\u95ee\u9898\u3002<\/p>\n

\u5148\u628a\u6734\u7d20 RAG \u5728\u4f01\u4e1a\u573a\u666f\u91cc\u7684\u56db\u5b97\u7f6a\u5217\u51fa\u6765，\u540e\u9762\u6240\u6709\u8bbe\u8ba1\u90fd\u662f\u5bf9\u7740\u8fd9\u56db\u6761\u6253\u7684：<\/p>\n\n\u95ee\u9898\u5177\u4f53\u8868\u73b0\u4e3a\u4ec0\u4e48\u6734\u7d20 RAG \u6cbb\u4e0d\u4e86<\/tr>\n\n\n\n\n\n
\u7f29\u5199\u591a\u4e49<\/td>\nCRM \u5728\u9500\u552e\u90e8\u95e8\u6307\u5ba2\u6237\u5173\u7cfb\u7ba1\u7406，\u5728\u8fd0\u7ef4\u90e8\u95e8\u6307 Cluster Resource Manager；\u540c\u4e00\u4efd\u8bed\u6599\u91cc\u4e24\u4e2a\u542b\u4e49\u6df7\u7740\u51fa\u73b0<\/td>\n\u5411\u91cf\u7a7a\u95f4\u91cc\u4e24\u4e2a\u542b\u4e49\u7684 embedding（\u5d4c\u5165\u5411\u91cf）\u8ddd\u79bb\u5f88\u8fd1，\u68c0\u7d22\u65e0\u6cd5\u533a\u5206<\/td>\n<\/tr>\n
\u672f\u8bed\u4e0d\u4e00\u81f4<\/td>\n\u540c\u4e00\u4e2a\u300c\u5ba2\u6237\u4e3b\u6570\u636e\u300d\u5728 A \u7cfb\u7edf\u53eb Customer，\u5728 B \u7cfb\u7edf\u53eb \u5ba2\u6237，\u5728 C \u7cfb\u7edf\u53eb CUST_MSTR<\/td>\n\u5355\u8def\u5411\u91cf\u68c0\u7d22\u53ea\u4f1a\u547d\u4e2d\u5176\u4e2d\u4e00\u79cd\u5199\u6cd5，\u53ec\u56de\u4e0d\u5168<\/td>\n<\/tr>\n
\u65e0\u6cd5\u56de\u7b54\u662f\u5426 grounded<\/td>\n\u7528\u6237\u95ee\u300c\u8fd9\u53e5\u8bdd\u51fa\u81ea\u54ea\u4efd\u6587\u6863\u7684\u54ea\u4e00\u8282\u300d，\u7cfb\u7edf\u7b54\u4e0d\u4e0a\u6765；\u7b54\u9519\u4e86\u4e5f\u4e0d\u77e5\u9053\u662f\u54ea\u4e00\u6b65\u9519\u7684<\/td>\n\u68c0\u7d22\u4e0e\u751f\u6210\u4e4b\u95f4\u6ca1\u6709\u53ef\u5ba1\u8ba1\u7684\u4e2d\u95f4\u6001，\u7f3a\u5c11\u5f15\u7528\u5f52\u5c5e<\/td>\n<\/tr>\n
\u5355\u8df3\u7b54\u4e0d\u4e86\u591a\u8df3\u95ee\u9898<\/td>\n\u300c\u5bf9\u6bd4 A \u4ea7\u54c1\u548c B \u4ea7\u54c1\u7684 SLA，\u5e76\u7ed9\u51fa\u8fc1\u79fb\u6b65\u9aa4\u300d\u300c\u8fd9\u53f0\u673a\u5668\u62a5\u8b66，\u53ef\u80fd\u7684\u539f\u56e0\u6709\u54ea\u4e9b，\u600e\u4e48\u6392\u67e5\u300d<\/td>\n\u8fd9\u7c7b\u95ee\u9898\u9700\u8981\u591a\u6b21\u68c0\u7d22 + \u4e2d\u95f4\u63a8\u7406，\u4e00\u6b21 top-k \u53ec\u56de\u62ff\u4e0d\u5230\u5168\u90e8\u524d\u63d0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u7b2c\u4e09\u6761\u7684\u5371\u5bb3\u6700\u5927，\u56e0\u4e3a\u5b83\u8ba9\u6574\u4e2a\u7cfb\u7edf\u65e0\u6cd5\u88ab\u4fe1\u4efb\u4e5f\u65e0\u6cd5\u88ab\u8c03\u8bd5\u3002\u4e00\u4e2a\u6ca1\u6709 grounded \u5224\u5b9a\u7684 RAG，\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u4f1a\u5f15\u7ecf\u636e\u5178\u5730\u6492\u8c0e\u7684\u7cfb\u7edf\u2014\u2014\u5b83\u7ed9\u51fa\u7684\u5f15\u7528\u770b\u8d77\u6765\u5f88\u771f，\u4f46\u53ef\u80fd\u6765\u81ea\u4e00\u4e2a\u8bed\u4e49\u76f8\u8fd1\u5374\u5b8c\u5168\u4e0d\u540c\u7684\u6587\u6863\u3002<\/p>\n

\u8fd9\u7bc7\u6587\u7ae0\u7684\u7ed3\u6784\u662f：\u7b2c\u4e8c\u8282\u628a\u4e09\u4ee3\u67b6\u6784\u653e\u5728\u4e00\u5f20\u56fe\u4e0a\u5bf9\u6bd4；\u7b2c\u4e09\u8282\u8bb2\u900f\u539f\u7406，\u5305\u62ec RRF \u7684\u516c\u5f0f\u4e0e\u63a8\u5bfc\u3001\u4e09\u5c42\u53bb\u91cd\u3001\u89c4\u5219\u5316\u5b9e\u4f53\u62bd\u53d6\u7684\u4e09\u8d9f\u5339\u914d\u3001\u4e58\u6cd5\u7f6e\u4fe1\u5ea6\u548c\u6709\u754c\u91cd\u8bd5\u72b6\u6001\u673a；\u7b2c\u56db\u3001\u4e94\u8282\u7ed9\u73af\u5883\u4e0e\u4e00\u5957\u5b8c\u6574\u53ef\u8dd1\u7684\u5b9e\u73b0；\u7b2c\u516d\u8282\u662f\u5b9e\u6d4b\u6570\u636e；\u7b2c\u4e03\u8282 8 \u6761\u8e29\u5751；\u7b2c\u516b\u8282\u8bb2\u4e09\u4e2a\u53cd\u6a21\u5f0f；\u7b2c\u4e5d\u8282\u8bb2\u4ec0\u4e48\u60c5\u51b5\u4e0b\u4e0d\u8be5\u4e0a\u7b2c\u4e09\u4ee3\u3002<\/p>\n

\u7ed3\u8bba\u5148\u6446：**\u5927\u591a\u6570\u56e2\u961f\u5e94\u8be5\u628a 80% \u7684\u7cbe\u529b\u653e\u5728\u7b2c\u4e00\u4ee3（\u6df7\u5408\u68c0\u7d22 + \u53bb\u91cd + \u91cd\u6392）\u4e0a，\u5b83\u80fd\u89e3\u51b3 70% \u7684\u95ee\u9898，\u800c\u4e14\u6ca1\u6709\u65b0\u589e\u4efb\u4f55\u4e0d\u53ef\u63a7\u56e0\u7d20\u3002**\u7b2c\u4e8c\u4ee3\u548c\u7b2c\u4e09\u4ee3\u662f\u5bf9\u7279\u5b9a\u95ee\u9898\u5f62\u6001\u7684\u8865\u5145，\u4e0d\u662f\u5347\u7ea7\u8def\u7ebf\u3002\u628a\u987a\u5e8f\u641e\u53cd\u2014\u2014\u8fd8\u5728\u7528\u5355\u8def\u5411\u91cf\u68c0\u7d22\u5c31\u53bb\u641e\u591a Agent \u7f16\u6392\u2014\u2014\u662f\u6700\u5e38\u89c1\u7684\u8d44\u6e90\u9519\u914d\u3002<\/p>\n

\u4e8c\u3001\u6280\u672f\u80cc\u666f\u4e0e\u73b0\u72b6\u68b3\u7406<\/h3>\n

2.1 \u4e09\u4ee3\u67b6\u6784\u5bf9\u7167<\/h4>\n\n\u7ef4\u5ea6\u7b2c\u4e00\u4ee3 \u6df7\u5408\u68c0\u7d22\u7b2c\u4e8c\u4ee3 GraphRAG\u7b2c\u4e09\u4ee3 Agentic<\/tr>\n\n\n\n\n\n\n\n\n
\u68c0\u7d22\u5355\u4f4d<\/td>\nchunk（\u6587\u672c\u5757）<\/td>\nchunk + \u5b9e\u4f53 + \u5173\u7cfb<\/td>\n\u5de5\u5177\u8c03\u7528\u5e8f\u5217<\/td>\n<\/tr>\n
\u53ec\u56de\u65b9\u5f0f<\/td>\ndense \u5411\u91cf + BM25 \u5e76\u884c<\/td>\n\u56fe\u904d\u5386 + \u5411\u91cf\u6df7\u5408<\/td>\n\u89c4\u5212\u5668\u51b3\u5b9a\u7684\u591a\u8f6e\u68c0\u7d22<\/td>\n<\/tr>\n
\u878d\u5408<\/td>\nRRF<\/td>\n\u56fe\u90bb\u57df\u6269\u5c55 + RRF<\/td>\n\u7f6e\u4fe1\u5ea6\u9a71\u52a8\u7684\u6709\u754c\u91cd\u8bd5<\/td>\n<\/tr>\n
\u5178\u578b\u67e5\u8be2<\/td>\n\u300cXX \u4ea7\u54c1\u7684\u9000\u8d27\u653f\u7b56\u662f\u4ec0\u4e48\u300d<\/td>\n\u300cCRM \u548c CUST_MSTR \u662f\u540c\u4e00\u4e2a\u4e1c\u897f\u5417\u300d<\/td>\n\u300c\u5bf9\u6bd4 A\/B \u4e24\u4ea7\u54c1\u7684 SLA \u5e76\u7ed9\u8fc1\u79fb\u6b65\u9aa4\u300d<\/td>\n<\/tr>\n
\u5ef6\u8fdf\u91cf\u7ea7<\/td>\n100~300ms<\/td>\n300~800ms<\/td>\n2~15s<\/td>\n<\/tr>\n
\u65b0\u589e\u4e0d\u53ef\u63a7\u56e0\u7d20<\/td>\n\u57fa\u672c\u6ca1\u6709<\/td>\n\u56fe\u6784\u5efa\u8d28\u91cf<\/td>\n\u89c4\u5212\u5668\u7684\u975e\u786e\u5b9a\u6027<\/td>\n<\/tr>\n
\u9002\u5408\u89c4\u6a21<\/td>\n\u4efb\u610f<\/td>\n\u8bed\u6599 > 5 \u4e07 chunk \u4e14\u5b9e\u4f53\u5bc6\u96c6<\/td>\n\u591a\u8df3 \/ \u8de8\u7cfb\u7edf\u4efb\u52a1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

2.2 \u4e3a\u4ec0\u4e48 dense \u548c sparse \u90fd\u4e0d\u80fd\u5355\u72ec\u7528<\/h4>\n

\u8fd9\u4ef6\u4e8b\u5728\u5de5\u7a0b\u4e0a\u53cd\u590d\u88ab\u9a8c\u8bc1，\u4f46\u503c\u5f97\u628a\u673a\u5236\u8bb2\u6e05\u695a，\u56e0\u4e3a\u300c\u4e3a\u4ec0\u4e48\u300d\u51b3\u5b9a\u4e86\u4f60\u600e\u4e48\u8c03\u53c2\u3002<\/p>\n

dense \u5411\u91cf\u628a\u6587\u672c\u6620\u5c04\u5230\u4e00\u4e2a\u8fde\u7eed\u7a7a\u95f4，\u8bed\u4e49\u76f8\u8fd1\u7684\u6587\u672c\u8ddd\u79bb\u8fd1\u3002\u5b83\u64c5\u957f\u7684\u662f\u300c\u540c\u4e49\u66ff\u6362\u300d：\u7528\u6237\u95ee\u300c\u600e\u4e48\u9000\u6b3e\u300d，\u6587\u6863\u5199\u300c\u9000\u8d27\u6b3e\u9879\u9000\u56de\u6d41\u7a0b\u300d，\u5411\u91cf\u80fd\u5339\u914d\u4e0a，\u5173\u952e\u8bcd\u5339\u914d\u4e0d\u4e0a\u3002\u5b83\u7684\u8f6f\u808b\u662f\u7cbe\u786e\u5339\u914d：\u4ea7\u54c1\u7801 AX-2039-B\u3001\u9519\u8bef\u7801 E_TIMEOUT_409\u3001\u7f29\u5199 CRM\u3001k8s，\u8fd9\u4e9b\u4e1c\u897f\u5728\u5411\u91cf\u7a7a\u95f4\u91cc\u51e0\u4e4e\u6ca1\u6709\u533a\u5206\u5ea6\u2014\u2014AX-2039-B \u548c AX-2039-C \u7684 embedding \u4f59\u5f26\u76f8\u4f3c\u5ea6\u53ef\u80fd\u5728 0.97 \u4ee5\u4e0a，\u56e0\u4e3a\u5b57\u7b26\u5c42\u9762 90% \u91cd\u5408\u3002<\/p>\n

**BM25（Best Matching 25，\u4e00\u79cd\u7ecf\u5178\u7684\u7a00\u758f\u8bcd\u9891-\u9006\u6587\u6863\u9891\u7387\u6392\u5e8f\u51fd\u6570）**\u8d70\u7684\u662f\u53e6\u4e00\u6761\u8def：\u5b83\u770b\u8bcd\u9879\u662f\u5426\u7cbe\u786e\u51fa\u73b0\u3001\u8bcd\u9891\u591a\u5c11\u3001\u6587\u6863\u591a\u957f\u3002\u5b83\u5bf9\u4ea7\u54c1\u7801\u3001\u9519\u8bef\u7801\u3001\u7f29\u5199\u8fd9\u7c7b\u300c\u4e00\u5b57\u4e0d\u5dee\u300d\u7684\u67e5\u8be2\u7a33\u51c6\u72e0，\u4f46\u5bf9\u540c\u4e49\u66ff\u6362\u65e0\u80fd\u4e3a\u529b\u2014\u2014\u7528\u6237\u8bf4\u300c\u9000\u6b3e\u300d，\u6587\u6863\u8bf4\u300c\u9000\u8d27\u300d，BM25 \u4e00\u5206\u522b\u4e0d\u7ed9\u3002<\/p>\n

\u6240\u4ee5\u4e24\u6761\u8def\u4e0d\u662f\u300c\u54ea\u4e2a\u66f4\u597d\u300d\u7684\u5173\u7cfb，\u662f\u4e92\u8865\u4e14\u5fc5\u987b\u540c\u65f6\u5f00\u7684\u5173\u7cfb\u30022026 \u5e74\u8fd8\u5728\u4e89\u8bba\u300c\u7528\u5411\u91cf\u8fd8\u662f\u7528\u5173\u952e\u8bcd\u300d\u7684\u56e2\u961f，\u901a\u5e38\u662f\u88ab\u67d0\u4e2a benchmark \u4e0a\u7684\u5355\u8def\u6307\u6807\u8bef\u5bfc\u4e86：\u90a3\u4e9b benchmark \u7684\u67e5\u8be2\u96c6\u5f80\u5f80\u504f\u5411\u5176\u4e2d\u4e00\u79cd\u7c7b\u578b\u3002<\/p>\n

\u771f\u5b9e\u7684\u5206\u5e03\u662f：\u4f01\u4e1a\u95ee\u7b54\u91cc\u5927\u7ea6 60% \u5230 70% \u7684\u67e5\u8be2\u662f\u8bed\u4e49\u578b\u7684，20% \u5230 30% \u5e26\u660e\u786e\u7684\u7cbe\u786e\u6807\u8bc6\u7b26，\u5269\u4e0b 5% \u5230 10% \u662f\u4e24\u8005\u7684\u6df7\u5408（\u300cAX-2039-B \u7684\u9000\u6b3e\u6d41\u7a0b\u300d）\u3002\u4efb\u4f55\u5355\u8def\u65b9\u6848\u90fd\u4f1a\u5728\u5bf9\u5e94\u90a3\u90e8\u5206\u67e5\u8be2\u4e0a\u660e\u663e\u6389\u961f\u3002<\/p>\n

2.3 \u5e76\u53d1\u6267\u884c\u662f\u5ef6\u8fdf\u7684\u5173\u952e，\u4e0d\u662f\u68c0\u7d22\u7b97\u6cd5<\/h4>\n

\u5f88\u591a\u56e2\u961f\u628a\u6df7\u5408\u68c0\u7d22\u5b9e\u73b0\u6210\u4e32\u884c：\u5148\u67e5\u5411\u91cf（180ms），\u518d\u67e5 BM25（90ms），\u518d\u878d\u5408（20ms），\u603b\u5ef6\u8fdf 290ms\u3002\u8fd9\u662f\u6d6a\u8d39\u2014\u2014\u4e24\u8def\u68c0\u7d22\u4e4b\u95f4\u6ca1\u6709\u4efb\u4f55\u4f9d\u8d56\u5173\u7cfb，\u5b8c\u5168\u53ef\u4ee5\u5e76\u53d1\u3002<\/p>\n

\u6539\u6210 asyncio.gather \u4e4b\u540e，\u603b\u5ef6\u8fdf\u53d8\u6210 max(180, 90) + 20 = 200ms，\u964d 31%\u3002\u5982\u679c\u5411\u91cf\u68c0\u7d22\u90a3\u4e00\u8def\u672c\u8eab\u5c31\u80fd\u4f18\u5316\u5230 120ms，\u603b\u5ef6\u8fdf\u662f 140ms，\u964d\u5e45\u8d85\u8fc7 50%\u3002\u6211\u5728\u81ea\u5df1\u7684\u73af\u5883\u91cc\u5b9e\u6d4b\u4ece 287ms \u964d\u5230 168ms，\u964d 41.5%\u2014\u2014\u8fd9\u5c31\u662f\u6807\u9898\u91cc\u90a3\u4e2a 40%+ \u7684\u6765\u6e90\u3002<\/p>\n

\u503c\u5f97\u6ce8\u610f\u7684\u662f，\u8fd9\u4e2a\u4f18\u5316\u548c\u68c0\u7d22\u8d28\u91cf\u5b8c\u5168\u65e0\u5173，\u7eaf\u7cb9\u662f\u5de5\u7a0b\u6392\u5e03\u3002\u4f46\u5b83\u7ecf\u5e38\u88ab\u5ffd\u7565，\u56e0\u4e3a\u5927\u5bb6\u5728\u4ee3\u7801\u91cc\u5199\u7684\u662f\u987a\u5e8f\u6267\u884c\u7684\u540c\u6b65\u51fd\u6570，\u770b\u8d77\u6765\u300c\u5f88\u81ea\u7136\u300d\u3002<\/p>\n

2.4 \u53bb\u91cd\u4e3a\u4ec0\u4e48\u5fc5\u987b\u505a\u5728 rerank \u4e4b\u524d<\/h4>\n

cross-encoder（\u4ea4\u53c9\u7f16\u7801\u5668）reranker \u7684\u6210\u672c\u662f O(\u5019\u9009\u6570 x \u5355\u6b21\u524d\u5411)，\u800c\u4e14\u5b83\u662f\u4f4d\u7f6e\u654f\u611f\u7684：\u5982\u679c\u4e24\u4e2a\u5019\u9009\u662f\u540c\u4e00\u6bb5\u6587\u672c\u7684\u4e0d\u540c\u7248\u672c，\u5b83\u4eec\u4f1a\u5360\u636e\u5019\u9009\u5217\u8868\u7684\u4e24\u4e2a\u4f4d\u7f6e，\u628a\u771f\u6b63\u76f8\u5173\u7684\u7b2c\u4e09\u4e2a\u5019\u9009\u6324\u5230\u540e\u9762\u53bb\u3002<\/p>\n

\u66f4\u7cdf\u7684\u662f，\u91cd\u6392\u6a21\u578b\u5728\u770b\u5230\u4e24\u4e2a\u51e0\u4e4e\u76f8\u540c\u7684\u5019\u9009\u65f6，\u5f80\u5f80\u7ed9\u5b83\u4eec\u76f8\u8fd1\u7684\u9ad8\u5206，\u4e8e\u662f top-5 \u91cc\u4e09\u4e2a\u4f4d\u7f6e\u88ab\u540c\u4e00\u6bb5\u5185\u5bb9\u5360\u6ee1，\u6709\u6548\u4fe1\u606f\u91cf\u76f4\u63a5\u8170\u65a9\u3002\u6240\u4ee5\u987a\u5e8f\u5fc5\u987b\u662f：\u53ec\u56de -> \u4e09\u5c42\u53bb\u91cd -> rerank\u3002\u5148 rerank \u518d\u53bb\u91cd\u662f\u9519\u7684，\u56e0\u4e3a\u4f60\u4ed8\u4e86\u4e24\u500d\u7684\u91cd\u6392\u6210\u672c，\u7136\u540e\u6254\u6389\u4e00\u534a\u7ed3\u679c\u3002<\/p>\n

\u4e09\u3001\u6838\u5fc3\u539f\u7406\u62c6\u89e3<\/h3>\n

3.1 \u7b2c\u4e00\u4ee3：\u6df7\u5408\u68c0\u7d22\u7684\u5b8c\u6574\u6570\u636e\u6d41<\/h4>\n

query
\n |
\n +————-+————-+
\n | |
\n [dense \u5206\u652f] [sparse \u5206\u652f]
\n query -> embedding query -> \u5206\u8bcd
\n ANN \u68c0\u7d22 top-50 BM25 \u68c0\u7d22 top-50
\n | |
\n +————-+————-+
\n |
\n [RRF \u878d\u5408] score(d) = sum_i 1\/(k + rank_i(d))
\n |
\n [\u4e09\u5c42\u53bb\u91cd]
\n 1. \u552f\u4e00 ID
\n 2. \u6765\u6e90\u4f4d\u7f6e（doc_id + \u7ae0\u8282 + \u5b57\u7b26\u533a\u95f4）
\n 3. \u5185\u5bb9\u6307\u7eb9（\u5f52\u4e00\u5316\u540e simhash \/ sha1）
\n |
\n [cross-encoder rerank] top-50 -> top-8
\n |
\n [\u4e0a\u4e0b\u6587\u7ec4\u88c5 + \u5f15\u7528\u5f52\u5c5e]
\n |
\n \u751f\u6210<\/p>\n

\u6570\u636e\u6d41\u91cc\u6bcf\u4e00\u6b65\u90fd\u6709\u660e\u786e\u7684\u5b58\u5728\u7406\u7531，\u5220\u6389\u4efb\u4f55\u4e00\u6b65\u90fd\u4f1a\u6709\u5bf9\u5e94\u7684\u5931\u8d25\u6a21\u5f0f\u3002\u4e0b\u9762\u9010\u4e2a\u62c6\u3002<\/p>\n

3.2 RRF：\u516c\u5f0f\u3001\u63a8\u5bfc\u4e0e\u4e3a\u4ec0\u4e48\u5b83\u4e0d\u9700\u8981\u5206\u6570\u5f52\u4e00\u5316<\/h4>\n

RRF \u6765\u81ea Cormack\u3001Clarke\u3001Buettcher \u5728 SIGIR 2009 \u4e0a\u7684\u8bba\u6587\u300aReciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods\u300b\u3002\u516c\u5f0f\u6781\u7b80：<\/p>\n

RRF_score(d) = sum over i in 1..n of 1 \/ (k + rank_i(d))<\/p>\n

\u5176\u4e2d：
\n d = \u5019\u9009\u6587\u6863
\n n = \u68c0\u7d22\u8def\u6570（\u8fd9\u91cc\u662f 2：dense \u4e0e sparse）
\n rank_i(d) = \u6587\u6863 d \u5728\u7b2c i \u8def\u7ed3\u679c\u4e2d\u7684\u6392\u540d，\u4ece 1 \u5f00\u59cb
\n k = \u5e38\u6570，\u539f\u8bba\u6587\u7528 60；d \u53ea\u5728\u67d0\u51e0\u8def\u51fa\u73b0\u65f6，\u5176\u4f59\u8def\u89c6\u4e3a rank = \u65e0\u7a77（\u8be5\u9879\u8d21\u732e 0）<\/p>\n

\u4e3a\u4ec0\u4e48\u5b83\u597d\u7528，\u4e09\u4e2a\u7406\u7531：<\/p>\n

**\u7b2c\u4e00，\u5b83\u53ea\u7528\u6392\u540d，\u4e0d\u7528\u5206\u6570，\u6240\u4ee5\u4e0d\u9700\u8981\u5f52\u4e00\u5316\u3002**\u8fd9\u662f\u6700\u5173\u952e\u7684\u5de5\u7a0b\u4ef7\u503c\u3002dense \u5206\u652f\u8fd4\u56de\u7684\u662f\u4f59\u5f26\u76f8\u4f3c\u5ea6，\u8303\u56f4\u5927\u7ea6\u662f 0.3 \u5230 0.95；sparse \u5206\u652f\u8fd4\u56de\u7684 BM25 \u5206\u6570\u662f\u65e0\u4e0a\u754c\u7684\u6b63\u6570，\u5b9e\u6d4b\u8303\u56f4\u53ef\u80fd\u4ece 2.1 \u5230 34.7\u3002\u8981\u628a\u8fd9\u4e24\u4e2a\u5206\u6570\u878d\u5728\u4e00\u8d77，\u4f60\u5fc5\u987b\u5148\u5f52\u4e00\u5316\u2014\u2014\u800c\u5f52\u4e00\u5316\u9700\u8981\u77e5\u9053\u5206\u6570\u7684\u5206\u5e03，\u5206\u5e03\u4f1a\u968f\u8bed\u6599\u53d8\u5316，\u4e8e\u662f\u4f60\u8981\u7ef4\u62a4\u4e00\u4e2a\u968f\u65f6\u4f1a\u8fc7\u65f6\u7684\u5f52\u4e00\u5316\u53c2\u6570\u3002RRF \u5b8c\u5168\u7ed5\u5f00\u4e86\u8fd9\u4e2a\u95ee\u9898，\u628a\u300c\u5206\u6570\u4e0d\u53ef\u6bd4\u300d\u8fd9\u4e2a\u96be\u9898\u6362\u6210\u4e86\u300c\u6392\u540d\u53ef\u6bd4\u300d\u8fd9\u4e2a\u663e\u7136\u6210\u7acb\u7684\u524d\u63d0\u3002<\/p>\n

**\u7b2c\u4e8c，\u5b83\u5bf9\u5355\u8def\u7ed3\u679c\u7684\u5206\u6570\u6f02\u79fb\u514d\u75ab\u3002**\u4f60\u6362\u4e86\u4e00\u4e2a embedding \u6a21\u578b，\u4f59\u5f26\u76f8\u4f3c\u5ea6\u7684\u7edd\u5bf9\u503c\u5168\u53d8\u4e86，\u4f46\u6392\u540d\u5927\u4f53\u7a33\u5b9a\u3002\u7528\u52a0\u6743\u6c42\u548c\u878d\u5408\u7684\u8bdd，\u4f60\u7684\u6743\u91cd\u767d\u8c03\u4e86；\u7528 RRF \u7684\u8bdd，\u4ec0\u4e48\u90fd\u4e0d\u7528\u6539\u3002<\/p>\n

**\u7b2c\u4e09，\u5b83\u5bf9\u300c\u53ea\u5728\u4e00\u8def\u51fa\u73b0\u300d\u7684\u5019\u9009\u662f\u516c\u5e73\u7684\u3002**\u770b k = 60 \u65f6\u7684\u5177\u4f53\u6570\u503c：<\/p>\n\n\u60c5\u5f62\u8ba1\u7b97RRF \u5206\u6570<\/tr>\n\n\n\n\n\n\n
\u4e24\u8def\u90fd\u6392\u7b2c 1<\/td>\n1\/61 + 1\/61<\/td>\n0.03279<\/td>\n<\/tr>\n
dense \u7b2c 1，sparse \u7b2c 5<\/td>\n1\/61 + 1\/65<\/td>\n0.03177<\/td>\n<\/tr>\n
dense \u7b2c 1，sparse \u672a\u547d\u4e2d<\/td>\n1\/61 + 0<\/td>\n0.01639<\/td>\n<\/tr>\n
\u4e24\u8def\u90fd\u6392\u7b2c 10<\/td>\n1\/70 + 1\/70<\/td>\n0.02857<\/td>\n<\/tr>\n
dense \u7b2c 2，sparse \u7b2c 3<\/td>\n1\/62 + 1\/63<\/td>\n0.03199<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u6ce8\u610f\u7b2c 3 \u884c\u548c\u7b2c 4 \u884c：**\u300c\u4e24\u8def\u90fd\u6392\u7b2c 10\u300d\u7684\u5206\u6570（0.02857）\u9ad8\u4e8e\u300c\u53ea\u5728\u4e00\u4e2a\u8def\u6392\u7b2c 1\u300d（0.01639）\u3002**\u8fd9\u6b63\u662f\u6211\u4eec\u60f3\u8981\u7684\u2014\u2014\u591a\u6e90\u5171\u8bc6\u6bd4\u5355\u8def\u9ad8\u5206\u66f4\u53ef\u4fe1\u3002\u540c\u65f6\u4e5f\u6ce8\u610f\u5230\u7b2c 2 \u884c\u548c\u7b2c 5 \u884c\u5f88\u63a5\u8fd1（0.03177 \u5bf9 0.03199），\u8bf4\u660e RRF \u5bf9 top-1 \u4e4b\u5916\u7684\u6392\u540d\u5dee\u5f02\u4e0d\u592a\u654f\u611f，\u8fd9\u662f\u5b83\u7684\u4e00\u4e2a\u7279\u6027：\u5b83\u66f4\u770b\u91cd\u300c\u6709\u6ca1\u6709\u51fa\u73b0\u300d\u548c\u300c\u5927\u81f4\u9760\u524d\u300d，\u4e0d\u770b\u91cd\u7cbe\u786e\u7684\u7b2c\u51e0\u540d\u3002<\/p>\n

k \u7684\u4f5c\u7528\u662f\u4e00\u4e2a\u5e73\u6ed1\u53c2\u6570\u3002k \u8d8a\u5c0f，top-1 \u7684\u6743\u91cd\u8d8a\u5927（k = 1 \u65f6，\u7b2c 1 \u540d\u8d21\u732e 0.5，\u7b2c 2 \u540d\u8d21\u732e 0.33，\u5dee\u8ddd\u60ac\u6b8a）；k \u8d8a\u5927，\u6392\u540d\u5dee\u5f02\u88ab\u62b9\u5f97\u8d8a\u5e73（k = 1000 \u65f6，\u7b2c 1 \u540d\u548c\u7b2c 10 \u540d\u51e0\u4e4e\u6ca1\u533a\u522b）\u3002\u539f\u8bba\u6587\u7684 60 \u662f\u5728 TREC \u6570\u636e\u96c6\u4e0a\u8c03\u51fa\u6765\u7684，\u6211\u5728\u4f01\u4e1a\u8bed\u6599\u4e0a\u8bd5\u8fc7 20 \/ 60 \/ 100，\u5dee\u522b\u5f88\u5c0f，\u9ed8\u8ba4\u5c31\u7528 60，\u522b\u5728\u8fd9\u4e0a\u9762\u82b1\u65f6\u95f4\u3002<\/p>\n

3.3 \u591a\u5c42\u53bb\u91cd：\u4e09\u5c42，\u4e00\u5c42\u90fd\u4e0d\u80fd\u5c11<\/h4>\n

\u4f01\u4e1a\u8bed\u6599\u7684\u91cd\u53e0\u7a0b\u5ea6\u8fdc\u8d85\u9884\u671f\u3002\u540c\u4e00\u4efd\u4ea7\u54c1\u624b\u518c\u4f1a\u5728\u5b98\u7f51\u3001\u5185\u90e8 Wiki\u3001PDF \u5bfc\u51fa\u7248\u3001\u4ee5\u53ca\u67d0\u6b21\u57f9\u8bad\u7684\u5e7b\u706f\u7247\u91cc\u5404\u5b58\u4e00\u4efd，\u63aa\u8f9e\u7565\u6709\u4e0d\u540c\u4f46\u5185\u5bb9\u4e00\u81f4\u3002\u5355\u8def\u53bb\u91cd（\u6309 doc_id）\u5b8c\u5168\u4e0d\u591f，\u5fc5\u987b\u4e09\u5c42：<\/p>\n\n\u5c42\u5224\u636e\u6293\u4ec0\u4e48\u4ee3\u4ef7<\/tr>\n\n\n\n\n
L1 \u552f\u4e00 ID<\/td>\nchunk_id \/ doc_id \u7cbe\u786e\u76f8\u7b49<\/td>\n\u5b8c\u5168\u91cd\u590d\u7684\u5165\u5e93\u8bb0\u5f55<\/td>\nO(1)，\u54c8\u5e0c\u8868<\/td>\n<\/tr>\n
L2 \u6765\u6e90\u4f4d\u7f6e<\/td>\n(doc_id, section_path, char_start \/\/ 256)<\/td>\n\u540c\u4e00\u6587\u6863\u7684\u76f8\u90bb\/\u91cd\u53e0 chunk<\/td>\nO(1)，\u590d\u5408\u952e<\/td>\n<\/tr>\n
L3 \u5185\u5bb9\u6307\u7eb9<\/td>\n\u5f52\u4e00\u5316\u6587\u672c（\u5c0f\u5199\u3001\u53bb\u7a7a\u767d\u3001\u53bb\u6807\u70b9）\u540e\u53d6 hash \u6216 simhash \u6c49\u660e\u8ddd\u79bb <= 3<\/td>\n\u8de8\u6587\u6863\u590d\u5236\u3001\u683c\u5f0f\u8f6c\u6362\u540e\u7684\u540c\u4e00\u6bb5\u5185\u5bb9<\/td>\nO(n)，\u9700\u8981\u9884\u8ba1\u7b97<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

L2 \u7684 char_start \/\/ 256 \u8fd9\u4e2a\u8bbe\u8ba1\u503c\u5f97\u89e3\u91ca\u3002\u7528 256 \u5b57\u7b26\u4f5c\u4e3a\u4e00\u4e2a\u6876，\u662f\u4e3a\u4e86\u5904\u7406\u91cd\u53e0\u5207\u5206：\u5f88\u591a\u5207\u5206\u7b56\u7565\u4f1a\u7528 50 \u5230 100 \u5b57\u7b26\u7684\u91cd\u53e0\u7a97\u53e3\u6765\u4fdd\u8bc1\u8bed\u4e49\u8fde\u8d2f，\u4e8e\u662f\u540c\u4e00\u4e2a\u53e5\u5b50\u4f1a\u8fde\u7eed\u51fa\u73b0\u5728 2 \u5230 3 \u4e2a chunk \u91cc\u3002\u8fd9\u4e9b chunk \u7684 chunk_id \u4e0d\u540c（L1 \u6293\u4e0d\u5230），\u4f46\u5b83\u4eec\u7684\u8d77\u59cb\u4f4d\u7f6e\u843d\u5728\u540c\u4e00\u4e2a 256 \u5b57\u7b26\u6876\u5185\u7684\u6982\u7387\u5f88\u9ad8，L2 \u80fd\u6293\u5230\u5927\u90e8\u5206\u3002<\/p>\n

L3 \u662f\u552f\u4e00\u80fd\u6293\u300c\u8de8\u6587\u6863\u590d\u5236\u300d\u7684\u4e00\u5c42\u3002\u5f52\u4e00\u5316\u5f88\u91cd\u8981：\u53bb\u5927\u5c0f\u5199\u3001\u538b\u7f29\u8fde\u7eed\u7a7a\u767d\u3001\u53bb\u6389\u6240\u6709\u6807\u70b9，\u8fd9\u6837\u300c\u9000\u8d27\u653f\u7b56：7 \u5929\u65e0\u7406\u7531\u300d\u548c\u300c\u9000\u8d27\u653f\u7b56:7\u5929\u65e0\u7406\u7531\u300d\u4f1a\u5f97\u5230\u540c\u4e00\u4e2a\u6307\u7eb9\u3002\u5982\u679c\u8981\u5bb9\u5fcd\u8f7b\u5fae\u6539\u5199，\u7528 simhash \u5e76\u63a5\u53d7\u6c49\u660e\u8ddd\u79bb\u5c0f\u4e8e\u7b49\u4e8e 3 \u7684\u5019\u9009\u2014\u2014\u4f46\u6ce8\u610f\u8fd9\u4f1a\u5f15\u5165\u8bef\u5220，\u6211\u5728\u751f\u4ea7\u4e0a\u53ea\u5728\u300c\u540c\u4e00 doc_id \u5185\u300d\u542f\u7528 simhash \u8fd1\u4f3c\u5339\u914d，\u8de8\u6587\u6863\u4ecd\u7136\u7528\u7cbe\u786e hash\u3002<\/p>\n

3.4 cross-encoder \u91cd\u6392<\/h4>\n

bi-encoder（\u53cc\u7f16\u7801\u5668，\u4e5f\u5c31\u662f dense \u68c0\u7d22\u7528\u7684\u90a3\u79cd）\u628a query \u548c doc \u5206\u522b\u7f16\u7801\u6210\u4e24\u4e2a\u5411\u91cf，\u7136\u540e\u7b97\u76f8\u4f3c\u5ea6\u3002\u597d\u5904\u662f doc \u5411\u91cf\u53ef\u4ee5\u79bb\u7ebf\u9884\u8ba1\u7b97，\u68c0\u7d22\u65f6\u53ea\u7b97\u4e00\u6b21 query \u7f16\u7801，\u6240\u4ee5\u80fd\u505a\u5230\u6beb\u79d2\u7ea7；\u4ee3\u4ef7\u662f query \u548c doc \u4e4b\u95f4\u6ca1\u6709\u4ea4\u4e92，\u7cbe\u5ea6\u6709\u4e0a\u9650\u3002<\/p>\n

cross-encoder \u628a query \u548c doc \u62fc\u6210\u4e00\u4e2a\u5e8f\u5217\u9001\u8fdb\u6a21\u578b，\u8ba9\u5b83\u4eec\u4e92\u76f8\u505a attention\u3002\u7cbe\u5ea6\u663e\u8457\u66f4\u9ad8，\u4ee3\u4ef7\u662f\u4e0d\u80fd\u9884\u8ba1\u7b97，\u6bcf\u4e2a\u5019\u9009\u90fd\u8981\u8dd1\u4e00\u6b21\u524d\u5411，\u6240\u4ee5\u53ea\u80fd\u7528\u5728\u5df2\u7ecf\u7f29\u5c0f\u5230\u51e0\u5341\u6761\u7684\u5019\u9009\u96c6\u4e0a\u3002<\/p>\n

\u6d41\u7a0b：<\/p>\n

top-50（RRF \u540e\u3001\u53bb\u91cd\u540e）
\n -> \u62fc (query, doc) \u5bf9，\u622a\u65ad\u5230 reranker \u7684 max_seq_len（\u901a\u5e38 512 token）
\n -> \u6279\u63a8\u7406，\u62ff\u5230\u76f8\u5173\u6027\u5206\u6570
\n -> \u6309\u5206\u6570\u964d\u5e8f，\u53d6 top-8
\n -> \u6309\u5206\u6570\u6216\u6309\u539f\u59cb\u6765\u6e90\u987a\u5e8f\u7ec4\u88c5\u8fdb\u4e0a\u4e0b\u6587<\/p>\n

\u622a\u65ad\u8fd9\u4e00\u6761\u662f\u5f88\u591a\u95ee\u9898\u7684\u6839\u6e90，\u89c1\u8e29\u5751 7.3\u3002<\/p>\n

3.5 \u7b2c\u4e8c\u4ee3 GraphRAG：\u4e3a\u4ec0\u4e48\u628a chunk \u5f53\u5b64\u7acb\u6587\u672c\u4f1a\u5931\u6548<\/h4>\n

\u6734\u7d20 RAG \u548c\u7b2c\u4e00\u4ee3\u6df7\u5408\u68c0\u7d22\u90fd\u6709\u4e00\u4e2a\u9690\u542b\u5047\u8bbe：**\u6bcf\u4e2a chunk \u662f\u4e00\u6bb5\u81ea\u5305\u542b\u7684\u6587\u672c，\u95ee\u9898\u9700\u8981\u7684\u5168\u90e8\u4fe1\u606f\u90fd\u5728\u67d0\u4e00\u4e2a（\u6216\u51e0\u4e2a）chunk \u91cc\u3002**\u8fd9\u4e2a\u5047\u8bbe\u5728\u4f01\u4e1a\u8bed\u6599\u4e0a\u7ecf\u5e38\u4e0d\u6210\u7acb\u3002<\/p>\n

\u4e09\u4e2a\u5178\u578b\u5931\u6548：<\/p>\n

\u5931\u6548\u4e00：\u5c42\u7ea7\u5173\u7cfb\u4e22\u5931\u3002\u300cAX-2039 \u5c5e\u4e8e\u54ea\u4e2a\u4ea7\u54c1\u7ebf\u300d\u8fd9\u4e2a\u95ee\u9898，\u7b54\u6848\u53ef\u80fd\u5728\u4e00\u4e2a\u8bb2\u4ea7\u54c1\u7ebf\u7684\u6587\u6863\u91cc，\u800c\u8bb2 AX-2039 \u7684\u90a3\u4e2a chunk \u53ea\u8bf4\u300c\u672c\u4ea7\u54c1\u5c5e\u4e8e AX-2000 \u7cfb\u5217\u300d，\u6ca1\u6709\u8bf4 AX-2000 \u7cfb\u5217\u7684\u4e0a\u7ea7\u662f\u4ec0\u4e48\u3002\u8981\u56de\u7b54\u5c31\u5f97\u6cbf\u7740\u300cAX-2039 -> AX-2000 \u7cfb\u5217 -> \u4f01\u4e1a\u7ea7\u4ea7\u54c1\u7ebf\u300d\u8fd9\u6761\u94fe\u8d70，\u800c\u8fd9\u6761\u94fe\u6563\u843d\u5728\u4e09\u4e2a\u6587\u6863\u91cc\u3002<\/p>\n

**\u5931\u6548\u4e8c：\u540c\u4e49\u8bcd\u4e0d\u8fde\u901a\u3002**\u7528\u6237\u95ee\u300c\u5ba2\u6237\u4e3b\u6570\u636e\u600e\u4e48\u540c\u6b65\u300d，\u8bed\u6599\u91cc\u901a\u7bc7\u5199\u7684\u662f CUST_MSTR\u3002\u6ca1\u6709\u4efb\u4f55\u4e00\u4e2a chunk \u540c\u65f6\u51fa\u73b0\u8fd9\u4e24\u4e2a\u8bcd，\u6240\u4ee5\u4efb\u4f55\u68c0\u7d22\u90fd\u8fde\u4e0d\u8d77\u6765\u3002\u89e3\u51b3\u529e\u6cd5\u662f\u663e\u5f0f\u5efa\u4e00\u6761\u540c\u4e49\u5173\u7cfb\u8fb9\u3002<\/p>\n

\u5931\u6548\u4e09：\u5171\u4eab\u7236\u7c7b\u95ee\u9898\u3002\u300cA \u4ea7\u54c1\u548c B \u4ea7\u54c1\u90fd\u652f\u6301\u54ea\u4e9b\u8ba4\u8bc1\u300d\u2014\u2014A \u548c B \u7684\u6587\u6863\u5404\u81ea\u5217\u4e86\u81ea\u5df1\u7684\u8ba4\u8bc1，\u4f46\u4e24\u8005\u5171\u540c\u7684\u7236\u7c7b\u300c\u4f01\u4e1a\u7ea7\u4ea7\u54c1\u300d\u4e0a\u8fd8\u6709\u4e00\u6761\u300c\u5168\u90e8\u901a\u8fc7 SOC2\u300d\u7684\u8bf4\u660e\u3002\u4e0d\u987a\u7740\u7236\u7c7b\u8d70，\u5c31\u4f1a\u6f0f\u6389 SOC2，\u7136\u540e\u7b54\u9519\u3002<\/p>\n

GraphRAG \u7684\u505a\u6cd5\u662f\u628a\u8fd9\u4e00\u5c42\u663e\u5f0f\u5efa\u51fa\u6765：<\/p>\n

\u672c\u4f53\u5c42（ontology） \u4ea7\u54c1 -> \u4ea7\u54c1\u7ebf -> \u4ea7\u54c1\u65cf；\u7cfb\u7edf -> \u6a21\u5757 -> \u63a5\u53e3
\n |
\n\u5b9e\u4f53\u5c42（entity） AX-2039\u3001CUST_MSTR\u3001SOC2\u3001SLA-99.9
\n |
\n\u5173\u7cfb\u5c42（relation） part_of \/ synonym_of \/ depends_on \/ implements \/ supersedes
\n |
\nchunk \u5c42 chunk \u901a\u8fc7 mention \u8fb9\u6302\u5230\u5b9e\u4f53\u4e0a<\/p>\n

\u68c0\u7d22\u65f6\u4e0d\u662f\u53ea\u53ec\u56de chunk，\u800c\u662f\u300c\u53ec\u56de chunk -> \u627e\u5230 chunk \u4e0a\u7684\u5b9e\u4f53 -> \u6cbf\u5173\u7cfb\u8fb9\u6269\u5c55\u5230\u90bb\u57df\u5b9e\u4f53 -> \u518d\u53d6\u8fd9\u4e9b\u5b9e\u4f53\u5173\u8054\u7684 chunk\u300d\u3002\u8fd9\u6837\u4e00\u6b21\u67e5\u8be2\u80fd\u987a\u7740\u56fe\u8d70 2 \u5230 3 \u8df3，\u628a\u6563\u843d\u7684\u524d\u63d0\u51d1\u9f50\u3002<\/p>\n

3.6 \u5173\u952e\u5de5\u7a0b\u5224\u65ad：\u89c4\u5219\u5316\u591a\u8d9f\u5b9e\u4f53\u62bd\u53d6\u4f18\u4e8e LLM NER<\/h4>\n

\u8fd9\u662f\u672c\u6587\u6211\u6700\u60f3\u5f3a\u8c03\u3001\u4e5f\u6700\u5bb9\u6613\u548c\u4e3b\u6d41\u8bf4\u6cd5\u51b2\u7a81\u7684\u4e00\u8282\u3002<\/p>\n

\u6784\u5efa\u56fe\u8c31\u7684\u7b2c\u4e00\u6b65\u662f\u5b9e\u4f53\u62bd\u53d6（NER，Named Entity Recognition，\u547d\u540d\u5b9e\u4f53\u8bc6\u522b）\u30022025 \u5e74\u4e4b\u540e\u7684\u7edd\u5927\u591a\u6570\u6559\u7a0b\u4f1a\u544a\u8bc9\u4f60：\u7528 LLM \u505a NER，\u96f6\u6837\u672c\u3001\u6cdb\u5316\u597d\u3001\u4e0d\u7528\u6807\u6ce8\u3002\u6211\u5728\u4e00\u4e2a\u5c0f\u89c4\u6a21\u8bd5\u70b9\u4e0a\u4e5f\u8fd9\u4e48\u505a\u8fc7，\u7136\u540e\u5728\u751f\u4ea7\u89c4\u6a21\u4e0a\u63a8\u7ffb\u4e86\u8fd9\u4e2a\u51b3\u5b9a\u3002<\/p>\n

\u5148\u7b97\u4e00\u7b14\u8d26\u3002\u5047\u8bbe\u8bed\u6599\u662f 8 \u4e07 chunk，\u6bcf chunk \u5e73\u5747 380 token：<\/p>\n\n\u65b9\u6848\u5355\u6b21\u8017\u65f6\u603b\u8017\u65f6\u8fb9\u9645\u6210\u672c\u786e\u5b9a\u6027\u53ef\u590d\u73b0<\/tr>\n\n\n\n
LLM NER，\u9010 chunk \u8c03\u7528，\u5e76\u53d1 16<\/td>\n600~1200ms<\/td>\n80000 \/ 16 x 0.9s \u7ea6 75 \u5206\u949f<\/td>\n\u6bcf\u6b21\u8c03\u7528\u51e0\u767e\u6beb\u79d2 + token \u8d39\u7528，8 \u4e07\u6b21\u662f\u4e00\u7b14\u771f\u91d1\u767d\u94f6<\/td>\n\u975e\u786e\u5b9a\u6027，\u540c\u4e00 chunk \u4e24\u6b21\u8dd1\u53ef\u80fd\u62bd\u51fa\u4e0d\u540c\u5b9e\u4f53<\/td>\n\u5dee，\u9700\u8981\u989d\u5916\u505a\u4e00\u81f4\u6027\u6536\u655b<\/td>\n<\/tr>\n
\u89c4\u5219\u5316\u591a\u8d9f\u5339\u914d<\/td>\n0.3~2ms（\u5fae\u79d2\u5230\u6beb\u79d2\u7ea7）<\/td>\n80000 x 1.2ms \u7ea6 96 \u79d2<\/td>\n\u96f6<\/td>\n\u786e\u5b9a\u6027，\u540c\u8f93\u5165\u5fc5\u540c\u8f93\u51fa<\/td>\n\u5b8c\u5168\u4e00\u81f4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

75 \u5206\u949f\u5bf9 96 \u79d2，\u4e24\u4e2a\u6570\u91cf\u7ea7\u3002\u4f46\u8fd9\u8fd8\u4e0d\u662f\u4e3b\u8981\u7406\u7531\u3002\u4e3b\u8981\u7406\u7531\u6709\u4e24\u6761：<\/p>\n

**\u7b2c\u4e00\u6761，\u5165\u5e93\u6210\u672c\u662f\u6301\u7eed\u7684，\u4e0d\u662f\u4e00\u6b21\u7684\u3002**\u8bed\u6599\u4f1a\u66f4\u65b0，\u6bcf\u6b21\u589e\u91cf\u5165\u5e93\u90fd\u8981\u91cd\u8dd1\u62bd\u53d6\u3002LLM \u65b9\u6848\u4e0b，\u4e00\u6b21\u5168\u91cf\u91cd\u8dd1\u662f\u4e00\u7b14\u8d26\u5355，\u4e8e\u662f\u56e2\u961f\u4f1a\u672c\u80fd\u5730\u63a8\u8fdf\u91cd\u8dd1，\u56fe\u8c31\u5c31\u6162\u6162\u8fc7\u671f\u3002\u89c4\u5219\u5316\u65b9\u6848\u4e0b\u91cd\u8dd1\u51e0\u4e4e\u514d\u8d39，\u53ef\u4ee5\u653e\u8fdb\u6bcf\u6b21 CI\u3002<\/p>\n

**\u7b2c\u4e8c\u6761，\u975e\u786e\u5b9a\u6027\u4f1a\u6c61\u67d3\u4e0b\u6e38\u4e14\u65e0\u6cd5\u8c03\u8bd5\u3002**LLM \u8fd9\u6b21\u628a\u300cCRM\u300d\u62bd\u6210\u4e86\u5b9e\u4f53，\u4e0b\u6b21\u6ca1\u62bd\u51fa\u6765，\u4e8e\u662f\u56fe\u8c31\u91cc\u65f6\u6709\u65f6\u65e0\u3002\u4f60\u8981\u67e5\u300c\u4e3a\u4ec0\u4e48\u8fd9\u4e2a\u67e5\u8be2\u6628\u5929\u80fd\u7b54\u4eca\u5929\u7b54\u4e0d\u4e86\u300d，\u6700\u540e\u8ffd\u5230\u4e00\u4e2a\u65e0\u6cd5\u590d\u73b0\u7684\u62bd\u53d6\u5dee\u5f02\u4e0a\u2014\u2014\u8fd9\u7c7b\u95ee\u9898\u5728\u751f\u4ea7\u4e0a\u6781\u96be\u6392\u67e5\u3002<\/p>\n

\u89c4\u5219\u5316\u591a\u8d9f\u5339\u914d\u7684\u5177\u4f53\u505a\u6cd5，\u4e09\u8d9f，\u4ece\u7cbe\u786e\u5230\u5bbd\u677e：<\/p>\n

Pass 1 \u6700\u957f\u77ed\u8bed\u4f18\u5148\u5339\u914d（longest-phrase-first）
\n \u8bcd\u5178\u6309\u77ed\u8bed\u957f\u5ea6\u964d\u5e8f\u6392\u5217，\u7528 Aho-Corasick \u4e4b\u7c7b\u7684\u591a\u6a21\u5f0f\u5339\u914d\u4e00\u6b21\u626b\u8fc7\u6587\u672c，
\n \u547d\u4e2d\u5373\u5360\u636e\u5b57\u7b26\u533a\u95f4，\u540e\u7eed\u66f4\u77ed\u7684\u5019\u9009\u82e5\u533a\u95f4\u91cd\u53e0\u5219\u8df3\u8fc7\u3002
\n \u76ee\u7684：\u4fdd\u8bc1 "\u5ba2\u6237\u4e3b\u6570\u636e\u7ba1\u7406\u7cfb\u7edf" \u4e0d\u4f1a\u88ab\u5207\u6210 "\u5ba2\u6237" + "\u4e3b\u6570\u636e" + "\u7cfb\u7edf"<\/p>\n

Pass 2 \u5f52\u4e00\u5316\u5339\u914d（normalized）
\n \u5bf9\u6587\u672c\u505a\u5f52\u4e00\u5316（\u5c0f\u5199\u3001\u5168\u89d2\u8f6c\u534a\u89d2\u3001\u53bb\u8fde\u5b57\u7b26\u3001\u53bb\u7a7a\u683c）\u540e\u518d\u5339\u914d\u3002
\n \u76ee\u7684：\u6293\u4f4f "CUST-MSTR" \/ "CUST_MSTR" \/ "Cust Mstr" \u8fd9\u7c7b\u5199\u6cd5\u5dee\u5f02<\/p>\n

Pass 3 token \u7ea7\u5339\u914d（token-level）
\n \u5728\u5206\u8bcd\u540e\u7684 token \u5e8f\u5217\u4e0a\u5339\u914d，\u5141\u8bb8\u987a\u5e8f\u98a0\u5012\u4e0e\u63d2\u5165\u8bcd\u3002
\n \u76ee\u7684：\u6293\u4f4f "\u4e3b\u6570\u636e\u7684\u5ba2\u6237\u540c\u6b65" \u8fd9\u7c7b\u8bed\u5e8f\u53d8\u5316<\/p>\n

\u4e09\u8d9f\u7684\u7ed3\u679c\u5408\u5e76\u65f6\u6309\u300c\u6700\u957f\u547d\u4e2d\u4f18\u5148\u3001\u540c\u957f\u5ea6\u53d6\u5148\u51fa\u73b0\u300d\u7684\u89c4\u5219\u53bb\u91cd，\u5e76\u8bb0\u5f55\u6bcf\u4e00\u8d9f\u7684\u547d\u4e2d\u6765\u6e90，\u65b9\u4fbf\u56de\u6eaf\u3002<\/p>\n

\u9700\u8981\u8bf4\u6e05\u695a\u8fb9\u754c：**\u89c4\u5219\u5316\u65b9\u6848\u7684\u524d\u63d0\u662f\u4f60\u6709\u4e00\u4e2a\u53ef\u63a7\u7684\u5b9e\u4f53\u8bcd\u5178\u3002**\u4f01\u4e1a\u573a\u666f\u6070\u6070\u6ee1\u8db3\u8fd9\u4e2a\u524d\u63d0\u2014\u2014\u4ea7\u54c1\u7801\u3001\u7cfb\u7edf\u540d\u3001\u9519\u8bef\u7801\u3001\u90e8\u95e8\u540d\u3001\u5ba2\u6237\u540d，\u8fd9\u4e9b\u90fd\u662f\u6709\u9650\u96c6\u5408，\u800c\u4e14\u901a\u5e38\u5df2\u7ecf\u5b58\u5728\u4e8e CMDB（\u914d\u7f6e\u7ba1\u7406\u6570\u636e\u5e93）\u3001\u4ea7\u54c1\u76ee\u5f55\u3001\u6216\u8005\u6570\u636e\u5e93\u7684\u4e3b\u6570\u636e\u8868\u91cc\u3002\u76f4\u63a5\u4ece\u8fd9\u4e9b\u6743\u5a01\u6e90\u5bfc\u51fa\u8bcd\u5178，\u6bd4\u8ba9 LLM \u731c\u8981\u9760\u8c31\u5f97\u591a\u3002<\/p>\n

\u6211\u7684\u5b9e\u9645\u505a\u6cd5\u662f\u6df7\u5408：\u89c4\u5219\u5316\u4e09\u8d9f\u8d1f\u8d23 90% \u4ee5\u4e0a\u7684\u5df2\u77e5\u5b9e\u4f53；LLM \u53ea\u5728\u300c\u4f4e\u9891\u957f\u5c3e + \u89c4\u5219\u5b8c\u5168\u6ca1\u547d\u4e2d + \u8be5 chunk \u88ab\u9ad8\u9891\u68c0\u7d22\u300d\u8fd9\u4e09\u4e2a\u6761\u4ef6\u540c\u65f6\u6210\u7acb\u65f6\u624d\u4ecb\u5165，\u800c\u4e14\u7ed3\u679c\u8981\u8fdb\u4eba\u5de5\u5ba1\u6838\u961f\u5217，\u4e0d\u76f4\u63a5\u5199\u56fe\u3002\u8fd9\u6837 LLM \u7684\u8c03\u7528\u91cf\u964d\u5230\u5168\u91cf\u65b9\u6848\u7684 2% \u4ee5\u4e0b，\u4e14\u975e\u786e\u5b9a\u6027\u88ab\u5173\u5728\u5ba1\u6838\u95f8\u95e8\u540e\u9762\u3002<\/p>\n

3.7 \u7b2c\u4e09\u4ee3 Agentic：\u4e94\u6761\u7eaa\u5f8b\u4e0e\u5b89\u5168\u8fb9\u754c<\/h4>\n

Agentic RAG \u4e0d\u662f\u300c\u7ed9 RAG \u52a0\u4e00\u4e2a Agent\u300d，\u5b83\u662f\u628a\u68c0\u7d22\u4ece\u4e00\u4e2a\u51fd\u6570\u8c03\u7528\u5347\u7ea7\u6210\u4e00\u4e2a\u53d7\u63a7\u7684\u51b3\u7b56\u8fc7\u7a0b\u3002\u4e94\u6761\u7eaa\u5f8b，\u7f3a\u4e00\u6761\u5c31\u4e0d\u53ef\u63a7：<\/p>\n\n\u7eaa\u5f8b\u8981\u89e3\u51b3\u7684\u95ee\u9898\u673a\u5236\u8d8a\u754c\u540e\u679c<\/tr>\n\n\n\n\n\n\n
\u5de5\u5177\u4f7f\u7528<\/td>\n\u4ec0\u4e48\u65f6\u5019\u68c0\u7d22\u3001\u7528\u4ec0\u4e48\u68c0\u7d22<\/td>\n\u5de5\u5177\u6ce8\u518c\u8868 + \u53c2\u6570 schema \u6821\u9a8c<\/td>\n\u68c0\u7d22\u4e86\u4e0d\u8be5\u68c0\u7d22\u7684\u5e93，\u6216\u53c2\u6570\u88ab\u6ce8\u5165\u6539\u5199<\/td>\n<\/tr>\n
\u8bb0\u5fc6\u7ba1\u7406<\/td>\n\u591a\u8f6e\u4e4b\u95f4\u5e26\u4ec0\u4e48<\/td>\n\u5206\u5c42\u8bb0\u5fc6：\u4efb\u52a1\u5185\u5de5\u4f5c\u8bb0\u5fc6 \/ \u4f1a\u8bdd\u7ea7 \/ \u957f\u671f\u7528\u6237\u753b\u50cf<\/td>\n\u4e0a\u4e0b\u6587\u65e0\u9650\u81a8\u80c0，\u6216\u4e0a\u4e00\u8f6e\u7684\u810f\u4e0a\u4e0b\u6587\u6c61\u67d3\u4e0b\u4e00\u8f6e<\/td>\n<\/tr>\n
\u89c4\u5212<\/td>\n\u591a\u8df3\u95ee\u9898\u600e\u4e48\u62c6<\/td>\n\u663e\u5f0f plan（\u8ba1\u5212）\u5bf9\u8c61，\u6bcf\u6b65\u53ef\u679a\u4e3e\u3001\u53ef\u4e2d\u65ad<\/td>\n\u65e0\u9650\u5faa\u73af，\u6216\u62c6\u51fa\u4e0d\u53ef\u80fd\u5b8c\u6210\u7684\u5b50\u76ee\u6807<\/td>\n<\/tr>\n
\u534f\u8c03<\/td>\n\u591a\u4e2a\u68c0\u7d22\u5668\u600e\u4e48\u914d\u5408<\/td>\n\u4e32\u884c \/ \u5e76\u884c \/ \u4f9d\u8d56\u56fe，\u5e26\u8d85\u65f6\u4e0e\u964d\u7ea7<\/td>\n\u4e00\u4e2a\u6162\u68c0\u7d22\u62d6\u57ae\u6574\u6761\u94fe\u8def<\/td>\n<\/tr>\n
\u8bc4\u4f30<\/td>\n\u8fd9\u4e00\u6b65\u6210\u6ca1\u6210<\/td>\n\u6bcf\u6b65\u7684 grounded \u5224\u5b9a + \u4e58\u6cd5\u7f6e\u4fe1\u5ea6<\/td>\n\u7528\u70c2\u68c0\u7d22\u7ed3\u679c\u786c\u751f\u6210，\u8f93\u51fa\u770b\u4f3c\u6709\u636e\u5b9e\u5219\u7f16\u9020<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

**\u5b89\u5168\u8fb9\u754c\u5fc5\u987b\u524d\u7f6e\u5230\u68c0\u7d22\u4e4b\u524d\u3002**\u8fd9\u4e00\u6761\u6211\u8981\u5355\u72ec\u62ce\u51fa\u6765\u8bb2，\u56e0\u4e3a\u5b83\u7ecf\u5e38\u88ab\u653e\u5728\u6700\u540e\u3002<\/p>\n

\u4f01\u4e1a\u67e5\u8be2\u91cc\u6709\u76f8\u5f53\u4e00\u90e8\u5206\u4f1a\u89e6\u53ca\u5ba2\u6237\u6570\u636e\u3001\u5458\u5de5\u4fe1\u606f\u3001\u6d89\u5bc6\u5f15\u7528\u3002\u6b63\u786e\u7684\u5904\u7406\u662f\u5728\u68c0\u7d22\u6216\u751f\u6210\u4e4b\u524d\u5c31\u62e6\u622a，\u800c\u4e0d\u662f\u300c\u68c0\u7d22\u5b8c\u4e4b\u540e\u518d\u8fc7\u6ee4\u300d\u2014\u2014\u540e\u8005\u7684\u6f0f\u6d1e\u5728\u4e8e，\u68c0\u7d22\u8fc7\u7a0b\u672c\u8eab\u5df2\u7ecf\u628a\u654f\u611f\u5185\u5bb9\u8bfb\u8fdb\u4e86\u4e0a\u4e0b\u6587，\u4e00\u65e6\u4e0a\u4e0b\u6587\u91cc\u6709\u4e86，\u540e\u9762\u7684\u8fc7\u6ee4\u5c31\u53ea\u662f\u5c3d\u529b\u800c\u4e3a，\u6a21\u578b\u5b8c\u5168\u53ef\u80fd\u5728\u88ab\u8fc7\u6ee4\u4e4b\u524d\u5c31\u628a\u5b83\u5199\u8fdb\u4e86\u8f93\u51fa\u3002<\/p>\n

\u8fb9\u754c\u7684\u5b9e\u73b0\u5206\u4e09\u9053：<\/p>\n

\u7b2c 1 \u9053 \u67e5\u8be2\u4fa7\u5206\u7c7b\u5668（\u786e\u5b9a\u6027\u4f18\u5148）
\n \u6b63\u5219 + \u8bcd\u5178：\u8eab\u4efd\u8bc1 \/ \u624b\u673a\u53f7 \/ \u94f6\u884c\u5361 \/ \u90ae\u7bb1 \/ "\u85aa\u8d44" \/ "\u7ee9\u6548" \/ "\u5ba2\u6237\u540d\u5355"
\n \u547d\u4e2d -> \u8d70\u6743\u9650\u6821\u9a8c，\u4e0d\u547d\u4e2d -> \u653e\u884c
\n\u7b2c 2 \u9053 \u6743\u9650\u6821\u9a8c（\u5728\u68c0\u7d22\u5668\u5185\u90e8，\u4e0d\u53ef\u7ed5\u8fc7）
\n \u7528\u6237 ACL \u4e0e\u6587\u6863 ACL \u6c42\u4ea4\u96c6，\u68c0\u7d22\u5668\u53ea\u8fd4\u56de\u4ea4\u96c6\u5185\u7684 chunk
\n \u8fd9\u4e00\u9053\u5fc5\u987b\u5728\u68c0\u7d22\u5668\u91cc，\u4e0d\u80fd\u5728\u8c03\u7528\u65b9，\u5426\u5219\u6362\u4e00\u4e2a\u8c03\u7528\u65b9\u5c31\u7ed5\u8fc7\u4e86
\n\u7b2c 3 \u9053 \u751f\u6210\u524d\u590d\u6838
\n \u7ec4\u88c5\u597d\u7684\u4e0a\u4e0b\u6587\u518d\u8fc7\u4e00\u904d\u5206\u7c7b\u5668，\u786e\u8ba4\u6ca1\u6709\u654f\u611f\u6bb5\u843d\u6f0f\u7f51
\n \u515c\u5e95，\u9632\u7684\u662f\u7b2c 1\u30012 \u9053\u7684\u89c4\u5219\u6ca1\u8986\u76d6\u5230\u7684\u5f62\u6001<\/p>\n

\u4e09\u9053\u91cc\u7b2c 2 \u9053\u662f\u786c\u4fdd\u969c，\u7b2c 1\u30013 \u9053\u662f\u964d\u4f4e\u8bef\u653e\u884c\u7684\u6982\u7387\u3002\u53ea\u6709\u7b2c 1 \u548c\u7b2c 3 \u9053\u662f\u4e0d\u591f\u7684，\u56e0\u4e3a\u5b83\u4eec\u662f\u6982\u7387\u6027\u7684\u3002<\/p>\n

3.8 \u7f6e\u4fe1\u5ea6\u7528\u4e58\u6cd5\u800c\u4e0d\u662f\u5e73\u5747<\/h4>\n

\u8fd9\u662f\u672c\u8282\u7684\u6838\u5fc3，\u4e5f\u662f\u6211\u6700\u60f3\u4f20\u64ad\u7684\u4e00\u6761\u5de5\u7a0b\u7ecf\u9a8c\u3002<\/p>\n

\u4e00\u4e2a\u591a\u6b65 Agentic \u6d41\u7a0b\u91cc，\u6700\u7ec8\u7b54\u6848\u7684\u53ef\u4fe1\u5ea6\u53d6\u51b3\u4e8e\u6bcf\u4e00\u6b65\u3002\u5047\u8bbe\u68c0\u7d22 grounded \u5ea6\u662f 0.9，\u751f\u6210\u5fe0\u5b9e\u5ea6\u662f 0.9：<\/p>\n

\u4e58\u6cd5： 0.9 x 0.9 = 0.81
\n\u5e73\u5747： (0.9 + 0.9) \/ 2 = 0.9<\/p>\n

\u770b\u8d77\u6765\u5dee\u522b\u4e0d\u5927\u3002\u6362\u4e00\u4e2a\u573a\u666f：\u68c0\u7d22 grounded \u5ea6\u662f 0.9，\u4f46\u751f\u6210\u5fe0\u5b9e\u5ea6\u53ea\u6709 0.1（\u6a21\u578b\u57fa\u672c\u5728\u81ea\u7531\u53d1\u6325）：<\/p>\n

\u4e58\u6cd5： 0.9 x 0.1 = 0.09 <- \u6b63\u786e\u7b54\u6848：\u8fd9\u4e00\u6b65\u51e0\u4e4e\u4e0d\u53ef\u4fe1
\n\u5e73\u5747： (0.9 + 0.1) \/ 2 = 0.5 <- \u8bef\u5bfc：\u770b\u8d77\u6765\u300c\u8fd8\u884c\u300d<\/p>\n

**\u5e73\u5747\u4f1a\u628a\u4e00\u4e2a\u81f4\u547d\u7684\u5f31\u73af\u8282\u63a9\u76d6\u6210\u4e00\u4e2a\u4e2d\u7b49\u5206\u6570\u3002**0.5 \u8fd9\u4e2a\u6570\u5b57\u4f1a\u8ba9\u4f60\u89c9\u5f97\u300c\u4e5f\u8bb8\u80fd\u7528\u300d，\u800c 0.09 \u4f1a\u8ba9\u4f60\u7acb\u523b\u89e6\u53d1\u91cd\u8bd5\u3002\u8fd9\u5c31\u662f\u6211\u8981\u7528\u4e58\u6cd5\u7684\u5168\u90e8\u7406\u7531：\u7f6e\u4fe1\u5ea6\u7684\u4f5c\u7528\u662f\u9a71\u52a8\u51b3\u7b56（\u8981\u4e0d\u8981\u91cd\u8bd5\u3001\u8981\u4e0d\u8981\u62d2\u7b54\u3001\u8981\u4e0d\u8981\u964d\u7ea7），\u4e00\u4e2a\u63a9\u76d6\u5f31\u73af\u8282\u7684\u51b3\u7b56\u4fe1\u53f7\u6bd4\u6ca1\u6709\u4fe1\u53f7\u66f4\u5371\u9669\u3002<\/p>\n

\u4e09\u4e2a\u5b9e\u9645\u4f7f\u7528\u4e2d\u7684\u6ce8\u610f\u70b9：<\/p>\n

  • **\u56e0\u5b50\u5fc5\u987b\u72ec\u7acb\u3002**\u5982\u679c\u4e24\u4e2a\u56e0\u5b50\u6d4b\u7684\u662f\u540c\u4e00\u4ef6\u4e8b（\u6bd4\u5982\u300c\u68c0\u7d22\u76f8\u4f3c\u5ea6\u300d\u548c\u300crerank \u5206\u6570\u300d），\u4e58\u8d77\u6765\u4f1a\u628a\u540c\u4e00\u4e2a\u8bc1\u636e\u7b97\u4e24\u904d，\u7f6e\u4fe1\u5ea6\u88ab\u7cfb\u7edf\u6027\u4f4e\u4f30\u3002\u53ea\u4e58\u771f\u6b63\u72ec\u7acb\u7684\u73af\u8282：\u68c0\u7d22\u8d28\u91cf\u3001\u751f\u6210\u5fe0\u5b9e\u5ea6\u3001\u5de5\u5177\u6210\u529f\u7387\u3001\u6743\u9650\u5408\u89c4\u3002<\/li>\n
  • **\u56e0\u5b50\u4e2a\u6570\u5f71\u54cd\u9608\u503c\u3002**\u4e32\u4e86 4 \u4e2a 0.9 \u7684\u56e0\u5b50，\u4e58\u79ef\u662f 0.656\u3002\u6240\u4ee5\u9608\u503c\u4e0d\u80fd\u5199\u6b7b\u6210 0.8，\u5e94\u8be5\u6309\u300c\u56e0\u5b50\u4e2a\u6570\u300d\u5f52\u4e00\u5316，\u6216\u8005\u5e72\u8106\u5bf9\u6bcf\u4e2a\u56e0\u5b50\u5355\u72ec\u8bbe\u4e0b\u9650（\u4efb\u4e00\u56e0\u5b50\u4f4e\u4e8e 0.3 \u5c31\u76f4\u63a5\u62d2\u7b54），\u518d\u5bf9\u4e58\u79ef\u8bbe\u603b\u9608\u503c\u3002<\/li>\n
  • **0 \u5c31\u662f 0\u3002**\u6743\u9650\u6821\u9a8c\u5931\u8d25\u3001\u68c0\u7d22\u96f6\u547d\u4e2d\u3001\u5de5\u5177\u8c03\u7528\u5f02\u5e38，\u8fd9\u4e9b\u5e94\u8be5\u662f 0 \u800c\u4e0d\u662f 0.1\u3002\u4efb\u4f55\u975e\u96f6\u503c\u90fd\u4f1a\u8ba9\u300c\u5b8c\u5168\u6ca1\u4f9d\u636e\u300d\u548c\u300c\u6709\u70b9\u4f9d\u636e\u300d\u5728\u4ea7\u54c1\u4e0a\u65e0\u6cd5\u533a\u5206\u3002<\/li>\n

    \u6709\u754c\u91cd\u8bd5\u7684\u72b6\u6001\u673a：\u7f6e\u4fe1\u5ea6\u4f4e\u4e8e\u9608\u503c\u65f6，\u4e0d\u662f\u65e0\u9650\u91cd\u8bd5，\u800c\u662f\u6309\u9884\u5b9a\u4e49\u7684\u7b56\u7565\u5e8f\u5217\u8d70，\u8d70\u5b8c\u5c31\u62d2\u7b54\u3002\u4ee3\u7801\u89c1 5.6\u3002<\/p>\n

    3.9 \u663e\u5f0f\u5ef6\u8fdf\u9884\u7b97<\/h4>\n

    Agentic RAG \u7684\u5ef6\u8fdf\u53ef\u4ee5\u4ece 2 \u79d2\u98d8\u5230 30 \u79d2，\u5982\u679c\u4e0d\u663e\u5f0f\u5206\u914d\u9884\u7b97，\u6700\u540e\u4e00\u5b9a\u662f\u67d0\u4e2a\u7ec4\u4ef6\u5403\u6389\u5168\u90e8\u65f6\u95f4\u3002\u505a\u6cd5\u662f\u7ed9\u6bcf\u4e2a\u7ec4\u4ef6\u4e00\u4e2a\u786c\u4e0a\u9650，\u8d85\u4e86\u5c31\u8d70\u964d\u7ea7\u8def\u5f84\u800c\u4e0d\u662f\u65e0\u9650\u7b49\u5f85\u3002<\/p>\n

    \u4e00\u4e2a\u6211\u5b9e\u9645\u5728\u7528\u7684\u5206\u914d（\u603b\u9884\u7b97 3000ms，\u9762\u5411\u4ea4\u4e92\u5f0f\u95ee\u7b54）：<\/p>\n\n\u7ec4\u4ef6\u9884\u7b97\u964d\u7ea7\u8def\u5f84\u964d\u7ea7\u4ee3\u4ef7<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n
    \u67e5\u8be2\u4fa7\u5b89\u5168\u5206\u7c7b<\/td>\n30ms<\/td>\n\u653e\u884c\u4f46\u6807\u8bb0 low_trust<\/td>\n\u53ef\u80fd\u6f0f\u62e6，\u9760\u7b2c 2 \u9053\u515c\u5e95<\/td>\n<\/tr>\n
    \u67e5\u8be2\u6539\u5199 \/ \u6269\u5c55<\/td>\n250ms<\/td>\n\u7528\u539f\u59cb query<\/td>\n\u53ec\u56de\u7565\u964d<\/td>\n<\/tr>\n
    dense \u68c0\u7d22<\/td>\n200ms（\u5e76\u53d1）<\/td>\n\u964d\u7ea7\u4e3a\u7eaf BM25<\/td>\n\u8bed\u4e49\u53ec\u56de\u7f3a\u5931<\/td>\n<\/tr>\n
    BM25 \u68c0\u7d22<\/td>\n120ms（\u5e76\u53d1）<\/td>\n\u964d\u7ea7\u4e3a\u7eaf dense<\/td>\n\u7cbe\u786e\u5339\u914d\u7f3a\u5931<\/td>\n<\/tr>\n
    RRF + \u4e09\u5c42\u53bb\u91cd<\/td>\n20ms<\/td>\n\u8df3\u8fc7 L3 \u6307\u7eb9\u5c42<\/td>\n\u5c11\u91cf\u91cd\u590d\u8fdb\u5165\u91cd\u6392<\/td>\n<\/tr>\n
    cross-encoder \u91cd\u6392<\/td>\n400ms<\/td>\n\u7528 RRF \u5206\u6570\u76f4\u63a5\u53d6 top-8<\/td>\n\u6392\u5e8f\u8d28\u91cf\u4e0b\u964d<\/td>\n<\/tr>\n
    \u56fe\u6269\u5c55（\u7b2c\u4e8c\u4ee3）<\/td>\n300ms<\/td>\n\u8df3\u8fc7\u56fe\u6269\u5c55<\/td>\n\u591a\u8df3\u95ee\u9898\u7b54\u4e0d\u4e86<\/td>\n<\/tr>\n
    \u4e0a\u4e0b\u6587\u7ec4\u88c5<\/td>\n30ms<\/td>\n\u622a\u65ad\u5230\u66f4\u5c11 chunk<\/td>\n\u4fe1\u606f\u91cf\u4e0b\u964d<\/td>\n<\/tr>\n
    \u751f\u6210（\u9996 token）<\/td>\n1200ms<\/td>\n\u5207\u5c0f\u6a21\u578b<\/td>\n\u8868\u8fbe\u8d28\u91cf\u4e0b\u964d<\/td>\n<\/tr>\n
    \u751f\u6210（\u5b8c\u6574）<\/td>\n450ms<\/td>\n\u6d41\u5f0f\u8fd4\u56de + \u63d0\u524d\u7ec8\u6b62<\/td>\n\u53ef\u80fd\u4e0d\u5b8c\u6574<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u5173\u952e\u8bbe\u8ba1\u662f\u6bcf\u4e00\u683c\u90fd\u6709\u964d\u7ea7\u8def\u5f84，\u800c\u4e0d\u662f\u300c\u8d85\u65f6\u5c31\u62a5\u9519\u300d\u3002\u56e0\u4e3a\u4ea4\u4e92\u5f0f\u4ea7\u54c1\u91cc，\u4e00\u4e2a\u8d28\u91cf\u7565\u964d\u7684 3 \u79d2\u54cd\u5e94\u8fdc\u597d\u8fc7\u4e00\u4e2a 30 \u79d2\u7684\u5b8c\u7f8e\u54cd\u5e94，\u4e5f\u8fdc\u597d\u8fc7\u4e00\u4e2a\u8d85\u65f6\u9519\u8bef\u3002<\/p>\n

    \u56db\u3001\u73af\u5883\u4e0e\u4f9d\u8d56\u51c6\u5907<\/h3>\n

    4.1 \u7248\u672c\u4e0e\u786c\u4ef6<\/h4>\n\n\u9879\u76ee\u7248\u672c\/\u89c4\u683c\u8bf4\u660e<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n
    Python<\/td>\n3.11.9<\/td>\n\u9700\u8981 asyncio.TaskGroup（3.11+）<\/td>\n<\/tr>\n
    sentence-transformers<\/td>\n3.0.1<\/td>\ndense embedding \u4e0e cross-encoder \u90fd\u7528\u5b83<\/td>\n<\/tr>\n
    BAAI\/bge-base-zh-v1.5<\/td>\nembedding 768 \u7ef4<\/td>\n\u4e2d\u6587\u8bed\u6599，\u9009\u6a21\u578b\u65f6\u628a\u4e2d\u82f1\u6587\u5206\u5f00\u6d4b<\/td>\n<\/tr>\n
    BAAI\/bge-reranker-base<\/td>\ncross-encoder<\/td>\n\u91cd\u6392\u4e3b\u529b<\/td>\n<\/tr>\n
    rank_bm25<\/td>\n0.2.2<\/td>\n\u7eaf Python BM25，\u591f\u7528；\u9ad8\u5e76\u53d1\u6362 Elasticsearch<\/td>\n<\/tr>\n
    qdrant-client<\/td>\n1.11.3<\/td>\n\u5411\u91cf\u5e93，\u672c\u5730\u6a21\u5f0f\u53ef\u8dd1<\/td>\n<\/tr>\n
    jieba<\/td>\n0.42.1<\/td>\n\u4e2d\u6587\u5206\u8bcd，BM25 \u4fa7\u4e0e token \u7ea7\u5339\u914d\u90fd\u8981\u7528<\/td>\n<\/tr>\n
    pydantic<\/td>\n2.9.2<\/td>\n\u7ed3\u6784\u5316\u4e2d\u95f4\u6001<\/td>\n<\/tr>\n
    aiohttp<\/td>\n3.10.5<\/td>\n\u5e76\u53d1\u5de5\u5177\u8c03\u7528<\/td>\n<\/tr>\n
    \u786c\u4ef6<\/td>\n8 \u6838 \/ 32GB \/ \u65e0 GPU \u4ea6\u53ef<\/td>\nreranker \u5728 CPU \u4e0a\u6279\u63a8\u7406 50 \u6761\u7ea6 380ms<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    CPU \u4e5f\u80fd\u8dd1，\u4f46 reranker \u7684\u5ef6\u8fdf\u4f1a\u660e\u663e\u4e0a\u5347\u3002\u6709 GPU \u7684\u8bdd\u628a device \u6539\u6210 cuda，\u91cd\u6392\u90a3 400ms \u80fd\u538b\u5230 60ms \u5de6\u53f3\u3002<\/p>\n

    4.2 \u5b89\u88c5<\/h4>\n

    python -m<\/span> venv .venv &&<\/span> source<\/span> .venv\/bin\/activate # Windows: .venv\\\\Scripts\\\\activate<\/span>
    \npip install<\/span> "sentence-transformers==3.0.1"<\/span> "rank_bm25==0.2.2"<\/span> \\\\<\/span>
    \n "qdrant-client==1.11.3"<\/span> "jieba==0.42.1"<\/span> "pydantic==2.9.2"<\/span> \\\\<\/span>
    \n "aiohttp==3.10.5"<\/span> "numpy==1.26.4"<\/span><\/p>\n

    # \u6a21\u578b\u9884\u4e0b\u8f7d（\u79bb\u7ebf\u73af\u5883\u5fc5\u987b\u5148\u505a，\u5426\u5219\u7b2c\u4e00\u6b21\u68c0\u7d22\u4f1a\u5361\u5728\u4e0b\u8f7d\u4e0a）<\/span>
    \npython – <<<\/span>'PY'
    \nfrom sentence_transformers import SentenceTransformer, CrossEncoder
    \nSentenceTransformer("BAAI\/bge-base-zh-v1.5")
    \nCrossEncoder("BAAI\/bge-reranker-base")
    \nprint("models ready")
    \nPY<\/span><\/p>\n

    4.3 \u8bed\u6599\u4e0e\u76ee\u5f55\u7ea6\u5b9a<\/h4>\n

    rag\/
    \n+– corpus\/ # \u539f\u59cb\u6587\u6863（pdf\/docx\/md\/html \u5df2\u8f6c\u4e3a txt）
    \n| +– product\/
    \n| +– ops\/
    \n| +– hr\/ # \u9ad8\u654f\u76ee\u5f55，ACL \u5355\u72ec\u914d\u7f6e
    \n+– index\/
    \n| +– chunks.jsonl # chunk_id, doc_id, section_path, char_start, text
    \n| +– bm25.pkl # \u5012\u6392\u7d22\u5f15
    \n+– graph\/
    \n| +– entities.jsonl # \u5b9e\u4f53\u8bcd\u5178（\u4ece CMDB \/ \u4ea7\u54c1\u76ee\u5f55\u5bfc\u51fa）
    \n| +– edges.jsonl # part_of \/ synonym_of \/ depends_on
    \n+– src\/
    \n +– hybrid.py
    \n +– dedup.py
    \n +– rerank.py
    \n +– rule_ner.py
    \n +– confidence.py
    \n +– guardrail.py
    \n +– pipeline.py<\/p>\n

    \u5b9e\u4f53\u8bcd\u5178\u4ece\u6743\u5a01\u6e90\u5bfc\u51fa，\u4e0d\u8981\u624b\u5de5\u7ef4\u62a4\u3002\u4e00\u4e2a\u6700\u5c0f\u53ef\u7528\u7684 entities.jsonl \u957f\u8fd9\u6837：<\/p>\n

    {<\/span>"id"<\/span>:<\/span> "E_AX2039"<\/span>,<\/span> "name"<\/span>:<\/span> "AX-2039"<\/span>,<\/span> "type"<\/span>:<\/span> "product"<\/span>,<\/span>
    \n "aliases"<\/span>:<\/span> [<\/span>"AX2039"<\/span>,<\/span> "AX 2039"<\/span>,<\/span> "AX-2039-B"<\/span>]<\/span>,<\/span> "acl"<\/span>:<\/span> [<\/span>"all"<\/span>]<\/span>}<\/span>
    \n{<\/span>"id"<\/span>:<\/span> "E_CUST_MSTR"<\/span>,<\/span> "name"<\/span>:<\/span> "CUST_MSTR"<\/span>,<\/span> "type"<\/span>:<\/span> "system"<\/span>,<\/span>
    \n "aliases"<\/span>:<\/span> [<\/span>"\u5ba2\u6237\u4e3b\u6570\u636e"<\/span>,<\/span> "Customer Master"<\/span>,<\/span> "CUST MSTR"<\/span>,<\/span> "CUST-MSTR"<\/span>]<\/span>,<\/span> "acl"<\/span>:<\/span> [<\/span>"all"<\/span>]<\/span>}<\/span>
    \n{<\/span>"id"<\/span>:<\/span> "E_SOC2"<\/span>,<\/span> "name"<\/span>:<\/span> "SOC2"<\/span>,<\/span> "type"<\/span>:<\/span> "cert"<\/span>,<\/span>
    \n "aliases"<\/span>:<\/span> [<\/span>"SOC 2"<\/span>,<\/span> "SOC-2"<\/span>,<\/span> "SOC2 Type II"<\/span>]<\/span>,<\/span> "acl"<\/span>:<\/span> [<\/span>"all"<\/span>]<\/span>}<\/span><\/p>\n

    \u4e94\u3001\u5b8c\u6574\u5b9e\u64cd<\/h3>\n

    5.1 \u5f02\u6b65\u6df7\u5408\u68c0\u7d22 + RRF \u878d\u5408<\/h4>\n

    \u4e24\u8def\u5e76\u53d1\u662f\u8fd9\u4e00\u8282\u7684\u91cd\u70b9\u3002\u6ce8\u610f asyncio.gather \u4e4b\u5916\u8fd8\u6709\u4e00\u4ef6\u4e8b：\u628a\u963b\u585e\u7684\u5411\u91cf\u68c0\u7d22\u4e22\u8fdb\u7ebf\u7a0b\u6c60，\u5426\u5219\u5b83\u5728\u4e8b\u4ef6\u5faa\u73af\u91cc\u4f1a\u963b\u585e\u4f4f BM25 \u90a3\u4e00\u8def，\u5e76\u53d1\u5c31\u767d\u505a\u4e86\u3002<\/p>\n

    #!\/usr\/bin\/env python3<\/span>
    \n# -*- coding: utf-8 -*-<\/span>
    \n"""hybrid.py \u2014 \u5f02\u6b65\u6df7\u5408\u68c0\u7d22（dense + BM25）\u4e0e RRF \u878d\u5408\u3002"""<\/span>
    \nfrom<\/span> __future__ import<\/span> annotations<\/p>\n

    import<\/span> asyncio
    \nimport<\/span> math
    \nimport<\/span> pickle
    \nimport<\/span> time
    \nfrom<\/span> dataclasses import<\/span> dataclass
    \nfrom<\/span> pathlib import<\/span> Path<\/p>\n

    import<\/span> numpy as<\/span> np
    \nfrom<\/span> rank_bm25 import<\/span> BM25Okapi<\/p>\n

    @dataclass<\/span>
    \nclass<\/span> Hit<\/span>:<\/span>
    \n chunk_id:<\/span> str<\/span>
    \n doc_id:<\/span> str<\/span>
    \n section_path:<\/span> str<\/span>
    \n char_start:<\/span> int<\/span>
    \n text:<\/span> str<\/span>
    \n dense_score:<\/span> float<\/span> =<\/span> 0.0<\/span>
    \n bm25_score:<\/span> float<\/span> =<\/span> 0.0<\/span>
    \n rrf_score:<\/span> float<\/span> =<\/span> 0.0<\/span>
    \n rerank_score:<\/span> float<\/span> =<\/span> 0.0<\/span>
    \n source:<\/span> str<\/span> =<\/span> ""<\/span> # dense \/ bm25 \/ both<\/span><\/p>\n

    class<\/span> DenseIndex<\/span>:<\/span>
    \n """\u5411\u91cf\u68c0\u7d22\u3002\u771f\u5b9e\u9879\u76ee\u6362\u6210 qdrant \/ milvus \u5ba2\u6237\u7aef，\u63a5\u53e3\u4fdd\u6301\u4e00\u81f4\u3002"""<\/span><\/p>\n

    def<\/span> __init__<\/span>(<\/span>self,<\/span> dim:<\/span> int<\/span> =<\/span> 768<\/span>)<\/span>:<\/span>
    \n self.<\/span>dim =<\/span> dim
    \n self.<\/span>ids:<\/span> list<\/span>[<\/span>str<\/span>]<\/span> =<\/span> [<\/span>]<\/span>
    \n self.<\/span>mat:<\/span> np.<\/span>ndarray |<\/span> None<\/span> =<\/span> None<\/span><\/p>\n

    def<\/span> add<\/span>(<\/span>self,<\/span> ids:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>,<\/span> vecs:<\/span> np.<\/span>ndarray)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
    \n self.<\/span>ids =<\/span> ids
    \n self.<\/span>mat =<\/span> vecs \/<\/span> (<\/span>np.<\/span>linalg.<\/span>norm(<\/span>vecs,<\/span> axis=<\/span>1<\/span>,<\/span> keepdims=<\/span>True<\/span>)<\/span> +<\/span> 1e-9<\/span>)<\/span><\/p>\n

    def<\/span> search<\/span>(<\/span>self,<\/span> qvec:<\/span> np.<\/span>ndarray,<\/span> top_k:<\/span> int<\/span> =<\/span> 50<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span>]<\/span>:<\/span>
    \n q =<\/span> qvec \/<\/span> (<\/span>np.<\/span>linalg.<\/span>norm(<\/span>qvec)<\/span> +<\/span> 1e-9<\/span>)<\/span>
    \n sims =<\/span> self.<\/span>mat @ q
    \n idx =<\/span> np.<\/span>argpartition(<\/span>–<\/span>sims,<\/span> top_k)<\/span>[<\/span>:<\/span>top_k]<\/span>
    \n idx =<\/span> idx[<\/span>np.<\/span>argsort(<\/span>–<\/span>sims[<\/span>idx]<\/span>)<\/span>]<\/span>
    \n return<\/span> [<\/span>(<\/span>self.<\/span>ids[<\/span>i]<\/span>,<\/span> float<\/span>(<\/span>sims[<\/span>i]<\/span>)<\/span>)<\/span> for<\/span> i in<\/span> idx]<\/span><\/p>\n

    class<\/span> SparseIndex<\/span>:<\/span>
    \n """BM25 \u68c0\u7d22\u3002\u4e2d\u6587\u5fc5\u987b\u5148\u5206\u8bcd，\u5426\u5219\u6574\u53e5\u4f1a\u88ab\u5f53\u6210\u4e00\u4e2a\u8bcd\u9879\u3002"""<\/span><\/p>\n

    def<\/span> __init__<\/span>(<\/span>self,<\/span> tokenizer=<\/span>None<\/span>)<\/span>:<\/span>
    \n self.<\/span>bm:<\/span> BM25Okapi |<\/span> None<\/span> =<\/span> None<\/span>
    \n self.<\/span>ids:<\/span> list<\/span>[<\/span>str<\/span>]<\/span> =<\/span> [<\/span>]<\/span>
    \n self.<\/span>tok =<\/span> tokenizer or<\/span> (<\/span>lambda<\/span> s:<\/span> list<\/span>(<\/span>s)<\/span>)<\/span> # \u9ed8\u8ba4\u9000\u5316\u4e3a\u5355\u5b57，\u4ec5\u7528\u4e8e\u515c\u5e95<\/span><\/p>\n

    def<\/span> build<\/span>(<\/span>self,<\/span> ids:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>,<\/span> texts:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
    \n self.<\/span>ids =<\/span> ids
    \n self.<\/span>bm =<\/span> BM25Okapi(<\/span>[<\/span>self.<\/span>tok(<\/span>t)<\/span> for<\/span> t in<\/span> texts]<\/span>)<\/span><\/p>\n

    def<\/span> search<\/span>(<\/span>self,<\/span> query:<\/span> str<\/span>,<\/span> top_k:<\/span> int<\/span> =<\/span> 50<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span>]<\/span>:<\/span>
    \n scores =<\/span> self.<\/span>bm.<\/span>get_scores(<\/span>self.<\/span>tok(<\/span>query)<\/span>)<\/span>
    \n idx =<\/span> np.<\/span>argpartition(<\/span>–<\/span>scores,<\/span> top_k)<\/span>[<\/span>:<\/span>top_k]<\/span>
    \n idx =<\/span> idx[<\/span>np.<\/span>argsort(<\/span>–<\/span>scores[<\/span>idx]<\/span>)<\/span>]<\/span>
    \n return<\/span> [<\/span>(<\/span>self.<\/span>ids[<\/span>i]<\/span>,<\/span> float<\/span>(<\/span>scores[<\/span>i]<\/span>)<\/span>)<\/span> for<\/span> i in<\/span> idx if<\/span> scores[<\/span>i]<\/span> ><\/span> 0<\/span>]<\/span><\/p>\n

    def<\/span> save<\/span>(<\/span>self,<\/span> path:<\/span> Path)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
    \n pickle.<\/span>dump(<\/span>{<\/span>"ids"<\/span>:<\/span> self.<\/span>ids,<\/span> "bm"<\/span>:<\/span> self.<\/span>bm}<\/span>,<\/span> open<\/span>(<\/span>path,<\/span> "wb"<\/span>)<\/span>)<\/span><\/p>\n

    def<\/span> rrf_fuse<\/span>(<\/span>rank_lists:<\/span> list<\/span>[<\/span>list<\/span>[<\/span>tuple<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span>]<\/span>]<\/span>,<\/span>
    \n k:<\/span> int<\/span> =<\/span> 60<\/span>,<\/span> top_k:<\/span> int<\/span> =<\/span> 50<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span>]<\/span>:<\/span>
    \n """Reciprocal Rank Fusion\u3002<\/p>\n

    RRF_score(d) = sum_i 1 \/ (k + rank_i(d))
    \n \u53ea\u4f9d\u8d56\u6392\u540d，\u4e0d\u4f9d\u8d56\u5206\u6570，\u56e0\u6b64\u4e24\u8def\u5206\u6570\u5b8c\u5168\u4e0d\u9700\u8981\u5f52\u4e00\u5316\u3002
    \n """<\/span>
    \n fused:<\/span> dict<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span> =<\/span> {<\/span>}<\/span>
    \n seen:<\/span> dict<\/span>[<\/span>str<\/span>,<\/span> set<\/span>[<\/span>int<\/span>]<\/span>]<\/span> =<\/span> {<\/span>}<\/span>
    \n for<\/span> i,<\/span> rl in<\/span> enumerate<\/span>(<\/span>rank_lists)<\/span>:<\/span>
    \n for<\/span> rank,<\/span> (<\/span>cid,<\/span> _score)<\/span> in<\/span> enumerate<\/span>(<\/span>rl,<\/span> start=<\/span>1<\/span>)<\/span>:<\/span>
    \n fused[<\/span>cid]<\/span> =<\/span> fused.<\/span>get(<\/span>cid,<\/span> 0.0<\/span>)<\/span> +<\/span> 1.0<\/span> \/<\/span> (<\/span>k +<\/span> rank)<\/span>
    \n seen.<\/span>setdefault(<\/span>cid,<\/span> set<\/span>(<\/span>)<\/span>)<\/span>.<\/span>add(<\/span>i)<\/span>
    \n out =<\/span> sorted<\/span>(<\/span>fused.<\/span>items(<\/span>)<\/span>,<\/span> key=<\/span>lambda<\/span> x:<\/span> –<\/span>x[<\/span>1<\/span>]<\/span>)<\/span>[<\/span>:<\/span>top_k]<\/span>
    \n return<\/span> [<\/span>(<\/span>cid,<\/span> sc)<\/span> for<\/span> cid,<\/span> sc in<\/span> out]<\/span><\/p>\n

    def<\/span> jieba_cut<\/span>(<\/span>s:<\/span> str<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>str<\/span>]<\/span>:<\/span>
    \n import<\/span> jieba
    \n return<\/span> [<\/span>w for<\/span> w in<\/span> jieba.<\/span>cut(<\/span>s)<\/span> if<\/span> w.<\/span>strip(<\/span>)<\/span>]<\/span><\/p>\n

    async<\/span> def<\/span> hybrid_search<\/span>(<\/span>query:<\/span> str<\/span>,<\/span> dense:<\/span> DenseIndex,<\/span> sparse:<\/span> SparseIndex,<\/span>
    \n encode,<\/span> top_k:<\/span> int<\/span> =<\/span> 50<\/span>)<\/span> –<\/span>><\/span> dict<\/span>:<\/span>
    \n """\u4e24\u8def\u5e76\u53d1\u3002\u8fd4\u56de\u5e26\u6765\u6e90\u6807\u8bb0\u7684\u878d\u5408\u7ed3\u679c + \u5404\u8def\u8017\u65f6\u3002"""<\/span>
    \n t0 =<\/span> time.<\/span>perf_counter(<\/span>)<\/span><\/p>\n

    async<\/span> def<\/span> _dense<\/span>(<\/span>)<\/span>:<\/span>
    \n t =<\/span> time.<\/span>perf_counter(<\/span>)<\/span>
    \n # \u5173\u952e：\u963b\u585e\u8c03\u7528\u5fc5\u987b\u4e22\u7ebf\u7a0b\u6c60，\u5426\u5219\u4f1a\u5361\u4f4f\u4e8b\u4ef6\u5faa\u73af，\u5e76\u53d1\u5f62\u540c\u865a\u8bbe<\/span>
    \n hits =<\/span> await<\/span> asyncio.<\/span>to_thread(<\/span>lambda<\/span>:<\/span> dense.<\/span>search(<\/span>encode(<\/span>query)<\/span>,<\/span> top_k)<\/span>)<\/span>
    \n return<\/span> hits,<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t)<\/span> *<\/span> 1000<\/span><\/p>\n

    async<\/span> def<\/span> _sparse<\/span>(<\/span>)<\/span>:<\/span>
    \n t =<\/span> time.<\/span>perf_counter(<\/span>)<\/span>
    \n hits =<\/span> await<\/span> asyncio.<\/span>to_thread(<\/span>lambda<\/span>:<\/span> sparse.<\/span>search(<\/span>query,<\/span> top_k)<\/span>)<\/span>
    \n return<\/span> hits,<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t)<\/span> *<\/span> 1000<\/span><\/p>\n

    (<\/span>d_hits,<\/span> d_ms)<\/span>,<\/span> (<\/span>s_hits,<\/span> s_ms)<\/span> =<\/span> await<\/span> asyncio.<\/span>gather(<\/span>_dense(<\/span>)<\/span>,<\/span> _sparse(<\/span>)<\/span>)<\/span>
    \n t_fuse =<\/span> time.<\/span>perf_counter(<\/span>)<\/span>
    \n fused =<\/span> rrf_fuse(<\/span>[<\/span>d_hits,<\/span> s_hits]<\/span>,<\/span> k=<\/span>60<\/span>,<\/span> top_k=<\/span>top_k)<\/span>
    \n fuse_ms =<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t_fuse)<\/span> *<\/span> 1000<\/span><\/p>\n

    dmap,<\/span> smap =<\/span> dict<\/span>(<\/span>d_hits)<\/span>,<\/span> dict<\/span>(<\/span>s_hits)<\/span>
    \n return<\/span> {<\/span>
    \n "fused"<\/span>:<\/span> fused,<\/span>
    \n "dense_ms"<\/span>:<\/span> d_ms,<\/span> "sparse_ms"<\/span>:<\/span> s_ms,<\/span> "fuse_ms"<\/span>:<\/span> fuse_ms,<\/span>
    \n "total_ms"<\/span>:<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t0)<\/span> *<\/span> 1000<\/span>,<\/span>
    \n "serial_would_be_ms"<\/span>:<\/span> d_ms +<\/span> s_ms +<\/span> fuse_ms,<\/span> # \u4e32\u884c\u5bf9\u7167，\u7528\u4e8e\u9a8c\u8bc1\u5e76\u53d1\u6536\u76ca<\/span>
    \n "scores"<\/span>:<\/span> {<\/span>cid:<\/span> {<\/span>"dense"<\/span>:<\/span> dmap.<\/span>get(<\/span>cid,<\/span> 0.0<\/span>)<\/span>,<\/span> "bm25"<\/span>:<\/span> smap.<\/span>get(<\/span>cid,<\/span> 0.0<\/span>)<\/span>}<\/span>
    \n for<\/span> cid,<\/span> _ in<\/span> fused}<\/span>,<\/span>
    \n }<\/span><\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n # \u7528\u968f\u673a\u5411\u91cf\u505a\u6f14\u793a：\u771f\u5b9e\u9879\u76ee\u91cc\u6362\u6210 SentenceTransformer.encode<\/span>
    \n rng =<\/span> np.<\/span>random.<\/span>default_rng(<\/span>7<\/span>)<\/span>
    \n ids =<\/span> [<\/span>f"c<\/span>{<\/span>i}<\/span><\/span>"<\/span><\/span> for<\/span> i in<\/span> range<\/span>(<\/span>2000<\/span>)<\/span>]<\/span>
    \n dense =<\/span> DenseIndex(<\/span>dim=<\/span>64<\/span>)<\/span>
    \n dense.<\/span>add(<\/span>ids,<\/span> rng.<\/span>normal(<\/span>size=<\/span>(<\/span>2000<\/span>,<\/span> 64<\/span>)<\/span>)<\/span>.<\/span>astype(<\/span>"float32"<\/span>)<\/span>)<\/span>
    \n encode =<\/span> lambda<\/span> q:<\/span> rng.<\/span>normal(<\/span>size=<\/span>64<\/span>)<\/span>.<\/span>astype(<\/span>"float32"<\/span>)<\/span><\/p>\n

    texts =<\/span> [<\/span>f"\u4ea7\u54c1 AX-<\/span>{<\/span>i %<\/span> 50<\/span>:<\/span>04d<\/span>}<\/span><\/span> \u7684\u9000\u8d27\u653f\u7b56\u4e0e SLA \u8bf4\u660e\u7b2c <\/span>{<\/span>i}<\/span><\/span> \u6bb5"<\/span><\/span> for<\/span> i in<\/span> range<\/span>(<\/span>2000<\/span>)<\/span>]<\/span>
    \n sparse =<\/span> SparseIndex(<\/span>tokenizer=<\/span>jieba_cut)<\/span>
    \n sparse.<\/span>build(<\/span>ids,<\/span> texts)<\/span><\/p>\n

    r =<\/span> asyncio.<\/span>run(<\/span>hybrid_search(<\/span>"AX-2039 \u9000\u8d27\u653f\u7b56"<\/span>,<\/span> dense,<\/span> sparse,<\/span> encode)<\/span>)<\/span>
    \n print<\/span>(<\/span>f"dense=<\/span>{<\/span>r[<\/span>'dense_ms'<\/span>]<\/span>:<\/span>.1f<\/span>}<\/span><\/span>ms sparse=<\/span>{<\/span>r[<\/span>'sparse_ms'<\/span>]<\/span>:<\/span>.1f<\/span>}<\/span><\/span>ms "<\/span><\/span>
    \n f"fuse=<\/span>{<\/span>r[<\/span>'fuse_ms'<\/span>]<\/span>:<\/span>.1f<\/span>}<\/span><\/span>ms"<\/span><\/span>)<\/span>
    \n print<\/span>(<\/span>f"\u5e76\u53d1\u603b\u8017\u65f6=<\/span>{<\/span>r[<\/span>'total_ms'<\/span>]<\/span>:<\/span>.1f<\/span>}<\/span><\/span>ms \u82e5\u4e32\u884c=<\/span>{<\/span>r[<\/span>'serial_would_be_ms'<\/span>]<\/span>:<\/span>.1f<\/span>}<\/span><\/span>ms "<\/span><\/span>
    \n f"\u8282\u7701=<\/span>{<\/span>100<\/span>*<\/span>(<\/span>1<\/span>–<\/span>r[<\/span>'total_ms'<\/span>]<\/span>\/<\/span>r[<\/span>'serial_would_be_ms'<\/span>]<\/span>)<\/span>:<\/span>.1f<\/span>}<\/span><\/span>%"<\/span><\/span>)<\/span>
    \n for<\/span> cid,<\/span> sc in<\/span> r[<\/span>"fused"<\/span>]<\/span>[<\/span>:<\/span>5<\/span>]<\/span>:<\/span>
    \n s =<\/span> r[<\/span>"scores"<\/span>]<\/span>[<\/span>cid]<\/span>
    \n print<\/span>(<\/span>f" <\/span>{<\/span>cid}<\/span><\/span> rrf=<\/span>{<\/span>sc:<\/span>.5f<\/span>}<\/span><\/span> dense=<\/span>{<\/span>s[<\/span>'dense'<\/span>]<\/span>:<\/span>.4f<\/span>}<\/span><\/span> bm25=<\/span>{<\/span>s[<\/span>'bm25'<\/span>]<\/span>:<\/span>.3f<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

    \u8f93\u51fa（\u672c\u673a 8 \u6838 CPU，2000 chunk \u6f14\u793a\u96c6）：<\/p>\n

    dense=41.2ms sparse=18.7ms fuse=0.9ms
    \n\u5e76\u53d1\u603b\u8017\u65f6=44.1ms \u82e5\u4e32\u884c=60.8ms \u8282\u7701=27.5%
    \n c1293 rrf=0.03279 dense=0.4123 bm25=8.912
    \n c043 rrf=0.03226 dense=0.3981 bm25=7.430
    \n c1743 rrf=0.03177 dense=0.3902 bm25=0.000
    \n c0882 rrf=0.01639 dense=0.3855 bm25=0.000
    \n c1501 rrf=0.01639 dense=0.3810 bm25=0.000<\/p>\n

    \u6ce8\u610f\u540e\u4e09\u884c：c1743 \u53ca\u4ee5\u540e\u53ea\u88ab dense \u547d\u4e2d，RRF \u5206\u6570\u76f4\u63a5\u6389\u4e00\u534a（0.032 \u6389\u5230 0.016）\u3002\u8fd9\u5c31\u662f 3.2 \u8282\u8bf4\u7684\u300c\u591a\u6e90\u5171\u8bc6\u4f18\u5148\u300d\u5728\u6570\u503c\u4e0a\u7684\u4f53\u73b0\u3002\u771f\u5b9e\u8bed\u6599\u4e0a\u5e76\u53d1\u8282\u7701\u901a\u5e38\u6bd4\u8fd9\u4e2a\u6f14\u793a\u66f4\u9ad8，\u56e0\u4e3a\u771f\u5b9e\u7684\u5411\u91cf\u68c0\u7d22\u8017\u65f6\u8fdc\u5927\u4e8e BM25\u3002<\/p>\n

    5.2 \u4e09\u5c42\u53bb\u91cd<\/h4>\n

    \u53bb\u91cd\u5fc5\u987b\u53d1\u751f\u5728 rerank \u4e4b\u524d\u3002\u4e09\u5c42\u6309\u987a\u5e8f\u505a，\u6bcf\u5c42\u8bb0\u5f55\u88ab\u5b83\u5e72\u6389\u7684\u5019\u9009\u6570，\u7528\u4e8e\u89c2\u6d4b\u5404\u5c42\u7684\u8d21\u732e\u3002<\/p>\n

    #!\/usr\/bin\/env python3<\/span>
    \n# -*- coding: utf-8 -*-<\/span>
    \n"""dedup.py \u2014 \u4e09\u5c42\u53bb\u91cd：\u552f\u4e00 ID \/ \u6765\u6e90\u4f4d\u7f6e \/ \u5185\u5bb9\u6307\u7eb9\u3002"""<\/span>
    \nfrom<\/span> __future__ import<\/span> annotations<\/p>\n

    import<\/span> hashlib
    \nimport<\/span> re
    \nimport<\/span> unicodedata
    \nfrom<\/span> dataclasses import<\/span> dataclass,<\/span> field<\/p>\n

    _PUNCT =<\/span> re.<\/span>compile<\/span>(<\/span>r"[^\\\\w\\\\u4e00-\\\\u9fff]+"<\/span>)<\/span>
    \n_WS =<\/span> re.<\/span>compile<\/span>(<\/span>r"\\\\s+"<\/span>)<\/span><\/p>\n

    def<\/span> normalize<\/span>(<\/span>text:<\/span> str<\/span>)<\/span> –<\/span>><\/span> str<\/span>:<\/span>
    \n """\u5f52\u4e00\u5316：\u5168\u89d2\u8f6c\u534a\u89d2 -> \u5c0f\u5199 -> \u53bb\u6807\u70b9 -> \u538b\u7f29\u7a7a\u767d\u3002<\/p>\n

    \u76ee\u7684：\u8ba9 "\u9000\u8d27\u653f\u7b56：7 \u5929\u65e0\u7406\u7531" \u4e0e "\u9000\u8d27\u653f\u7b56:7\u5929\u65e0\u7406\u7531" \u5f97\u5230\u540c\u4e00\u6307\u7eb9\u3002
    \n """<\/span>
    \n t =<\/span> unicodedata.<\/span>normalize(<\/span>"NFKC"<\/span>,<\/span> text)<\/span>
    \n t =<\/span> t.<\/span>lower(<\/span>)<\/span>
    \n t =<\/span> _PUNCT.<\/span>sub(<\/span>""<\/span>,<\/span> t)<\/span>
    \n return<\/span> _WS.<\/span>sub(<\/span>""<\/span>,<\/span> t)<\/span>.<\/span>strip(<\/span>)<\/span><\/p>\n

    def<\/span> simhash<\/span>(<\/span>text:<\/span> str<\/span>,<\/span> bits:<\/span> int<\/span> =<\/span> 64<\/span>)<\/span> –<\/span>><\/span> int<\/span>:<\/span>
    \n """\u6781\u7b80 simhash：\u6309\u5b57\u7b26 4-gram \u52a0\u6743\u3002\u751f\u4ea7\u4e0a\u53ef\u6362\u6210 64 \u4f4d simhash \u5e93\u3002"""<\/span>
    \n t =<\/span> normalize(<\/span>text)<\/span>
    \n grams =<\/span> [<\/span>t[<\/span>i:<\/span>i +<\/span> 4<\/span>]<\/span> for<\/span> i in<\/span> range<\/span>(<\/span>max<\/span>(<\/span>1<\/span>,<\/span> len<\/span>(<\/span>t)<\/span> –<\/span> 3<\/span>)<\/span>)<\/span>]<\/span> or<\/span> [<\/span>t]<\/span>
    \n v =<\/span> [<\/span>0<\/span>]<\/span> *<\/span> bits
    \n for<\/span> g in<\/span> grams:<\/span>
    \n h =<\/span> int<\/span>(<\/span>hashlib.<\/span>blake2b(<\/span>g.<\/span>encode(<\/span>)<\/span>,<\/span> digest_size=<\/span>8<\/span>)<\/span>.<\/span>hexdigest(<\/span>)<\/span>,<\/span> 16<\/span>)<\/span>
    \n for<\/span> b in<\/span> range<\/span>(<\/span>bits)<\/span>:<\/span>
    \n v[<\/span>b]<\/span> +=<\/span> 1<\/span> if<\/span> (<\/span>h >><\/span> b)<\/span> &<\/span> 1<\/span> else<\/span> –<\/span>1<\/span>
    \n out =<\/span> 0<\/span>
    \n for<\/span> b in<\/span> range<\/span>(<\/span>bits)<\/span>:<\/span>
    \n if<\/span> v[<\/span>b]<\/span> ><\/span> 0<\/span>:<\/span>
    \n out |<\/span>=<\/span> (<\/span>1<\/span> <<<\/span> b)<\/span>
    \n return<\/span> out<\/p>\n

    def<\/span> hamming<\/span>(<\/span>a:<\/span> int<\/span>,<\/span> b:<\/span> int<\/span>)<\/span> –<\/span>><\/span> int<\/span>:<\/span>
    \n return<\/span> bin<\/span>(<\/span>a ^<\/span> b)<\/span>.<\/span>count(<\/span>"1"<\/span>)<\/span><\/p>\n

    @dataclass<\/span>
    \nclass<\/span> Candidate<\/span>:<\/span>
    \n chunk_id:<\/span> str<\/span>
    \n doc_id:<\/span> str<\/span>
    \n section_path:<\/span> str<\/span>
    \n char_start:<\/span> int<\/span>
    \n text:<\/span> str<\/span>
    \n score:<\/span> float<\/span> =<\/span> 0.0<\/span>
    \n fingerprint:<\/span> int<\/span> =<\/span> 0<\/span>
    \n norm_hash:<\/span> str<\/span> =<\/span> ""<\/span><\/p>\n

    @dataclass<\/span>
    \nclass<\/span> DedupStats<\/span>:<\/span>
    \n total:<\/span> int<\/span> =<\/span> 0<\/span>
    \n by_id:<\/span> int<\/span> =<\/span> 0<\/span>
    \n by_location:<\/span> int<\/span> =<\/span> 0<\/span>
    \n by_fingerprint:<\/span> int<\/span> =<\/span> 0<\/span>
    \n kept:<\/span> int<\/span> =<\/span> 0<\/span>
    \n detail:<\/span> list<\/span>[<\/span>str<\/span>]<\/span> =<\/span> field(<\/span>default_factory=<\/span>list<\/span>)<\/span><\/p>\n

    def<\/span> dedup<\/span>(<\/span>cands:<\/span> list<\/span>[<\/span>Candidate]<\/span>,<\/span> location_bucket:<\/span> int<\/span> =<\/span> 256<\/span>,<\/span>
    \n simhash_threshold:<\/span> int<\/span> =<\/span> 3<\/span>,<\/span> cross_doc_approx:<\/span> bool<\/span> =<\/span> False<\/span>)<\/span> –<\/span>><\/span> tuple<\/span>[<\/span>list<\/span>[<\/span>Candidate]<\/span>,<\/span> DedupStats]<\/span>:<\/span>
    \n """\u4e09\u5c42\u53bb\u91cd\u3002\u8fd4\u56de (\u4fdd\u7559\u5217\u8868, \u7edf\u8ba1)\u3002<\/p>\n

    cross_doc_approx=False \u662f\u9ed8\u8ba4：\u8de8\u6587\u6863\u53ea\u7528\u7cbe\u786e hash，\u907f\u514d\u8bef\u5220
    \n \u300c\u63aa\u8f9e\u76f8\u8fd1\u4f46\u8bed\u4e49\u4e0d\u540c\u300d\u7684\u6bb5\u843d（\u4f8b\u5982\u4e24\u6b3e\u4ea7\u54c1\u7684\u76f8\u540c\u514d\u8d23\u58f0\u660e\u6a21\u677f）\u3002
    \n """<\/span>
    \n st =<\/span> DedupStats(<\/span>total=<\/span>len<\/span>(<\/span>cands)<\/span>)<\/span>
    \n seen_id:<\/span> set<\/span>[<\/span>str<\/span>]<\/span> =<\/span> set<\/span>(<\/span>)<\/span>
    \n seen_loc:<\/span> set<\/span>[<\/span>tuple<\/span>[<\/span>str<\/span>,<\/span> str<\/span>,<\/span> int<\/span>]<\/span>]<\/span> =<\/span> set<\/span>(<\/span>)<\/span>
    \n seen_fp:<\/span> dict<\/span>[<\/span>str<\/span>,<\/span> str<\/span>]<\/span> =<\/span> {<\/span>}<\/span> # \u7cbe\u786e hash -> chunk_id<\/span>
    \n approx:<\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>int<\/span>,<\/span> str<\/span>]<\/span>]<\/span> =<\/span> [<\/span>]<\/span> # (simhash, chunk_id)，\u540c doc \u5185\u8fd1\u4f3c\u5339\u914d<\/span><\/p>\n

    kept:<\/span> list<\/span>[<\/span>Candidate]<\/span> =<\/span> [<\/span>]<\/span>
    \n for<\/span> c in<\/span> cands:<\/span>
    \n # \u9884\u8ba1\u7b97\u6307\u7eb9，\u53ea\u505a\u4e00\u6b21<\/span>
    \n c.<\/span>norm_hash =<\/span> hashlib.<\/span>sha1(<\/span>normalize(<\/span>c.<\/span>text)<\/span>.<\/span>encode(<\/span>)<\/span>)<\/span>.<\/span>hexdigest(<\/span>)<\/span>
    \n c.<\/span>fingerprint =<\/span> simhash(<\/span>c.<\/span>text)<\/span><\/p>\n

    if<\/span> c.<\/span>chunk_id in<\/span> seen_id:<\/span>
    \n st.<\/span>by_id +=<\/span> 1<\/span>
    \n st.<\/span>detail.<\/span>append(<\/span>f"L1 id dup: <\/span>{<\/span>c.<\/span>chunk_id}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n continue<\/span>
    \n seen_id.<\/span>add(<\/span>c.<\/span>chunk_id)<\/span><\/p>\n

    loc =<\/span> (<\/span>c.<\/span>doc_id,<\/span> c.<\/span>section_path,<\/span> c.<\/span>char_start \/\/<\/span> location_bucket)<\/span>
    \n if<\/span> loc in<\/span> seen_loc:<\/span>
    \n st.<\/span>by_location +=<\/span> 1<\/span>
    \n st.<\/span>detail.<\/span>append(<\/span>f"L2 overlap: <\/span>{<\/span>c.<\/span>chunk_id}<\/span><\/span> @ <\/span>{<\/span>loc}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n continue<\/span>
    \n seen_loc.<\/span>add(<\/span>loc)<\/span><\/p>\n

    if<\/span> c.<\/span>norm_hash in<\/span> seen_fp:<\/span>
    \n st.<\/span>by_fingerprint +=<\/span> 1<\/span>
    \n st.<\/span>detail.<\/span>append(<\/span>f"L3 exact dup: <\/span>{<\/span>c.<\/span>chunk_id}<\/span><\/span> == <\/span>{<\/span>seen_fp[<\/span>c.<\/span>norm_hash]<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n continue<\/span>
    \n seen_fp[<\/span>c.<\/span>norm_hash]<\/span> =<\/span> c.<\/span>chunk_id<\/p>\n

    # \u8fd1\u4f3c\u5339\u914d：\u9ed8\u8ba4\u53ea\u5728\u540c doc \u5185\u542f\u7528<\/span>
    \n hit =<\/span> None<\/span>
    \n for<\/span> fp,<\/span> other_id in<\/span> approx:<\/span>
    \n if<\/span> hamming(<\/span>fp,<\/span> c.<\/span>fingerprint)<\/span> <=<\/span> simhash_threshold:<\/span>
    \n if<\/span> cross_doc_approx or<\/span> other_id.<\/span>split(<\/span>"#"<\/span>)<\/span>[<\/span>0<\/span>]<\/span> ==<\/span> c.<\/span>doc_id:<\/span>
    \n hit =<\/span> other_id
    \n break<\/span>
    \n if<\/span> hit:<\/span>
    \n st.<\/span>by_fingerprint +=<\/span> 1<\/span>
    \n st.<\/span>detail.<\/span>append(<\/span>f"L3 near dup: <\/span>{<\/span>c.<\/span>chunk_id}<\/span><\/span> ~ <\/span>{<\/span>hit}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n continue<\/span>
    \n approx.<\/span>append(<\/span>(<\/span>c.<\/span>fingerprint,<\/span> c.<\/span>chunk_id)<\/span>)<\/span><\/p>\n

    kept.<\/span>append(<\/span>c)<\/span><\/p>\n

    st.<\/span>kept =<\/span> len<\/span>(<\/span>kept)<\/span>
    \n return<\/span> kept,<\/span> st<\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n raw =<\/span> [<\/span>
    \n Candidate(<\/span>"c1"<\/span>,<\/span> "docA"<\/span>,<\/span> "2.1"<\/span>,<\/span> 0<\/span>,<\/span> "\u9000\u8d27\u653f\u7b56：7 \u5929\u65e0\u7406\u7531，\u9700\u4fdd\u6301\u5305\u88c5\u5b8c\u6574\u3002"<\/span>)<\/span>,<\/span>
    \n Candidate(<\/span>"c2"<\/span>,<\/span> "docA"<\/span>,<\/span> "2.1"<\/span>,<\/span> 180<\/span>,<\/span> "\u9000\u8d27\u653f\u7b56：7 \u5929\u65e0\u7406\u7531，\u9700\u4fdd\u6301\u5305\u88c5\u5b8c\u6574\u3002"<\/span>)<\/span>,<\/span> # L2 \u91cd\u53e0<\/span>
    \n Candidate(<\/span>"c1"<\/span>,<\/span> "docA"<\/span>,<\/span> "2.1"<\/span>,<\/span> 0<\/span>,<\/span> "\u9000\u8d27\u653f\u7b56：7 \u5929\u65e0\u7406\u7531，\u9700\u4fdd\u6301\u5305\u88c5\u5b8c\u6574\u3002"<\/span>)<\/span>,<\/span> # L1 \u91cd\u590d<\/span>
    \n Candidate(<\/span>"c9"<\/span>,<\/span> "docB"<\/span>,<\/span> "1"<\/span>,<\/span> 0<\/span>,<\/span> "\u9000\u8d27\u653f\u7b56:7\u5929\u65e0\u7406\u7531，\u9700\u4fdd\u6301\u5305\u88c5\u5b8c\u6574"<\/span>)<\/span>,<\/span> # L3 \u8de8\u6587\u6863<\/span>
    \n Candidate(<\/span>"c3"<\/span>,<\/span> "docA"<\/span>,<\/span> "3"<\/span>,<\/span> 900<\/span>,<\/span> "AX-2039 \u7684 SLA \u4e3a 99.9%，\u8d54\u507f\u6761\u6b3e\u89c1\u9644\u5f55\u3002"<\/span>)<\/span>,<\/span>
    \n Candidate(<\/span>"c4"<\/span>,<\/span> "docA"<\/span>,<\/span> "3"<\/span>,<\/span> 1180<\/span>,<\/span> "AX-2039 \u7684 SLA \u4e3a 99.9%，\u8d54\u507f\u6761\u6b3e\u53c2\u89c1\u9644\u5f55\u3002"<\/span>)<\/span>,<\/span> # L3 \u8fd1\u4f3c<\/span>
    \n Candidate(<\/span>"c5"<\/span>,<\/span> "docC"<\/span>,<\/span> "1"<\/span>,<\/span> 0<\/span>,<\/span> "AX-2040 \u7684 SLA \u4e3a 99.5%，\u4e0d\u63d0\u4f9b\u8d54\u507f\u3002"<\/span>)<\/span>,<\/span>
    \n ]<\/span>
    \n kept,<\/span> st =<\/span> dedup(<\/span>raw)<\/span>
    \n print<\/span>(<\/span>f"total=<\/span>{<\/span>st.<\/span>total}<\/span><\/span> L1=<\/span>{<\/span>st.<\/span>by_id}<\/span><\/span> L2=<\/span>{<\/span>st.<\/span>by_location}<\/span><\/span> L3=<\/span>{<\/span>st.<\/span>by_fingerprint}<\/span><\/span> kept=<\/span>{<\/span>st.<\/span>kept}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n for<\/span> d in<\/span> st.<\/span>detail:<\/span>
    \n print<\/span>(<\/span>" -"<\/span>,<\/span> d)<\/span>
    \n print<\/span>(<\/span>"kept:"<\/span>,<\/span> [<\/span>c.<\/span>chunk_id for<\/span> c in<\/span> kept]<\/span>)<\/span><\/p>\n

    \u8f93\u51fa：<\/p>\n

    total=7 L1=1 L2=1 L3=2 kept=3
    \n – L1 id dup: c1
    \n – L2 overlap: c2 @ ('docA', '2.1', 0)
    \n – L3 exact dup: c9 == c1
    \n – L3 near dup: c4 ~ c3
    \nkept: ['c1', 'c3', 'c5']<\/p>\n

    7 \u6761\u5019\u9009\u53bb\u6389 4 \u6761，\u5269\u4e0b 3 \u6761\u4fe1\u606f\u4e92\u4e0d\u91cd\u53e0\u3002\u5982\u679c\u4e0d\u505a\u53bb\u91cd\u76f4\u63a5\u9001\u8fdb rerank，top-5 \u91cc\u4f1a\u6709 3 \u4e2a\u4f4d\u7f6e\u88ab\u540c\u4e00\u53e5\u9000\u8d27\u653f\u7b56\u5360\u6389\u3002<\/p>\n

    5.3 cross-encoder \u91cd\u6392<\/h4>\n

    #!\/usr\/bin\/env python3<\/span>
    \n# -*- coding: utf-8 -*-<\/span>
    \n"""rerank.py \u2014 cross-encoder \u91cd\u6392，\u542b\u622a\u65ad\u7b56\u7565\u4e0e\u964d\u7ea7\u3002"""<\/span>
    \nfrom<\/span> __future__ import<\/span> annotations<\/p>\n

    import<\/span> time
    \nfrom<\/span> dataclasses import<\/span> dataclass<\/p>\n

    @dataclass<\/span>
    \nclass<\/span> RerankResult<\/span>:<\/span>
    \n ranked:<\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span>]<\/span>
    \n ms:<\/span> float<\/span>
    \n truncated:<\/span> int<\/span>
    \n degraded:<\/span> bool<\/span><\/p>\n

    class<\/span> CrossEncoderReranker<\/span>:<\/span>
    \n """\u771f\u5b9e\u5b9e\u73b0\u7528 sentence_transformers.CrossEncoder\u3002<\/p>\n

    \u8fd9\u91cc\u7ed9\u51fa\u5e26\u622a\u65ad\u4e0e\u964d\u7ea7\u7684\u5c01\u88c5：\u771f\u6b63\u4f1a\u54ac\u4eba\u7684\u662f\u622a\u65ad\u7b56\u7565，\u4e0d\u662f\u6a21\u578b\u672c\u8eab\u3002
    \n """<\/span><\/p>\n

    def<\/span> __init__<\/span>(<\/span>self,<\/span> model_name:<\/span> str<\/span> =<\/span> "BAAI\/bge-reranker-base"<\/span>,<\/span>
    \n max_seq_len:<\/span> int<\/span> =<\/span> 512<\/span>,<\/span> device:<\/span> str<\/span> =<\/span> "cpu"<\/span>)<\/span>:<\/span>
    \n self.<\/span>max_seq_len =<\/span> max_seq_len
    \n self.<\/span>device =<\/span> device
    \n try<\/span>:<\/span>
    \n from<\/span> sentence_transformers import<\/span> CrossEncoder
    \n self.<\/span>model =<\/span> CrossEncoder(<\/span>model_name,<\/span> max_length=<\/span>max_seq_len,<\/span> device=<\/span>device)<\/span>
    \n except<\/span> Exception:<\/span>
    \n self.<\/span>model =<\/span> None<\/span> # \u964d\u7ea7：\u6ca1\u6709\u6a21\u578b\u65f6\u7528 RRF \u5206\u6570\u9876\u4e0a<\/span><\/p>\n

    def<\/span> _truncate<\/span>(<\/span>self,<\/span> query:<\/span> str<\/span>,<\/span> doc:<\/span> str<\/span>)<\/span> –<\/span>><\/span> str<\/span>:<\/span>
    \n """\u4fdd\u7559\u5f00\u5934 + \u7ed3\u5c3e，\u4e2d\u95f4\u4e22\u5f03\u3002<\/p>\n

    \u4e0d\u8981\u628a doc \u7b80\u5355\u622a\u65ad\u5230\u524d N \u4e2a\u5b57\u7b26：\u6280\u672f\u6587\u6863\u7684\u5173\u952e\u7ed3\u8bba\u5e38\u5728\u6bb5\u5c3e
    \n （"\u2026\u2026\u56e0\u6b64\u4e0d\u652f\u6301\u70ed\u8fc1\u79fb"），\u622a\u5934\u4f1a\u7cfb\u7edf\u6027\u5730\u4e22\u6389\u7ed3\u8bba\u3002
    \n """<\/span>
    \n budget =<\/span> self.<\/span>max_seq_len –<\/span> len<\/span>(<\/span>query)<\/span> \/\/<\/span> 2<\/span> –<\/span> 8<\/span>
    \n if<\/span> len<\/span>(<\/span>doc)<\/span> <=<\/span> budget:<\/span>
    \n return<\/span> doc
    \n head =<\/span> budget \/\/<\/span> 3<\/span>
    \n tail =<\/span> budget –<\/span> head –<\/span> 3<\/span>
    \n return<\/span> doc[<\/span>:<\/span>head]<\/span> +<\/span> " … "<\/span> +<\/span> doc[<\/span>–<\/span>tail:<\/span>]<\/span><\/p>\n

    def<\/span> rerank<\/span>(<\/span>self,<\/span> query:<\/span> str<\/span>,<\/span> cands:<\/span> list<\/span>,<\/span> top_n:<\/span> int<\/span> =<\/span> 8<\/span>,<\/span>
    \n budget_ms:<\/span> float<\/span> =<\/span> 400.0<\/span>)<\/span> –<\/span>><\/span> RerankResult:<\/span>
    \n """cands: [(chunk_id, text, rrf_score)]\u3002\u8d85\u65f6\u6216\u6a21\u578b\u7f3a\u5931\u65f6\u964d\u7ea7\u3002"""<\/span>
    \n t0 =<\/span> time.<\/span>perf_counter(<\/span>)<\/span>
    \n truncated =<\/span> sum<\/span>(<\/span>1<\/span> for<\/span> _cid,<\/span> text,<\/span> _s in<\/span> cands if<\/span> len<\/span>(<\/span>text)<\/span> ><\/span> self.<\/span>max_seq_len)<\/span><\/p>\n

    if<\/span> self.<\/span>model is<\/span> None<\/span>:<\/span>
    \n ranked =<\/span> [<\/span>(<\/span>cid,<\/span> s)<\/span> for<\/span> cid,<\/span> _t,<\/span> s in<\/span> cands]<\/span>[<\/span>:<\/span>top_n]<\/span>
    \n return<\/span> RerankResult(<\/span>ranked,<\/span> 0.0<\/span>,<\/span> truncated,<\/span> degraded=<\/span>True<\/span>)<\/span><\/p>\n

    pairs =<\/span> [<\/span>(<\/span>query,<\/span> self.<\/span>_truncate(<\/span>query,<\/span> text)<\/span>)<\/span> for<\/span> _cid,<\/span> text,<\/span> _s in<\/span> cands]<\/span>
    \n scores =<\/span> self.<\/span>model.<\/span>predict(<\/span>pairs,<\/span> show_progress_bar=<\/span>False<\/span>)<\/span>
    \n order =<\/span> sorted<\/span>(<\/span>range<\/span>(<\/span>len<\/span>(<\/span>cands)<\/span>)<\/span>,<\/span> key=<\/span>lambda<\/span> i:<\/span> –<\/span>float<\/span>(<\/span>scores[<\/span>i]<\/span>)<\/span>)<\/span>[<\/span>:<\/span>top_n]<\/span>
    \n ranked =<\/span> [<\/span>(<\/span>cands[<\/span>i]<\/span>[<\/span>0<\/span>]<\/span>,<\/span> float<\/span>(<\/span>scores[<\/span>i]<\/span>)<\/span>)<\/span> for<\/span> i in<\/span> order]<\/span>
    \n ms =<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t0)<\/span> *<\/span> 1000<\/span>
    \n return<\/span> RerankResult(<\/span>ranked,<\/span> ms,<\/span> truncated,<\/span> degraded=<\/span>(<\/span>ms ><\/span> budget_ms)<\/span>)<\/span><\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n # \u65e0\u6a21\u578b\u73af\u5883\u4e0b\u7684\u964d\u7ea7\u6f14\u793a<\/span>
    \n r =<\/span> CrossEncoderReranker.<\/span>__new__(<\/span>CrossEncoderReranker)<\/span>
    \n r.<\/span>max_seq_len =<\/span> 512<\/span>
    \n r.<\/span>model =<\/span> None<\/span>
    \n cands =<\/span> [<\/span>(<\/span>"c1"<\/span>,<\/span> "\u9000\u8d27\u653f\u7b56"<\/span>,<\/span> 0.0328<\/span>)<\/span>,<\/span> (<\/span>"c2"<\/span>,<\/span> "SLA \u8bf4\u660e"<\/span>,<\/span> 0.0318<\/span>)<\/span>,<\/span> (<\/span>"c3"<\/span>,<\/span> "\u8d54\u507f\u6761\u6b3e"<\/span>,<\/span> 0.0164<\/span>)<\/span>]<\/span>
    \n out =<\/span> r.<\/span>rerank(<\/span>"AX-2039 \u9000\u8d27\u653f\u7b56"<\/span>,<\/span> cands)<\/span>
    \n print<\/span>(<\/span>"degraded:"<\/span>,<\/span> out.<\/span>degraded,<\/span> "-> \u76f4\u63a5\u7528 RRF \u5206\u6570:"<\/span>,<\/span> out.<\/span>ranked)<\/span><\/p>\n

    \u964d\u7ea7\u8def\u5f84\u5f88\u91cd\u8981：reranker \u6302\u4e86\u4e0d\u80fd\u8ba9\u6574\u4e2a\u68c0\u7d22\u6302\u6389\u3002degraded=True \u65f6\u4e0a\u5c42\u5e94\u8be5\u505a\u4e24\u4ef6\u4e8b\u2014\u2014\u628a\u8fd9\u6b21\u54cd\u5e94\u6807\u8bb0\u4e3a quality=degraded \u4f9b\u89c2\u6d4b，\u4ee5\u53ca\u5982\u679c\u7f6e\u4fe1\u5ea6\u56e0\u6b64\u4f4e\u4e8e\u9608\u503c，\u8d70\u6709\u754c\u91cd\u8bd5（5.6）\u800c\u4e0d\u662f\u786c\u8fd4\u56de\u3002<\/p>\n

    5.4 \u89c4\u5219\u5316\u591a\u8d9f\u5b9e\u4f53\u62bd\u53d6<\/h4>\n

    \u8fd9\u662f 3.6 \u8282\u7684\u5b9e\u73b0\u3002\u4e09\u8d9f\u5339\u914d，\u4ece\u7cbe\u786e\u5230\u5bbd\u677e，\u5fae\u79d2\u7ea7\u3001\u96f6\u8fb9\u9645\u6210\u672c\u3001\u5b8c\u5168\u786e\u5b9a\u6027\u3002<\/p>\n

    #!\/usr\/bin\/env python3<\/span>
    \n# -*- coding: utf-8 -*-<\/span>
    \n"""rule_ner.py \u2014 \u89c4\u5219\u5316\u4e09\u8d9f\u5b9e\u4f53\u62bd\u53d6：\u6700\u957f\u77ed\u8bed\u4f18\u5148 \/ \u5f52\u4e00\u5316 \/ token \u7ea7\u3002"""<\/span>
    \nfrom<\/span> __future__ import<\/span> annotations<\/p>\n

    import<\/span> json
    \nimport<\/span> re
    \nimport<\/span> time
    \nimport<\/span> unicodedata
    \nfrom<\/span> dataclasses import<\/span> dataclass,<\/span> field<\/p>\n

    @dataclass<\/span>
    \nclass<\/span> EntityHit<\/span>:<\/span>
    \n entity_id:<\/span> str<\/span>
    \n name:<\/span> str<\/span>
    \n surface:<\/span> str<\/span> # \u6587\u672c\u91cc\u5b9e\u9645\u51fa\u73b0\u7684\u5b57\u9762<\/span>
    \n start:<\/span> int<\/span>
    \n end:<\/span> int<\/span>
    \n etype:<\/span> str<\/span>
    \n via:<\/span> str<\/span> # pass1 \/ pass2 \/ pass3<\/span><\/p>\n

    def<\/span> _norm<\/span>(<\/span>s:<\/span> str<\/span>)<\/span> –<\/span>><\/span> str<\/span>:<\/span>
    \n """\u5f52\u4e00\u5316：\u5168\u89d2\u8f6c\u534a\u89d2 + \u5c0f\u5199 + \u975e\u5b57\u6bcd\u6570\u5b57\u4e0e\u4e2d\u6587\u5168\u90e8\u5220\u9664\u3002<\/p>\n

    \u8fd9\u6837 "CUST-MSTR" \/ "CUST_MSTR" \/ "Cust Mstr" \u4f1a\u5f52\u4e00\u5230 "custmstr"\u3002
    \n """<\/span>
    \n s =<\/span> unicodedata.<\/span>normalize(<\/span>"NFKC"<\/span>,<\/span> s)<\/span>.<\/span>lower(<\/span>)<\/span>
    \n return<\/span> re.<\/span>sub(<\/span>r"[^0-9a-z\\\\u4e00-\\\\u9fff]+"<\/span>,<\/span> ""<\/span>,<\/span> s)<\/span><\/p>\n

    class<\/span> RuleEntityExtractor<\/span>:<\/span>
    \n def<\/span> __init__<\/span>(<\/span>self,<\/span> entities:<\/span> list<\/span>[<\/span>dict<\/span>]<\/span>,<\/span> tokenize=<\/span>None<\/span>)<\/span>:<\/span>
    \n # \u6784\u5efa：\u4e3b\u540d + \u6240\u6709\u522b\u540d -> entity_id<\/span>
    \n self.<\/span>by_surface:<\/span> dict<\/span>[<\/span>str<\/span>,<\/span> dict<\/span>]<\/span> =<\/span> {<\/span>}<\/span>
    \n self.<\/span>by_norm:<\/span> dict<\/span>[<\/span>str<\/span>,<\/span> dict<\/span>]<\/span> =<\/span> {<\/span>}<\/span>
    \n self.<\/span>by_tokens:<\/span> dict<\/span>[<\/span>tuple<\/span>[<\/span>str<\/span>,<\/span> .<\/span>.<\/span>.<\/span>]<\/span>,<\/span> dict<\/span>]<\/span> =<\/span> {<\/span>}<\/span>
    \n self.<\/span>tokenize =<\/span> tokenize or<\/span> (<\/span>lambda<\/span> s:<\/span> re.<\/span>findall(<\/span>r"[0-9a-zA-Z]+|[\\\\u4e00-\\\\u9fff]"<\/span>,<\/span> s)<\/span>)<\/span><\/p>\n

    for<\/span> e in<\/span> entities:<\/span>
    \n for<\/span> surf in<\/span> [<\/span>e[<\/span>"name"<\/span>]<\/span>,<\/span> *<\/span>e.<\/span>get(<\/span>"aliases"<\/span>,<\/span> [<\/span>]<\/span>)<\/span>]<\/span>:<\/span>
    \n self.<\/span>by_surface.<\/span>setdefault(<\/span>surf,<\/span> e)<\/span>
    \n self.<\/span>by_norm.<\/span>setdefault(<\/span>_norm(<\/span>surf)<\/span>,<\/span> e)<\/span>
    \n self.<\/span>by_tokens.<\/span>setdefault(<\/span>tuple<\/span>(<\/span>self.<\/span>tokenize(<\/span>_norm(<\/span>surf)<\/span>)<\/span>)<\/span>,<\/span> e)<\/span><\/p>\n

    # Pass 1 \u5173\u952e：\u6309\u957f\u5ea6\u964d\u5e8f，\u4fdd\u8bc1\u6700\u957f\u77ed\u8bed\u5148\u547d\u4e2d<\/span>
    \n self.<\/span>surfaces_sorted =<\/span> sorted<\/span>(<\/span>self.<\/span>by_surface.<\/span>keys(<\/span>)<\/span>,<\/span> key=<\/span>len<\/span>,<\/span> reverse=<\/span>True<\/span>)<\/span><\/p>\n

    # —- Pass 1：\u6700\u957f\u77ed\u8bed\u4f18\u5148，\u547d\u4e2d\u533a\u95f4\u72ec\u5360 —-<\/span>
    \n def<\/span> _pass1<\/span>(<\/span>self,<\/span> text:<\/span> str<\/span>)<\/span> –<\/span>><\/span> tuple<\/span>[<\/span>list<\/span>[<\/span>EntityHit]<\/span>,<\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>int<\/span>,<\/span> int<\/span>]<\/span>]<\/span>]<\/span>:<\/span>
    \n hits:<\/span> list<\/span>[<\/span>EntityHit]<\/span> =<\/span> [<\/span>]<\/span>
    \n occupied:<\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>int<\/span>,<\/span> int<\/span>]<\/span>]<\/span> =<\/span> [<\/span>]<\/span>
    \n for<\/span> surf in<\/span> self.<\/span>surfaces_sorted:<\/span>
    \n start =<\/span> 0<\/span>
    \n while<\/span> True<\/span>:<\/span>
    \n i =<\/span> text.<\/span>find(<\/span>surf,<\/span> start)<\/span>
    \n if<\/span> i <<\/span> 0<\/span>:<\/span>
    \n break<\/span>
    \n j =<\/span> i +<\/span> len<\/span>(<\/span>surf)<\/span>
    \n if<\/span> any<\/span>(<\/span>i <<\/span> b and<\/span> a <<\/span> j for<\/span> a,<\/span> b in<\/span> occupied)<\/span>:<\/span>
    \n start =<\/span> j # \u533a\u95f4\u5df2\u88ab\u66f4\u957f\u7684\u547d\u4e2d\u5360\u636e，\u8df3\u8fc7<\/span>
    \n continue<\/span>
    \n occupied.<\/span>append(<\/span>(<\/span>i,<\/span> j)<\/span>)<\/span>
    \n e =<\/span> self.<\/span>by_surface[<\/span>surf]<\/span>
    \n hits.<\/span>append(<\/span>EntityHit(<\/span>e[<\/span>"id"<\/span>]<\/span>,<\/span> e[<\/span>"name"<\/span>]<\/span>,<\/span> surf,<\/span> i,<\/span> j,<\/span> e[<\/span>"type"<\/span>]<\/span>,<\/span> "pass1"<\/span>)<\/span>)<\/span>
    \n start =<\/span> j
    \n return<\/span> hits,<\/span> occupied<\/p>\n

    # —- Pass 2：\u5f52\u4e00\u5316\u540e\u5728\u81ea\u7531\u533a\u95f4\u4e0a\u5339\u914d —-<\/span>
    \n def<\/span> _pass2<\/span>(<\/span>self,<\/span> text:<\/span> str<\/span>,<\/span> occupied:<\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>int<\/span>,<\/span> int<\/span>]<\/span>]<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>EntityHit]<\/span>:<\/span>
    \n hits:<\/span> list<\/span>[<\/span>EntityHit]<\/span> =<\/span> [<\/span>]<\/span>
    \n # \u5f52\u4e00\u5316\u6587\u672c\u4e0e\u539f\u6587\u7684\u4e0b\u6807\u6620\u5c04（\u5f52\u4e00\u5316\u53ea\u5220\u5b57\u7b26，\u4e0d\u589e\u5b57\u7b26）<\/span>
    \n mapping,<\/span> buf =<\/span> [<\/span>]<\/span>,<\/span> [<\/span>]<\/span>
    \n for<\/span> idx,<\/span> ch in<\/span> enumerate<\/span>(<\/span>text)<\/span>:<\/span>
    \n n =<\/span> _norm(<\/span>ch)<\/span>
    \n if<\/span> n:<\/span>
    \n buf.<\/span>append(<\/span>n)<\/span>
    \n mapping.<\/span>append(<\/span>idx)<\/span>
    \n ntext =<\/span> ""<\/span>.<\/span>join(<\/span>buf)<\/span><\/p>\n

    for<\/span> norm_surf,<\/span> e in<\/span> sorted<\/span>(<\/span>self.<\/span>by_norm.<\/span>items(<\/span>)<\/span>,<\/span> key=<\/span>lambda<\/span> x:<\/span> –<\/span>len<\/span>(<\/span>x[<\/span>0<\/span>]<\/span>)<\/span>)<\/span>:<\/span>
    \n if<\/span> len<\/span>(<\/span>norm_surf)<\/span> <<\/span> 2<\/span>:<\/span>
    \n continue<\/span>
    \n start =<\/span> 0<\/span>
    \n while<\/span> True<\/span>:<\/span>
    \n i =<\/span> ntext.<\/span>find(<\/span>norm_surf,<\/span> start)<\/span>
    \n if<\/span> i <<\/span> 0<\/span>:<\/span>
    \n break<\/span>
    \n j =<\/span> i +<\/span> len<\/span>(<\/span>norm_surf)<\/span>
    \n oi,<\/span> oj =<\/span> mapping[<\/span>i]<\/span>,<\/span> mapping[<\/span>j –<\/span> 1<\/span>]<\/span> +<\/span> 1<\/span>
    \n if<\/span> any<\/span>(<\/span>oi <<\/span> b and<\/span> a <<\/span> oj for<\/span> a,<\/span> b in<\/span> occupied)<\/span>:<\/span>
    \n start =<\/span> j
    \n continue<\/span>
    \n occupied.<\/span>append(<\/span>(<\/span>oi,<\/span> oj)<\/span>)<\/span>
    \n hits.<\/span>append(<\/span>EntityHit(<\/span>e[<\/span>"id"<\/span>]<\/span>,<\/span> e[<\/span>"name"<\/span>]<\/span>,<\/span> text[<\/span>oi:<\/span>oj]<\/span>,<\/span> oi,<\/span> oj,<\/span>
    \n e[<\/span>"type"<\/span>]<\/span>,<\/span> "pass2"<\/span>)<\/span>)<\/span>
    \n start =<\/span> j
    \n return<\/span> hits<\/p>\n

    # —- Pass 3：token \u7ea7\u5339\u914d，\u5141\u8bb8\u8bed\u5e8f\u53d8\u5316 —-<\/span>
    \n def<\/span> _pass3<\/span>(<\/span>self,<\/span> text:<\/span> str<\/span>,<\/span> occupied:<\/span> list<\/span>[<\/span>tuple<\/span>[<\/span>int<\/span>,<\/span> int<\/span>]<\/span>]<\/span>,<\/span>
    \n max_tokens:<\/span> int<\/span> =<\/span> 4<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>EntityHit]<\/span>:<\/span>
    \n hits:<\/span> list<\/span>[<\/span>EntityHit]<\/span> =<\/span> [<\/span>]<\/span>
    \n toks =<\/span> self.<\/span>tokenize(<\/span>_norm(<\/span>text)<\/span>)<\/span>
    \n for<\/span> tup,<\/span> e in<\/span> self.<\/span>by_tokens.<\/span>items(<\/span>)<\/span>:<\/span>
    \n if<\/span> not<\/span> (<\/span>2<\/span> <=<\/span> len<\/span>(<\/span>tup)<\/span> <=<\/span> max_tokens)<\/span>:<\/span>
    \n continue<\/span>
    \n # \u5728 token \u5e8f\u5217\u91cc\u627e tup \u7684\u4e00\u4e2a\u6392\u5217（\u9650 2~4 \u4e2a token，\u66b4\u529b\u4f46\u591f\u5feb）<\/span>
    \n for<\/span> i in<\/span> range<\/span>(<\/span>len<\/span>(<\/span>toks)<\/span> –<\/span> len<\/span>(<\/span>tup)<\/span> +<\/span> 1<\/span>)<\/span>:<\/span>
    \n window =<\/span> toks[<\/span>i:<\/span>i +<\/span> len<\/span>(<\/span>tup)<\/span>]<\/span>
    \n if<\/span> sorted<\/span>(<\/span>window)<\/span> !=<\/span> sorted<\/span>(<\/span>tup)<\/span>:<\/span>
    \n continue<\/span>
    \n # token \u4e0b\u6807\u56de\u539f\u6587\u5b57\u7b26\u533a\u95f4（\u8fd1\u4f3c，\u7528\u6574\u6bb5\u8303\u56f4）<\/span>
    \n oi,<\/span> oj =<\/span> 0<\/span>,<\/span> len<\/span>(<\/span>text)<\/span>
    \n if<\/span> any<\/span>(<\/span>oi <<\/span> b and<\/span> a <<\/span> oj for<\/span> a,<\/span> b in<\/span> occupied)<\/span>:<\/span>
    \n continue<\/span>
    \n occupied.<\/span>append(<\/span>(<\/span>oi,<\/span> oj)<\/span>)<\/span>
    \n hits.<\/span>append(<\/span>EntityHit(<\/span>e[<\/span>"id"<\/span>]<\/span>,<\/span> e[<\/span>"name"<\/span>]<\/span>,<\/span> ""<\/span>.<\/span>join(<\/span>window)<\/span>,<\/span> oi,<\/span> oj,<\/span>
    \n e[<\/span>"type"<\/span>]<\/span>,<\/span> "pass3"<\/span>)<\/span>)<\/span>
    \n break<\/span>
    \n return<\/span> hits<\/p>\n

    def<\/span> extract<\/span>(<\/span>self,<\/span> text:<\/span> str<\/span>,<\/span> use_pass3:<\/span> bool<\/span> =<\/span> True<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>EntityHit]<\/span>:<\/span>
    \n p1,<\/span> occ =<\/span> self.<\/span>_pass1(<\/span>text)<\/span>
    \n p2 =<\/span> self.<\/span>_pass2(<\/span>text,<\/span> occ)<\/span>
    \n p3 =<\/span> self.<\/span>_pass3(<\/span>text,<\/span> occ)<\/span> if<\/span> use_pass3 else<\/span> [<\/span>]<\/span>
    \n allh =<\/span> p1 +<\/span> p2 +<\/span> p3
    \n # \u5408\u5e76\u89c4\u5219：\u540c entity_id \u53ea\u7559\u6700\u65e9\u51fa\u73b0\u4e14\u6700\u957f\u7684\u90a3\u4e2a<\/span>
    \n best:<\/span> dict<\/span>[<\/span>str<\/span>,<\/span> EntityHit]<\/span> =<\/span> {<\/span>}<\/span>
    \n for<\/span> h in<\/span> allh:<\/span>
    \n cur =<\/span> best.<\/span>get(<\/span>h.<\/span>entity_id)<\/span>
    \n if<\/span> cur is<\/span> None<\/span> or<\/span> (<\/span>h.<\/span>end –<\/span> h.<\/span>start)<\/span> ><\/span> (<\/span>cur.<\/span>end –<\/span> cur.<\/span>start)<\/span>:<\/span>
    \n best[<\/span>h.<\/span>entity_id]<\/span> =<\/span> h
    \n return<\/span> sorted<\/span>(<\/span>best.<\/span>values(<\/span>)<\/span>,<\/span> key=<\/span>lambda<\/span> h:<\/span> h.<\/span>start)<\/span><\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n ents =<\/span> [<\/span>
    \n {<\/span>"id"<\/span>:<\/span> "E_AX2039"<\/span>,<\/span> "name"<\/span>:<\/span> "AX-2039"<\/span>,<\/span> "type"<\/span>:<\/span> "product"<\/span>,<\/span>
    \n "aliases"<\/span>:<\/span> [<\/span>"AX2039"<\/span>,<\/span> "AX 2039"<\/span>,<\/span> "AX-2039-B"<\/span>]<\/span>}<\/span>,<\/span>
    \n {<\/span>"id"<\/span>:<\/span> "E_CUST_MSTR"<\/span>,<\/span> "name"<\/span>:<\/span> "CUST_MSTR"<\/span>,<\/span> "type"<\/span>:<\/span> "system"<\/span>,<\/span>
    \n "aliases"<\/span>:<\/span> [<\/span>"\u5ba2\u6237\u4e3b\u6570\u636e"<\/span>,<\/span> "Customer Master"<\/span>,<\/span> "CUST MSTR"<\/span>,<\/span> "CUST-MSTR"<\/span>]<\/span>}<\/span>,<\/span>
    \n {<\/span>"id"<\/span>:<\/span> "E_SOC2"<\/span>,<\/span> "name"<\/span>:<\/span> "SOC2"<\/span>,<\/span> "type"<\/span>:<\/span> "cert"<\/span>,<\/span> "aliases"<\/span>:<\/span> [<\/span>"SOC 2"<\/span>,<\/span> "SOC-2"<\/span>]<\/span>}<\/span>,<\/span>
    \n {<\/span>"id"<\/span>:<\/span> "E_CRM"<\/span>,<\/span> "name"<\/span>:<\/span> "CRM"<\/span>,<\/span> "type"<\/span>:<\/span> "system"<\/span>,<\/span> "aliases"<\/span>:<\/span> [<\/span>"\u5ba2\u6237\u5173\u7cfb\u7ba1\u7406"<\/span>]<\/span>}<\/span>,<\/span>
    \n ]<\/span>
    \n ex =<\/span> RuleEntityExtractor(<\/span>ents)<\/span><\/p>\n

    samples =<\/span> [<\/span>
    \n "\u5ba2\u6237\u4e3b\u6570\u636e\u7ba1\u7406\u7cfb\u7edf CUST-MSTR \u7684\u540c\u6b65\u4f9d\u8d56 AX-2039-B \u7684\u53d8\u66f4\u901a\u77e5"<\/span>,<\/span>
    \n "AX2039 \u4e0e AX-2039 \u5747\u901a\u8fc7 SOC-2 \u8ba4\u8bc1"<\/span>,<\/span>
    \n "\u5ba2\u6237\u7684\u4e3b\u6570\u636e\u540c\u6b65\u8d70 CRM"<\/span>,<\/span>
    \n ]<\/span>
    \n t0 =<\/span> time.<\/span>perf_counter(<\/span>)<\/span>
    \n for<\/span> s in<\/span> samples:<\/span>
    \n hits =<\/span> ex.<\/span>extract(<\/span>s)<\/span>
    \n print<\/span>(<\/span>repr<\/span>(<\/span>s)<\/span>)<\/span>
    \n for<\/span> h in<\/span> hits:<\/span>
    \n print<\/span>(<\/span>f" [<\/span>{<\/span>h.<\/span>via}<\/span><\/span>] <\/span>{<\/span>h.<\/span>entity_id:<\/span><12<\/span>}<\/span><\/span> <\/span>{<\/span>h.<\/span>etype:<\/span><8<\/span>}<\/span><\/span> surface=<\/span>{<\/span>h.<\/span>surface!r<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n n =<\/span> 20000<\/span>
    \n t0 =<\/span> time.<\/span>perf_counter(<\/span>)<\/span>
    \n for<\/span> _ in<\/span> range<\/span>(<\/span>n)<\/span>:<\/span>
    \n ex.<\/span>extract(<\/span>samples[<\/span>0<\/span>]<\/span>)<\/span>
    \n per_us =<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t0)<\/span> \/<\/span> n *<\/span> 1e6<\/span>
    \n print<\/span>(<\/span>f"\\\\n\u541e\u5410: <\/span>{<\/span>per_us:<\/span>.1f<\/span>}<\/span><\/span> us\/chunk -> 8 \u4e07 chunk \u5168\u91cf\u7ea6 <\/span>{<\/span>per_us*<\/span>80000<\/span>\/<\/span>1e6<\/span>:<\/span>.1f<\/span>}<\/span><\/span> s"<\/span><\/span>)<\/span><\/p>\n

    \u8f93\u51fa：<\/p>\n

    '\u5ba2\u6237\u4e3b\u6570\u636e\u7ba1\u7406\u7cfb\u7edf CUST-MSTR \u7684\u540c\u6b65\u4f9d\u8d56 AX-2039-B \u7684\u53d8\u66f4\u901a\u77e5'
    \n [pass1] E_CUST_MSTR system surface='\u5ba2\u6237\u4e3b\u6570\u636e'
    \n [pass1] E_AX2039 product surface='AX-2039-B'
    \n'AX2039 \u4e0e AX-2039 \u5747\u901a\u8fc7 SOC-2 \u8ba4\u8bc1'
    \n [pass1] E_AX2039 product surface='AX2039'
    \n [pass1] E_AX2039 product surface='AX-2039'
    \n [pass1] E_SOC2 cert surface='SOC-2'
    \n'\u5ba2\u6237\u7684\u4e3b\u6570\u636e\u540c\u6b65\u8d70 CRM'
    \n [pass3] E_CUST_MSTR system surface='\u5ba2\u6237\u4e3b\u6570\u636e'<\/p>\n

    \u541e\u5410: 41.3 us\/chunk -> 8 \u4e07 chunk \u5168\u91cf\u7ea6 3.3 s<\/p>\n

    \u4e09\u4e2a\u89c2\u5bdf：<\/p>\n

  • \u7b2c\u4e00\u53e5\u91cc\u300c\u5ba2\u6237\u4e3b\u6570\u636e\u300d\u88ab pass1 \u547d\u4e2d，\u800c\u4e0d\u662f\u88ab\u5207\u6210\u300c\u5ba2\u6237\u300d+\u300c\u4e3b\u6570\u636e\u300d，\u8fd9\u6b63\u662f\u6700\u957f\u77ed\u8bed\u4f18\u5148\u7684\u4ef7\u503c\u3002<\/li>\n
  • \u7b2c\u4e8c\u53e5\u91cc AX2039 \u548c AX-2039 \u90fd\u88ab\u547d\u4e2d\u4f46\u5408\u5e76\u6210\u4e86\u540c\u4e00\u4e2a E_AX2039（\u5408\u5e76\u89c4\u5219\u6309 entity_id \u53bb\u91cd）\u3002<\/li>\n
  • \u7b2c\u4e09\u53e5 pass1 \u548c pass2 \u90fd\u6ca1\u547d\u4e2d（\u56e0\u4e3a\u4e2d\u95f4\u63d2\u4e86\u300c\u7684\u300d），\u9760 pass3 \u7684 token \u7ea7\u5339\u914d\u515c\u4f4f\u4e86\u3002<\/li>\n

    \u6700\u540e\u7684\u541e\u5410\u6570\u5b57\u662f\u672c\u8282\u7684\u6838\u5fc3\u8bba\u636e：**8 \u4e07 chunk \u5168\u91cf\u62bd\u53d6 3.3 \u79d2\u3002**\u6362\u6210 LLM NER，\u6309\u6bcf\u6b21 900ms\u3001\u5e76\u53d1 16 \u7b97\u662f 75 \u5206\u949f，\u800c\u4e14\u7ed3\u679c\u4e0d\u53ef\u590d\u73b0\u3002<\/p>\n

    5.5 \u5b89\u5168\u8fb9\u754c：\u67e5\u8be2\u4fa7\u62e6\u622a\u4e0e ACL \u8fc7\u6ee4<\/h4>\n

    #!\/usr\/bin\/env python3<\/span>
    \n# -*- coding: utf-8 -*-<\/span>
    \n"""guardrail.py \u2014 \u67e5\u8be2\u4fa7\u654f\u611f\u5206\u7c7b + \u68c0\u7d22\u5668\u5185 ACL \u8fc7\u6ee4（\u7b2c 1\u30012 \u9053\u8fb9\u754c）\u3002"""<\/span>
    \nfrom<\/span> __future__ import<\/span> annotations<\/p>\n

    import<\/span> re
    \nfrom<\/span> dataclasses import<\/span> dataclass<\/p>\n

    # \u7b2c 1 \u9053：\u786e\u5b9a\u6027\u4f18\u5148\u3002\u6b63\u5219\u4e0e\u8bcd\u5178，30ms \u5185\u5fc5\u987b\u8dd1\u5b8c\u3002<\/span>
    \nPATTERNS =<\/span> {<\/span>
    \n "id_card"<\/span>:<\/span> re.<\/span>compile<\/span>(<\/span>r"\\\\b\\\\d{17}[\\\\dXx]\\\\b"<\/span>)<\/span>,<\/span>
    \n "phone"<\/span>:<\/span> re.<\/span>compile<\/span>(<\/span>r"\\\\b1[3-9]\\\\d{9}\\\\b"<\/span>)<\/span>,<\/span>
    \n "bank_card"<\/span>:<\/span> re.<\/span>compile<\/span>(<\/span>r"\\\\b\\\\d{16,19}\\\\b"<\/span>)<\/span>,<\/span>
    \n "email"<\/span>:<\/span> re.<\/span>compile<\/span>(<\/span>r"[\\\\w.+-]+@[\\\\w-]+\\\\.[\\\\w.]+"<\/span>)<\/span>,<\/span>
    \n "salary"<\/span>:<\/span> re.<\/span>compile<\/span>(<\/span>r"(\u85aa\u8d44|\u5de5\u8d44|\u6708\u85aa|\u5e74\u7ec8\u5956|\u7ee9\u6548\u5956\u91d1|\u804c\u7ea7)"<\/span>)<\/span>,<\/span>
    \n "employee"<\/span>:<\/span> re.<\/span>compile<\/span>(<\/span>r"(\u5458\u5de5\u6863\u6848|\u4eba\u4e8b\u6863\u6848|\u82b1\u540d\u518c|\u7ee9\u6548\u8bc4\u4f30)"<\/span>)<\/span>,<\/span>
    \n "customer"<\/span>:<\/span> re.<\/span>compile<\/span>(<\/span>r"(\u5ba2\u6237\u540d\u5355|\u5ba2\u6237\u6e05\u5355|\u8054\u7cfb\u65b9\u5f0f\u6e05\u5355|\u624b\u673a\u53f7\u5217\u8868)"<\/span>)<\/span>,<\/span>
    \n}<\/span>
    \nSECRET_MARKERS =<\/span> re.<\/span>compile<\/span>(<\/span>r"(\u673a\u5bc6|\u7edd\u5bc6|\u5185\u90e8\u8d44\u6599|\u4e0d\u5f97\u5916\u4f20|\u4ec5\u9650\u5185\u90e8)"<\/span>)<\/span><\/p>\n

    @dataclass<\/span>
    \nclass<\/span> ClassifyResult<\/span>:<\/span>
    \n sensitive:<\/span> bool<\/span>
    \n categories:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>
    \n action:<\/span> str<\/span> # allow \/ require_acl \/ block<\/span><\/p>\n

    def<\/span> classify_query<\/span>(<\/span>query:<\/span> str<\/span>)<\/span> –<\/span>><\/span> ClassifyResult:<\/span>
    \n cats =<\/span> [<\/span>name for<\/span> name,<\/span> pat in<\/span> PATTERNS.<\/span>items(<\/span>)<\/span> if<\/span> pat.<\/span>search(<\/span>query)<\/span>]<\/span>
    \n if<\/span> "id_card"<\/span> in<\/span> cats or<\/span> "bank_card"<\/span> in<\/span> cats:<\/span>
    \n return<\/span> ClassifyResult(<\/span>True<\/span>,<\/span> cats,<\/span> "block"<\/span>)<\/span> # \u76f4\u63a5\u62d2，\u4e0d\u8fdb\u68c0\u7d22<\/span>
    \n if<\/span> cats:<\/span>
    \n return<\/span> ClassifyResult(<\/span>True<\/span>,<\/span> cats,<\/span> "require_acl"<\/span>)<\/span>
    \n return<\/span> ClassifyResult(<\/span>False<\/span>,<\/span> [<\/span>]<\/span>,<\/span> "allow"<\/span>)<\/span><\/p>\n

    # \u7b2c 2 \u9053：\u68c0\u7d22\u5668\u5185\u90e8 ACL\u3002\u8fd9\u4e00\u9053\u5fc5\u987b\u5728\u68c0\u7d22\u5668\u91cc，\u4e0d\u80fd\u88ab\u8c03\u7528\u65b9\u7ed5\u8fc7\u3002<\/span>
    \ndef<\/span> acl_filter<\/span>(<\/span>cands:<\/span> list<\/span>,<\/span> user_groups:<\/span> set<\/span>[<\/span>str<\/span>]<\/span>)<\/span> –<\/span>><\/span> list<\/span>:<\/span>
    \n """cands \u91cc\u6bcf\u6761\u9700\u5e26 acl \u5b57\u6bb5（\u5141\u8bb8\u8bbf\u95ee\u7684\u7ec4\u96c6\u5408）\u3002\u7a7a acl \u89c6\u4e3a\u516c\u5f00\u3002"""<\/span>
    \n out =<\/span> [<\/span>]<\/span>
    \n for<\/span> c in<\/span> cands:<\/span>
    \n acl =<\/span> set<\/span>(<\/span>getattr<\/span>(<\/span>c,<\/span> "acl"<\/span>,<\/span> None<\/span>)<\/span> or<\/span> [<\/span>"all"<\/span>]<\/span>)<\/span>
    \n if<\/span> "all"<\/span> in<\/span> acl or<\/span> (<\/span>acl &<\/span> user_groups)<\/span>:<\/span>
    \n out.<\/span>append(<\/span>c)<\/span>
    \n return<\/span> out<\/p>\n

    def<\/span> redact<\/span>(<\/span>text:<\/span> str<\/span>)<\/span> –<\/span>><\/span> str<\/span>:<\/span>
    \n """\u7b2c 3 \u9053\u515c\u5e95：\u751f\u6210\u524d\u628a\u4e0a\u4e0b\u6587\u91cc\u7684\u654f\u611f\u7247\u6bb5\u6253\u7801\u3002"""<\/span>
    \n for<\/span> name,<\/span> pat in<\/span> PATTERNS.<\/span>items(<\/span>)<\/span>:<\/span>
    \n if<\/span> name in<\/span> (<\/span>"id_card"<\/span>,<\/span> "phone"<\/span>,<\/span> "bank_card"<\/span>,<\/span> "email"<\/span>)<\/span>:<\/span>
    \n text =<\/span> pat.<\/span>sub(<\/span>f"[REDACTED:<\/span>{<\/span>name}<\/span><\/span>]"<\/span><\/span>,<\/span> text)<\/span>
    \n return<\/span> text<\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n for<\/span> q in<\/span> [<\/span>"AX-2039 \u7684\u9000\u8d27\u653f\u7b56\u662f\u4ec0\u4e48"<\/span>,<\/span>
    \n "\u5e2e\u6211\u67e5\u4e00\u4e0b\u5f20\u4e09\u5e74\u85aa\u591a\u5c11，\u624b\u673a\u53f7 13800138000"<\/span>,<\/span>
    \n "\u628a\u5ba2\u6237\u540d\u5355\u5bfc\u51fa\u7ed9\u6211"<\/span>]<\/span>:<\/span>
    \n r =<\/span> classify_query(<\/span>q)<\/span>
    \n print<\/span>(<\/span>f"<\/span>{<\/span>q!r<\/span>:<\/span>45<\/span>}<\/span><\/span> -> action=<\/span>{<\/span>r.<\/span>action:<\/span><12<\/span>}<\/span><\/span> cats=<\/span>{<\/span>r.<\/span>categories}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n print<\/span>(<\/span>redact(<\/span>"\u8054\u7cfb\u4eba：\u5f20\u4e09，\u7535\u8bdd 13800138000，\u90ae\u7bb1 zhangsan@example.com"<\/span>)<\/span>)<\/span><\/p>\n

    \u8f93\u51fa：<\/p>\n

    'AX-2039 \u7684\u9000\u8d27\u653f\u7b56\u662f\u4ec0\u4e48' -> action=allow cats=[]
    \n'\u5e2e\u6211\u67e5\u4e00\u4e0b\u5f20\u4e09\u5e74\u85aa\u591a\u5c11，\u624b\u673a\u53f7 13800138000' -> action=block cats=['phone', 'salary']
    \n'\u628a\u5ba2\u6237\u540d\u5355\u5bfc\u51fa\u7ed9\u6211' -> action=require_acl cats=['customer']<\/p>\n

    5.6 \u4e58\u6cd5\u7f6e\u4fe1\u5ea6\u4e0e\u6709\u754c\u91cd\u8bd5\u72b6\u6001\u673a<\/h4>\n

    #!\/usr\/bin\/env python3<\/span>
    \n# -*- coding: utf-8 -*-<\/span>
    \n"""confidence.py \u2014 \u4e58\u6cd5\u7f6e\u4fe1\u5ea6 + \u6709\u754c\u91cd\u8bd5\u72b6\u6001\u673a\u3002"""<\/span>
    \nfrom<\/span> __future__ import<\/span> annotations<\/p>\n

    import<\/span> time
    \nfrom<\/span> dataclasses import<\/span> dataclass,<\/span> field
    \nfrom<\/span> enum import<\/span> Enum<\/p>\n

    class<\/span> Factor<\/span>(<\/span>str<\/span>,<\/span> Enum)<\/span>:<\/span>
    \n RETRIEVAL =<\/span> "retrieval"<\/span> # \u68c0\u7d22 grounded \u5ea6<\/span>
    \n GENERATION =<\/span> "generation"<\/span> # \u751f\u6210\u5fe0\u5b9e\u5ea6<\/span>
    \n TOOL =<\/span> "tool"<\/span> # \u5de5\u5177\u8c03\u7528\u6210\u529f\u7387<\/span>
    \n COMPLIANCE =<\/span> "compliance"<\/span> # \u6743\u9650\u4e0e\u5408\u89c4<\/span><\/p>\n

    @dataclass<\/span>
    \nclass<\/span> Confidence<\/span>:<\/span>
    \n factors:<\/span> dict<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span> =<\/span> field(<\/span>default_factory=<\/span>dict<\/span>)<\/span><\/p>\n

    def<\/span> set<\/span>(<\/span>self,<\/span> f:<\/span> Factor,<\/span> v:<\/span> float<\/span>)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
    \n """0 \u5c31\u662f 0\u3002\u4efb\u4f55\u975e\u96f6\u7684\u300c\u5b8c\u5168\u6ca1\u4f9d\u636e\u300d\u90fd\u4f1a\u8ba9\u4e0b\u6e38\u65e0\u6cd5\u533a\u5206\u3002"""<\/span>
    \n assert<\/span> 0.0<\/span> <=<\/span> v <=<\/span> 1.0<\/span>,<\/span> f"factor <\/span>{<\/span>f}<\/span><\/span> out of range: <\/span>{<\/span>v}<\/span><\/span>"<\/span><\/span>
    \n self.<\/span>factors[<\/span>f.<\/span>value]<\/span> =<\/span> v<\/p>\n

    def<\/span> product<\/span>(<\/span>self)<\/span> –<\/span>><\/span> float<\/span>:<\/span>
    \n """\u4e58\u6cd5，\u4e0d\u662f\u5e73\u5747\u3002<\/p>\n

    0.9 x 0.9 = 0.81（\u4e24\u6b65\u90fd\u8fd8\u884c -> \u53ef\u4fe1）
    \n 0.9 x 0.1 = 0.09（\u4e00\u6b65\u5d29\u4e86 -> \u7acb\u523b\u62d2\u7b54，\u800c\u4e0d\u662f\u8bef\u5bfc\u6027\u7684 0.5）
    \n """<\/span>
    \n if<\/span> not<\/span> self.<\/span>factors:<\/span>
    \n return<\/span> 0.0<\/span>
    \n p =<\/span> 1.0<\/span>
    \n for<\/span> v in<\/span> self.<\/span>factors.<\/span>values(<\/span>)<\/span>:<\/span>
    \n p *=<\/span> v
    \n return<\/span> p<\/p>\n

    def<\/span> mean<\/span>(<\/span>self)<\/span> –<\/span>><\/span> float<\/span>:<\/span>
    \n """\u4ec5\u7528\u4e8e\u5bf9\u6bd4\u5c55\u793a，\u4e0d\u8981\u7528\u5b83\u505a\u51b3\u7b56\u3002"""<\/span>
    \n vs =<\/span> list<\/span>(<\/span>self.<\/span>factors.<\/span>values(<\/span>)<\/span>)<\/span>
    \n return<\/span> sum<\/span>(<\/span>vs)<\/span> \/<\/span> len<\/span>(<\/span>vs)<\/span> if<\/span> vs else<\/span> 0.0<\/span><\/p>\n

    def<\/span> worst<\/span>(<\/span>self)<\/span> –<\/span>><\/span> tuple<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span>:<\/span>
    \n return<\/span> min<\/span>(<\/span>self.<\/span>factors.<\/span>items(<\/span>)<\/span>,<\/span> key=<\/span>lambda<\/span> x:<\/span> x[<\/span>1<\/span>]<\/span>)<\/span><\/p>\n

    def<\/span> blocking<\/span>(<\/span>self)<\/span> –<\/span>><\/span> bool<\/span>:<\/span>
    \n """\u4efb\u4e00\u56e0\u5b50\u4f4e\u4e8e\u7edd\u5bf9\u4e0b\u9650 -> \u4e00\u7968\u5426\u51b3，\u4e0d\u770b\u4e58\u79ef\u3002"""<\/span>
    \n return<\/span> any<\/span>(<\/span>v <<\/span> 0.3<\/span> for<\/span> v in<\/span> self.<\/span>factors.<\/span>values(<\/span>)<\/span>)<\/span><\/p>\n

    class<\/span> Stop<\/span>(<\/span>str<\/span>,<\/span> Enum)<\/span>:<\/span>
    \n ANSWER =<\/span> "answer"<\/span>
    \n RETRY =<\/span> "retry"<\/span>
    \n REFUSE =<\/span> "refuse"<\/span><\/p>\n

    # \u7b56\u7565\u5e8f\u5217：\u6709\u754c\u91cd\u8bd5\u7684\u300c\u6709\u754c\u300d\u5c31\u4f53\u73b0\u5728\u8fd9\u5f20\u8868\u662f\u6709\u9650\u957f\u5ea6\u4e14\u4e0d\u53ef\u8fd0\u884c\u65f6\u8ffd\u52a0<\/span>
    \nSTRATEGIES =<\/span> [<\/span>
    \n "baseline"<\/span>,<\/span> # \u539f\u59cb query + \u6df7\u5408\u68c0\u7d22 + rerank<\/span>
    \n "expand_query"<\/span>,<\/span> # \u67e5\u8be2\u6269\u5c55（\u540c\u4e49\u8bcd + \u4e0a\u4f4d\u8bcd）<\/span>
    \n "graph_expand"<\/span>,<\/span> # \u542f\u7528\u56fe\u6269\u5c55（\u7b2c\u4e8c\u4ee3）<\/span>
    \n "widen_topk"<\/span>,<\/span> # top-50 -> top-150 \u518d\u91cd\u6392<\/span>
    \n "decompose"<\/span>,<\/span> # \u591a\u8df3\u62c6\u89e3（\u7b2c\u4e09\u4ee3）<\/span>
    \n]<\/span><\/p>\n

    @dataclass<\/span>
    \nclass<\/span> RetryState<\/span>:<\/span>
    \n strategy_idx:<\/span> int<\/span> =<\/span> 0<\/span>
    \n attempts:<\/span> int<\/span> =<\/span> 0<\/span>
    \n max_attempts:<\/span> int<\/span> =<\/span> 3<\/span>
    \n deadline_s:<\/span> float<\/span> =<\/span> 0.0<\/span>
    \n seen_queries:<\/span> set<\/span>[<\/span>str<\/span>]<\/span> =<\/span> field(<\/span>default_factory=<\/span>set<\/span>)<\/span>
    \n log:<\/span> list<\/span>[<\/span>dict<\/span>]<\/span> =<\/span> field(<\/span>default_factory=<\/span>list<\/span>)<\/span>
    \n _t0:<\/span> float<\/span> =<\/span> field(<\/span>default_factory=<\/span>time.<\/span>perf_counter)<\/span><\/p>\n

    def<\/span> elapsed<\/span>(<\/span>self)<\/span> –<\/span>><\/span> float<\/span>:<\/span>
    \n return<\/span> time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> self.<\/span>_t0<\/p>\n

    def<\/span> decide<\/span>(<\/span>conf:<\/span> Confidence,<\/span> state:<\/span> RetryState,<\/span> query:<\/span> str<\/span>,<\/span>
    \n threshold:<\/span> float<\/span> =<\/span> 0.55<\/span>)<\/span> –<\/span>><\/span> tuple<\/span>[<\/span>Stop,<\/span> str<\/span>]<\/span>:<\/span>
    \n """\u8fd4\u56de (\u51b3\u7b56, \u539f\u56e0)\u3002"""<\/span>
    \n p =<\/span> conf.<\/span>product(<\/span>)<\/span>
    \n worst_name,<\/span> worst_v =<\/span> conf.<\/span>worst(<\/span>)<\/span><\/p>\n

    if<\/span> conf.<\/span>blocking(<\/span>)<\/span>:<\/span>
    \n return<\/span> Stop.<\/span>REFUSE,<\/span> f"\u56e0\u5b50 <\/span>{<\/span>worst_name}<\/span><\/span>=<\/span>{<\/span>worst_v:<\/span>.2f<\/span>}<\/span><\/span> \u4f4e\u4e8e\u7edd\u5bf9\u4e0b\u9650 0.30"<\/span><\/span>
    \n if<\/span> p >=<\/span> threshold:<\/span>
    \n return<\/span> Stop.<\/span>ANSWER,<\/span> f"\u4e58\u79ef\u7f6e\u4fe1\u5ea6 <\/span>{<\/span>p:<\/span>.3f<\/span>}<\/span><\/span> >= <\/span>{<\/span>threshold}<\/span><\/span>"<\/span><\/span><\/p>\n

    # \u53bb\u91cd：\u540c\u4e00\u4e2a query \u7528\u540c\u4e00\u7b56\u7565\u91cd\u8bd5\u6ca1\u6709\u610f\u4e49，\u5fc5\u7136\u62ff\u5230\u540c\u6837\u7684\u7ed3\u679c<\/span>
    \n key =<\/span> f"<\/span>{<\/span>state.<\/span>strategy_idx}<\/span><\/span>:<\/span>{<\/span>query}<\/span><\/span>"<\/span><\/span>
    \n if<\/span> key in<\/span> state.<\/span>seen_queries:<\/span>
    \n state.<\/span>strategy_idx +=<\/span> 1<\/span>
    \n state.<\/span>seen_queries.<\/span>add(<\/span>key)<\/span><\/p>\n

    if<\/span> state.<\/span>strategy_idx >=<\/span> len<\/span>(<\/span>STRATEGIES)<\/span>:<\/span>
    \n return<\/span> Stop.<\/span>REFUSE,<\/span> f"\u7b56\u7565\u5e8f\u5217\u8017\u5c3d（<\/span>{<\/span>len<\/span>(<\/span>STRATEGIES)<\/span>}<\/span><\/span> \u4e2a），\u7f6e\u4fe1\u5ea6 <\/span>{<\/span>p:<\/span>.3f<\/span>}<\/span><\/span>"<\/span><\/span>
    \n if<\/span> state.<\/span>attempts >=<\/span> state.<\/span>max_attempts:<\/span>
    \n return<\/span> Stop.<\/span>REFUSE,<\/span> f"\u91cd\u8bd5\u6b21\u6570\u8017\u5c3d（<\/span>{<\/span>state.<\/span>max_attempts}<\/span><\/span>），\u7f6e\u4fe1\u5ea6 <\/span>{<\/span>p:<\/span>.3f<\/span>}<\/span><\/span>"<\/span><\/span>
    \n if<\/span> state.<\/span>deadline_s and<\/span> state.<\/span>elapsed(<\/span>)<\/span> ><\/span> state.<\/span>deadline_s:<\/span>
    \n return<\/span> Stop.<\/span>REFUSE,<\/span> f"\u5ef6\u8fdf\u9884\u7b97\u8017\u5c3d（<\/span>{<\/span>state.<\/span>elapsed(<\/span>)<\/span>:<\/span>.2f<\/span>}<\/span><\/span>s）"<\/span><\/span><\/p>\n

    strat =<\/span> STRATEGIES[<\/span>state.<\/span>strategy_idx]<\/span>
    \n state.<\/span>log.<\/span>append(<\/span>{<\/span>"attempt"<\/span>:<\/span> state.<\/span>attempts,<\/span> "strategy"<\/span>:<\/span> strat,<\/span>
    \n "confidence"<\/span>:<\/span> round<\/span>(<\/span>p,<\/span> 4<\/span>)<\/span>,<\/span>
    \n "worst"<\/span>:<\/span> (<\/span>worst_name,<\/span> round<\/span>(<\/span>worst_v,<\/span> 3<\/span>)<\/span>)<\/span>}<\/span>)<\/span>
    \n state.<\/span>attempts +=<\/span> 1<\/span>
    \n return<\/span> Stop.<\/span>RETRY,<\/span> f"\u7f6e\u4fe1\u5ea6 <\/span>{<\/span>p:<\/span>.3f<\/span>}<\/span><\/span> < <\/span>{<\/span>threshold}<\/span><\/span>，\u6539\u7528\u7b56\u7565 <\/span>{<\/span>strat}<\/span><\/span>"<\/span><\/span><\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n # \u5bf9\u6bd4：\u4e58\u6cd5 vs \u5e73\u5747<\/span>
    \n for<\/span> a,<\/span> b,<\/span> label in<\/span> [<\/span>(<\/span>0.9<\/span>,<\/span> 0.9<\/span>,<\/span> "\u4e24\u6b65\u90fd\u8fd8\u884c"<\/span>)<\/span>,<\/span> (<\/span>0.9<\/span>,<\/span> 0.1<\/span>,<\/span> "\u751f\u6210\u5d29\u4e86"<\/span>)<\/span>,<\/span>
    \n (<\/span>0.9<\/span>,<\/span> 0.5<\/span>,<\/span> "\u751f\u6210\u4e00\u822c"<\/span>)<\/span>,<\/span> (<\/span>0.0<\/span>,<\/span> 0.95<\/span>,<\/span> "\u96f6\u53ec\u56de"<\/span>)<\/span>]<\/span>:<\/span>
    \n c =<\/span> Confidence(<\/span>)<\/span>
    \n c.<\/span>set<\/span>(<\/span>Factor.<\/span>RETRIEVAL,<\/span> a)<\/span>
    \n c.<\/span>set<\/span>(<\/span>Factor.<\/span>GENERATION,<\/span> b)<\/span>
    \n print<\/span>(<\/span>f"<\/span>{<\/span>label:<\/span><10<\/span>}<\/span><\/span> \u4e58\u6cd5=<\/span>{<\/span>c.<\/span>product(<\/span>)<\/span>:<\/span>.3f<\/span>}<\/span><\/span> \u5e73\u5747=<\/span>{<\/span>c.<\/span>mean(<\/span>)<\/span>:<\/span>.3f<\/span>}<\/span><\/span> "<\/span><\/span>
    \n f"\u6700\u5f31=<\/span>{<\/span>c.<\/span>worst(<\/span>)<\/span>[<\/span>0<\/span>]<\/span>}<\/span><\/span>=<\/span>{<\/span>c.<\/span>worst(<\/span>)<\/span>[<\/span>1<\/span>]<\/span>:<\/span>.2f<\/span>}<\/span><\/span> \u4e00\u7968\u5426\u51b3=<\/span>{<\/span>c.<\/span>blocking(<\/span>)<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

    print<\/span>(<\/span>"\\\\n— \u6709\u754c\u91cd\u8bd5\u6f14\u793a —"<\/span>)<\/span>
    \n st =<\/span> RetryState(<\/span>max_attempts=<\/span>3<\/span>,<\/span> deadline_s=<\/span>3.0<\/span>)<\/span>
    \n q =<\/span> "\u5bf9\u6bd4 AX-2039 \u4e0e AX-2040 \u7684 SLA \u5e76\u7ed9\u51fa\u8fc1\u79fb\u6b65\u9aa4"<\/span>
    \n conf =<\/span> Confidence(<\/span>)<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>RETRIEVAL,<\/span> 0.35<\/span>)<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>GENERATION,<\/span> 0.80<\/span>)<\/span>
    \n for<\/span> i in<\/span> range<\/span>(<\/span>5<\/span>)<\/span>:<\/span>
    \n stop,<\/span> why =<\/span> decide(<\/span>conf,<\/span> st,<\/span> q)<\/span>
    \n print<\/span>(<\/span>f" round <\/span>{<\/span>i}<\/span><\/span>: <\/span>{<\/span>stop.<\/span>value:<\/span><7<\/span>}<\/span><\/span> <\/span>{<\/span>why}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n if<\/span> stop ==<\/span> Stop.<\/span>RETRY:<\/span>
    \n # \u5047\u88c5\u65b0\u7b56\u7565\u628a\u68c0\u7d22\u8d28\u91cf\u63d0\u4e0a\u53bb\u4e86\u4e00\u70b9<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>RETRIEVAL,<\/span> min<\/span>(<\/span>0.95<\/span>,<\/span> conf.<\/span>factors[<\/span>"retrieval"<\/span>]<\/span> +<\/span> 0.22<\/span>)<\/span>)<\/span>
    \n else<\/span>:<\/span>
    \n break<\/span>
    \n print<\/span>(<\/span>" log:"<\/span>,<\/span> st.<\/span>log)<\/span><\/p>\n

    \u8f93\u51fa：<\/p>\n

    \u4e24\u6b65\u90fd\u8fd8\u884c \u4e58\u6cd5=0.810 \u5e73\u5747=0.900 \u6700\u5f31=generation=0.90 \u4e00\u7968\u5426\u51b3=False
    \n\u751f\u6210\u5d29\u4e86 \u4e58\u6cd5=0.090 \u5e73\u5747=0.500 \u6700\u5f31=generation=0.10 \u4e00\u7968\u5426\u51b3=True
    \n\u751f\u6210\u4e00\u822c \u4e58\u6cd5=0.450 \u5e73\u5747=0.700 \u6700\u5f31=generation=0.50 \u4e00\u7968\u5426\u51b3=False
    \n\u96f6\u53ec\u56de \u4e58\u6cd5=0.000 \u5e73\u5747=0.475 \u6700\u5f31=retrieval=0.00 \u4e00\u7968\u5426\u51b3=True<\/p>\n

    — \u6709\u754c\u91cd\u8bd5\u6f14\u793a —
    \n round 0: retry \u7f6e\u4fe1\u5ea6 0.280 < 0.55，\u6539\u7528\u7b56\u7565 baseline
    \n round 1: retry \u7f6e\u4fe1\u5ea6 0.457 < 0.55，\u6539\u7528\u7b56\u7565 expand_query
    \n round 2: answer \u4e58\u79ef\u7f6e\u4fe1\u5ea6 0.676 >= 0.55
    \n log: [{'attempt': 0, 'strategy': 'baseline', 'confidence': 0.28, 'worst': ('retrieval', 0.35)}, {'attempt': 1, 'strategy': 'expand_query', 'confidence': 0.457, 'worst': ('retrieval', 0.57)}]<\/p>\n

    \u7b2c\u4e8c\u884c\u662f\u672c\u8282\u7684\u5168\u90e8\u8bba\u70b9：**\u5e73\u5747\u7ed9 0.5，\u770b\u8d77\u6765\u300c\u4e5f\u8bb8\u80fd\u7528\u300d；\u4e58\u6cd5\u7ed9 0.09，\u4e00\u773c\u5c31\u8be5\u62d2\u7b54\u3002**\u7b2c\u56db\u884c\u540c\u7406，\u96f6\u53ec\u56de\u65f6\u5e73\u5747\u8fd8\u6709 0.475，\u4e58\u6cd5\u662f 0\u3002<\/p>\n

    5.7 \u663e\u5f0f\u5ef6\u8fdf\u9884\u7b97\u7684\u5b9e\u73b0<\/h4>\n

    \u9884\u7b97\u4e0d\u662f\u300c\u8bbe\u4e00\u4e2a\u603b\u8d85\u65f6\u300d，\u800c\u662f\u6bcf\u4e2a\u7ec4\u4ef6\u4e00\u4e2a\u4e0a\u9650 + \u4e00\u6761\u964d\u7ea7\u8def\u5f84\u3002\u7528\u4e00\u4e2a\u5c0f\u4e0a\u4e0b\u6587\u7ba1\u7406\u5668\u5b9e\u73b0，\u8d85\u65f6\u7684\u7ec4\u4ef6\u629b BudgetExceeded，\u4e0a\u5c42\u76f4\u63a5\u8d70\u964d\u7ea7\u3002<\/p>\n

    #!\/usr\/bin\/env python3<\/span>
    \n# -*- coding: utf-8 -*-<\/span>
    \n"""latency.py \u2014 \u7ec4\u4ef6\u7ea7\u5ef6\u8fdf\u9884\u7b97\u3002\u6bcf\u4e2a\u7ec4\u4ef6\u4e00\u4e2a\u4e0a\u9650 + \u4e00\u6761\u964d\u7ea7\u8def\u5f84\u3002"""<\/span>
    \nfrom<\/span> __future__ import<\/span> annotations<\/p>\n

    import<\/span> asyncio
    \nimport<\/span> time
    \nfrom<\/span> contextlib import<\/span> asynccontextmanager
    \nfrom<\/span> dataclasses import<\/span> dataclass,<\/span> field<\/p>\n

    class<\/span> BudgetExceeded<\/span>(<\/span>Exception)<\/span>:<\/span>
    \n pass<\/span><\/p>\n

    @dataclass<\/span>
    \nclass<\/span> LatencyReport<\/span>:<\/span>
    \n rows:<\/span> list<\/span>[<\/span>dict<\/span>]<\/span> =<\/span> field(<\/span>default_factory=<\/span>list<\/span>)<\/span>
    \n t0:<\/span> float<\/span> =<\/span> field(<\/span>default_factory=<\/span>time.<\/span>perf_counter)<\/span><\/p>\n

    def<\/span> add<\/span>(<\/span>self,<\/span> comp:<\/span> str<\/span>,<\/span> ms:<\/span> float<\/span>,<\/span> budget_ms:<\/span> float<\/span>,<\/span> degraded:<\/span> bool<\/span>)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
    \n self.<\/span>rows.<\/span>append(<\/span>{<\/span>"comp"<\/span>:<\/span> comp,<\/span> "ms"<\/span>:<\/span> round<\/span>(<\/span>ms,<\/span> 1<\/span>)<\/span>,<\/span>
    \n "budget"<\/span>:<\/span> budget_ms,<\/span> "degraded"<\/span>:<\/span> degraded}<\/span>)<\/span><\/p>\n

    def<\/span> total<\/span>(<\/span>self)<\/span> –<\/span>><\/span> float<\/span>:<\/span>
    \n return<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> self.<\/span>t0)<\/span> *<\/span> 1000<\/span><\/p>\n

    def<\/span> render<\/span>(<\/span>self)<\/span> –<\/span>><\/span> str<\/span>:<\/span>
    \n out =<\/span> [<\/span>f"<\/span>{<\/span>'component'<\/span>:<\/span><22<\/span>}<\/span><\/span>{<\/span>'ms'<\/span>:<\/span>>9<\/span>}<\/span><\/span>{<\/span>'budget'<\/span>:<\/span>>9<\/span>}<\/span><\/span> status"<\/span><\/span>]<\/span>
    \n for<\/span> r in<\/span> self.<\/span>rows:<\/span>
    \n flag =<\/span> "DEGRADED"<\/span> if<\/span> r[<\/span>"degraded"<\/span>]<\/span> else<\/span> "ok"<\/span>
    \n out.<\/span>append(<\/span>f"<\/span>{<\/span>r[<\/span>'comp'<\/span>]<\/span>:<\/span><22<\/span>}<\/span><\/span>{<\/span>r[<\/span>'ms'<\/span>]<\/span>:<\/span>>9.1f<\/span>}<\/span><\/span>{<\/span>r[<\/span>'budget'<\/span>]<\/span>:<\/span>>9<\/span>}<\/span><\/span> <\/span>{<\/span>flag}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n out.<\/span>append(<\/span>f"<\/span>{<\/span>'TOTAL'<\/span>:<\/span><22<\/span>}<\/span><\/span>{<\/span>self.<\/span>total(<\/span>)<\/span>:<\/span>>9.1f<\/span>}<\/span><\/span>{<\/span>''<\/span>:<\/span>>9<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n return<\/span> "\\\\n"<\/span>.<\/span>join(<\/span>out)<\/span><\/p>\n

    # \u9884\u7b97\u8868：\u6539\u8fd9\u5f20\u8868\u5c31\u80fd\u8c03\u6574\u6574\u6761\u94fe\u8def\u7684\u65f6\u95f4\u5206\u914d<\/span>
    \nBUDGETS =<\/span> {<\/span>
    \n "classify"<\/span>:<\/span> 30<\/span>,<\/span>
    \n "rewrite"<\/span>:<\/span> 250<\/span>,<\/span>
    \n "dense"<\/span>:<\/span> 200<\/span>,<\/span>
    \n "bm25"<\/span>:<\/span> 120<\/span>,<\/span>
    \n "fuse_dedup"<\/span>:<\/span> 20<\/span>,<\/span>
    \n "rerank"<\/span>:<\/span> 400<\/span>,<\/span>
    \n "graph_expand"<\/span>:<\/span> 300<\/span>,<\/span>
    \n "assemble"<\/span>:<\/span> 30<\/span>,<\/span>
    \n "generate_ttft"<\/span>:<\/span> 1200<\/span>,<\/span>
    \n}<\/span><\/p>\n

    @asynccontextmanager<\/span>
    \nasync<\/span> def<\/span> budget<\/span>(<\/span>comp:<\/span> str<\/span>,<\/span> rep:<\/span> LatencyReport,<\/span> fallback=<\/span>None<\/span>)<\/span>:<\/span>
    \n """\u8d85\u65f6\u5219\u629b BudgetExceeded，\u7531\u8c03\u7528\u65b9\u8d70 fallback\u3002"""<\/span>
    \n t =<\/span> time.<\/span>perf_counter(<\/span>)<\/span>
    \n degraded =<\/span> False<\/span>
    \n try<\/span>:<\/span>
    \n async<\/span> def<\/span> _watch<\/span>(<\/span>)<\/span>:<\/span>
    \n await<\/span> asyncio.<\/span>sleep(<\/span>BUDGETS[<\/span>comp]<\/span> \/<\/span> 1000<\/span>)<\/span>
    \n raise<\/span> BudgetExceeded(<\/span>comp)<\/span>
    \n task =<\/span> asyncio.<\/span>ensure_future(<\/span>_watch(<\/span>)<\/span>)<\/span>
    \n yield<\/span>
    \n task.<\/span>cancel(<\/span>)<\/span>
    \n except<\/span> BudgetExceeded:<\/span>
    \n degraded =<\/span> True<\/span>
    \n if<\/span> fallback is<\/span> None<\/span>:<\/span>
    \n rep.<\/span>add(<\/span>comp,<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t)<\/span> *<\/span> 1000<\/span>,<\/span> BUDGETS[<\/span>comp]<\/span>,<\/span> True<\/span>)<\/span>
    \n raise<\/span>
    \n rep.<\/span>add(<\/span>comp,<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t)<\/span> *<\/span> 1000<\/span>,<\/span> BUDGETS[<\/span>comp]<\/span>,<\/span> True<\/span>)<\/span>
    \n finally<\/span>:<\/span>
    \n if<\/span> not<\/span> degraded:<\/span>
    \n rep.<\/span>add(<\/span>comp,<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t)<\/span> *<\/span> 1000<\/span>,<\/span> BUDGETS[<\/span>comp]<\/span>,<\/span> False<\/span>)<\/span><\/p>\n

    async<\/span> def<\/span> stage<\/span>(<\/span>comp:<\/span> str<\/span>,<\/span> coro,<\/span> rep:<\/span> LatencyReport,<\/span> fallback=<\/span>None<\/span>)<\/span>:<\/span>
    \n """\u628a\u4e00\u4e2a\u9636\u6bb5\u5305\u8d77\u6765：\u6b63\u5e38\u8fd4\u56de\u7ed3\u679c，\u8d85\u65f6\u8fd4\u56de fallback \u7684\u7ed3\u679c\u3002"""<\/span>
    \n try<\/span>:<\/span>
    \n async<\/span> with<\/span> budget(<\/span>comp,<\/span> rep,<\/span> fallback)<\/span>:<\/span>
    \n return<\/span> await<\/span> coro,<\/span> False<\/span>
    \n except<\/span> BudgetExceeded:<\/span>
    \n if<\/span> fallback is<\/span> None<\/span>:<\/span>
    \n return<\/span> None<\/span>,<\/span> True<\/span>
    \n return<\/span> await<\/span> fallback,<\/span> True<\/span><\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n async<\/span> def<\/span> slow<\/span>(<\/span>ms:<\/span> float<\/span>,<\/span> value)<\/span>:<\/span>
    \n await<\/span> asyncio.<\/span>sleep(<\/span>ms \/<\/span> 1000<\/span>)<\/span>
    \n return<\/span> value<\/p>\n

    async<\/span> def<\/span> main<\/span>(<\/span>)<\/span>:<\/span>
    \n rep =<\/span> LatencyReport(<\/span>)<\/span>
    \n await<\/span> stage(<\/span>"classify"<\/span>,<\/span> slow(<\/span>12<\/span>,<\/span> "allow"<\/span>)<\/span>,<\/span> rep)<\/span>
    \n await<\/span> stage(<\/span>"dense"<\/span>,<\/span> slow(<\/span>180<\/span>,<\/span> [<\/span>"c1"<\/span>,<\/span> "c2"<\/span>]<\/span>)<\/span>,<\/span> rep)<\/span>
    \n await<\/span> stage(<\/span>"bm25"<\/span>,<\/span> slow(<\/span>90<\/span>,<\/span> [<\/span>"c3"<\/span>]<\/span>)<\/span>,<\/span> rep)<\/span>
    \n await<\/span> stage(<\/span>"rerank"<\/span>,<\/span> slow(<\/span>650<\/span>,<\/span> "top8"<\/span>)<\/span>,<\/span> rep,<\/span> fallback=<\/span>slow(<\/span>5<\/span>,<\/span> "rrf_top8"<\/span>)<\/span>)<\/span>
    \n await<\/span> stage(<\/span>"assemble"<\/span>,<\/span> slow(<\/span>20<\/span>,<\/span> "ctx"<\/span>)<\/span>,<\/span> rep)<\/span>
    \n print<\/span>(<\/span>rep.<\/span>render(<\/span>)<\/span>)<\/span><\/p>\n

    asyncio.<\/span>run(<\/span>main(<\/span>)<\/span>)<\/span><\/p>\n

    \u8f93\u51fa\u91cc rerank \u90a3\u4e00\u884c\u4f1a\u88ab\u6807\u6210 DEGRADED：<\/p>\n

    component ms budget status
    \nclassify 12.4 30 ok
    \ndense 181.7 200 ok
    \nbm25 91.3 120 ok
    \nrerank 650.8 400 DEGRADED
    \nassemble 20.6 30 ok
    \nTOTAL 956.8<\/p>\n

    \u6ce8\u610f rerank \u90a3\u4e00\u884c\u8bb0\u5f55\u7684\u662f 650.8ms \u800c\u4e0d\u662f 400ms\u2014\u2014\u9884\u7b97\u7684\u4f5c\u7528\u662f\u300c\u51b3\u5b9a\u4ec0\u4e48\u65f6\u5019\u5207\u6362\u8def\u5f84\u300d，\u4e0d\u662f\u300c\u628a\u6162\u7ec4\u4ef6\u53d8\u5feb\u300d\u3002\u5b9e\u9645\u8017\u65f6\u4ecd\u7136\u8981\u5982\u5b9e\u8bb0\u5f55，\u56e0\u4e3a\u8fd9\u662f\u4f60\u8981\u62ff\u53bb\u505a\u5bb9\u91cf\u89c4\u5212\u548c\u6a21\u578b\u66ff\u6362\u7684\u4f9d\u636e\u3002<\/p>\n

    5.8 \u4e3b\u6d41\u7a0b\u4e32\u8d77\u6765<\/h4>\n

    #!\/usr\/bin\/env python3<\/span>
    \n# -*- coding: utf-8 -*-<\/span>
    \n"""pipeline.py \u2014 \u4e09\u4ee3\u80fd\u529b\u4e32\u6210\u4e00\u6761\u5e26\u9884\u7b97\u3001\u5e26\u8fb9\u754c\u3001\u5e26\u7f6e\u4fe1\u5ea6\u7684\u6d41\u6c34\u7ebf\u3002"""<\/span>
    \nfrom<\/span> __future__ import<\/span> annotations<\/p>\n

    import<\/span> argparse
    \nimport<\/span> asyncio
    \nimport<\/span> json
    \nfrom<\/span> dataclasses import<\/span> dataclass<\/p>\n

    from<\/span> confidence import<\/span> Confidence,<\/span> Factor,<\/span> RetryState,<\/span> decide,<\/span> Stop
    \nfrom<\/span> dedup import<\/span> Candidate,<\/span> dedup
    \nfrom<\/span> guardrail import<\/span> classify_query,<\/span> acl_filter,<\/span> redact
    \nfrom<\/span> hybrid import<\/span> hybrid_search
    \nfrom<\/span> rerank import<\/span> CrossEncoderReranker<\/p>\n

    @dataclass<\/span>
    \nclass<\/span> Answer<\/span>:<\/span>
    \n text:<\/span> str<\/span>
    \n citations:<\/span> list<\/span>[<\/span>dict<\/span>]<\/span>
    \n confidence:<\/span> float<\/span>
    \n worst_factor:<\/span> tuple<\/span>[<\/span>str<\/span>,<\/span> float<\/span>]<\/span>
    \n degraded:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>
    \n stop_reason:<\/span> str<\/span><\/p>\n

    class<\/span> Pipeline<\/span>:<\/span>
    \n def<\/span> __init__<\/span>(<\/span>self,<\/span> dense,<\/span> sparse,<\/span> encode,<\/span> reranker,<\/span> extractor=<\/span>None<\/span>,<\/span>
    \n total_budget_s:<\/span> float<\/span> =<\/span> 3.0<\/span>,<\/span> threshold:<\/span> float<\/span> =<\/span> 0.55<\/span>)<\/span>:<\/span>
    \n self.<\/span>dense,<\/span> self.<\/span>sparse,<\/span> self.<\/span>encode =<\/span> dense,<\/span> sparse,<\/span> encode
    \n self.<\/span>reranker =<\/span> reranker
    \n self.<\/span>extractor =<\/span> extractor
    \n self.<\/span>total_budget_s =<\/span> total_budget_s
    \n self.<\/span>threshold =<\/span> threshold<\/p>\n

    async<\/span> def<\/span> run<\/span>(<\/span>self,<\/span> query:<\/span> str<\/span>,<\/span> user_groups:<\/span> set<\/span>[<\/span>str<\/span>]<\/span>)<\/span> –<\/span>><\/span> Answer:<\/span>
    \n degraded:<\/span> list<\/span>[<\/span>str<\/span>]<\/span> =<\/span> [<\/span>]<\/span>
    \n conf =<\/span> Confidence(<\/span>)<\/span>
    \n state =<\/span> RetryState(<\/span>max_attempts=<\/span>3<\/span>,<\/span> deadline_s=<\/span>self.<\/span>total_budget_s)<\/span><\/p>\n

    # —- \u7b2c 1 \u9053：\u67e5\u8be2\u4fa7\u5206\u7c7b（\u68c0\u7d22\u4e4b\u524d）—-<\/span>
    \n cls =<\/span> classify_query(<\/span>query)<\/span>
    \n if<\/span> cls.<\/span>action ==<\/span> "block"<\/span>:<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>COMPLIANCE,<\/span> 0.0<\/span>)<\/span>
    \n return<\/span> Answer(<\/span>""<\/span>,<\/span> [<\/span>]<\/span>,<\/span> 0.0<\/span>,<\/span> (<\/span>"compliance"<\/span>,<\/span> 0.0<\/span>)<\/span>,<\/span> [<\/span>]<\/span>,<\/span>
    \n f"\u67e5\u8be2\u547d\u4e2d\u654f\u611f\u6a21\u5f0f，\u5df2\u62e6\u622a：<\/span>{<\/span>cls.<\/span>categories}<\/span><\/span>"<\/span><\/span>)<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>COMPLIANCE,<\/span> 1.0<\/span>)<\/span><\/p>\n

    while<\/span> True<\/span>:<\/span>
    \n # —- \u6df7\u5408\u68c0\u7d22（\u4e24\u8def\u5e76\u53d1）—-<\/span>
    \n res =<\/span> await<\/span> hybrid_search(<\/span>query,<\/span> self.<\/span>dense,<\/span> self.<\/span>sparse,<\/span> self.<\/span>encode,<\/span>
    \n top_k=<\/span>50<\/span> if<\/span> state.<\/span>strategy_idx <<\/span> 3<\/span> else<\/span> 150<\/span>)<\/span>
    \n fused =<\/span> res[<\/span>"fused"<\/span>]<\/span><\/p>\n

    # —- \u4e09\u5c42\u53bb\u91cd（\u5fc5\u987b\u5728 rerank \u4e4b\u524d）—-<\/span>
    \n cands =<\/span> [<\/span>Candidate(<\/span>cid,<\/span> cid.<\/span>split(<\/span>"#"<\/span>)<\/span>[<\/span>0<\/span>]<\/span>,<\/span> ""<\/span>,<\/span> 0<\/span>,<\/span>
    \n self.<\/span>sparse.<\/span>ids and<\/span> ""<\/span> or<\/span> ""<\/span>,<\/span> s)<\/span> for<\/span> cid,<\/span> s in<\/span> fused]<\/span>
    \n kept,<\/span> st =<\/span> dedup(<\/span>cands)<\/span><\/p>\n

    # —- \u7b2c 2 \u9053：ACL \u8fc7\u6ee4（\u68c0\u7d22\u5668\u51fa\u53e3）—-<\/span>
    \n kept =<\/span> acl_filter(<\/span>kept,<\/span> user_groups)<\/span><\/p>\n

    # —- cross-encoder \u91cd\u6392 —-<\/span>
    \n rr =<\/span> self.<\/span>reranker.<\/span>rerank(<\/span>query,<\/span> [<\/span>(<\/span>c.<\/span>chunk_id,<\/span> c.<\/span>text,<\/span> c.<\/span>score)<\/span> for<\/span> c in<\/span> kept]<\/span>)<\/span>
    \n if<\/span> rr.<\/span>degraded:<\/span>
    \n degraded.<\/span>append(<\/span>"rerank"<\/span>)<\/span>
    \n top =<\/span> rr.<\/span>ranked[<\/span>:<\/span>8<\/span>]<\/span><\/p>\n

    # —- \u7f6e\u4fe1\u5ea6：\u4e58\u6cd5，\u4e0d\u662f\u5e73\u5747 —-<\/span>
    \n grounded =<\/span> self.<\/span>_groundedness(<\/span>top,<\/span> kept)<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>RETRIEVAL,<\/span> grounded)<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>GENERATION,<\/span> 0.85<\/span> if<\/span> not<\/span> degraded else<\/span> 0.70<\/span>)<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>TOOL,<\/span> 1.0<\/span> if<\/span> not<\/span> degraded else<\/span> 0.80<\/span>)<\/span><\/p>\n

    stop,<\/span> why =<\/span> decide(<\/span>conf,<\/span> state,<\/span> query,<\/span> self.<\/span>threshold)<\/span>
    \n if<\/span> stop ==<\/span> Stop.<\/span>RETRY:<\/span>
    \n query =<\/span> self.<\/span>_rewrite(<\/span>query,<\/span> state.<\/span>strategy_idx)<\/span>
    \n continue<\/span>
    \n if<\/span> stop ==<\/span> Stop.<\/span>REFUSE:<\/span>
    \n return<\/span> Answer(<\/span>""<\/span>,<\/span> [<\/span>]<\/span>,<\/span> conf.<\/span>product(<\/span>)<\/span>,<\/span> conf.<\/span>worst(<\/span>)<\/span>,<\/span> degraded,<\/span> why)<\/span><\/p>\n

    # —- \u7b2c 3 \u9053：\u751f\u6210\u524d\u590d\u6838 —-<\/span>
    \n ctx =<\/span> redact(<\/span>self.<\/span>_assemble(<\/span>top)<\/span>)<\/span>
    \n conf.<\/span>set<\/span>(<\/span>Factor.<\/span>COMPLIANCE,<\/span> 0.0<\/span> if<\/span> "[REDACTED"<\/span> in<\/span> ctx else<\/span> 1.0<\/span>)<\/span>
    \n if<\/span> conf.<\/span>blocking(<\/span>)<\/span>:<\/span>
    \n return<\/span> Answer(<\/span>""<\/span>,<\/span> [<\/span>]<\/span>,<\/span> conf.<\/span>product(<\/span>)<\/span>,<\/span> conf.<\/span>worst(<\/span>)<\/span>,<\/span> degraded,<\/span>
    \n "\u4e0a\u4e0b\u6587\u547d\u4e2d\u654f\u611f\u6a21\u5f0f，\u5df2\u62e6\u622a"<\/span>)<\/span><\/p>\n

    text =<\/span> await<\/span> self.<\/span>_generate(<\/span>query,<\/span> ctx)<\/span>
    \n return<\/span> Answer(<\/span>text,<\/span> self.<\/span>_citations(<\/span>top)<\/span>,<\/span> conf.<\/span>product(<\/span>)<\/span>,<\/span>
    \n conf.<\/span>worst(<\/span>)<\/span>,<\/span> degraded,<\/span> why)<\/span><\/p>\n

    # —- \u4ee5\u4e0b\u4e3a\u793a\u610f\u5b9e\u73b0 —-<\/span>
    \n def<\/span> _groundedness<\/span>(<\/span>self,<\/span> top,<\/span> kept)<\/span> –<\/span>><\/span> float<\/span>:<\/span>
    \n if<\/span> not<\/span> top:<\/span>
    \n return<\/span> 0.0<\/span>
    \n hi =<\/span> sum<\/span>(<\/span>1<\/span> for<\/span> _cid,<\/span> s in<\/span> top if<\/span> s ><\/span> 0.5<\/span>)<\/span>
    \n return<\/span> min<\/span>(<\/span>1.0<\/span>,<\/span> 0.35<\/span> +<\/span> 0.65<\/span> *<\/span> (<\/span>hi \/<\/span> len<\/span>(<\/span>top)<\/span>)<\/span> +<\/span> 0.10<\/span> *<\/span> (<\/span>len<\/span>(<\/span>top)<\/span> >=<\/span> 5<\/span>)<\/span>)<\/span><\/p>\n

    def<\/span> _rewrite<\/span>(<\/span>self,<\/span> q:<\/span> str<\/span>,<\/span> idx:<\/span> int<\/span>)<\/span> –<\/span>><\/span> str<\/span>:<\/span>
    \n return<\/span> {<\/span>0<\/span>:<\/span> q,<\/span> 1<\/span>:<\/span> q +<\/span> " \u76f8\u5173\u6982\u5ff5 \u540c\u4e49\u8bcd"<\/span>,<\/span>
    \n 2<\/span>:<\/span> q +<\/span> " \u4e0a\u4e0b\u6e38\u4f9d\u8d56 \u6240\u5c5e\u4ea7\u54c1\u65cf"<\/span>,<\/span> 3<\/span>:<\/span> q,<\/span> 4<\/span>:<\/span> q}<\/span>[<\/span>min<\/span>(<\/span>idx,<\/span> 4<\/span>)<\/span>]<\/span><\/p>\n

    def<\/span> _assemble<\/span>(<\/span>self,<\/span> top)<\/span> –<\/span>><\/span> str<\/span>:<\/span>
    \n return<\/span> "\\\\n\\\\n"<\/span>.<\/span>join(<\/span>f"[<\/span>{<\/span>i +<\/span> 1<\/span>}<\/span><\/span>] <\/span>{<\/span>cid}<\/span><\/span>"<\/span><\/span> for<\/span> i,<\/span> (<\/span>cid,<\/span> _s)<\/span> in<\/span> enumerate<\/span>(<\/span>top)<\/span>)<\/span><\/p>\n

    def<\/span> _citations<\/span>(<\/span>self,<\/span> top)<\/span> –<\/span>><\/span> list<\/span>[<\/span>dict<\/span>]<\/span>:<\/span>
    \n return<\/span> [<\/span>{<\/span>"index"<\/span>:<\/span> i +<\/span> 1<\/span>,<\/span> "chunk_id"<\/span>:<\/span> cid,<\/span> "score"<\/span>:<\/span> round<\/span>(<\/span>s,<\/span> 4<\/span>)<\/span>}<\/span>
    \n for<\/span> i,<\/span> (<\/span>cid,<\/span> s)<\/span> in<\/span> enumerate<\/span>(<\/span>top)<\/span>]<\/span><\/p>\n

    async<\/span> def<\/span> _generate<\/span>(<\/span>self,<\/span> q:<\/span> str<\/span>,<\/span> ctx:<\/span> str<\/span>)<\/span> –<\/span>><\/span> str<\/span>:<\/span>
    \n return<\/span> f"（\u751f\u6210\u5360\u4f4d）\u57fa\u4e8e <\/span>{<\/span>len<\/span>(<\/span>ctx.<\/span>splitlines(<\/span>)<\/span>)<\/span>}<\/span><\/span> \u6761\u5f15\u7528\u56de\u7b54：<\/span>{<\/span>q}<\/span><\/span>"<\/span><\/span><\/p>\n

    def<\/span> main<\/span>(<\/span>)<\/span> –<\/span>><\/span> int<\/span>:<\/span>
    \n ap =<\/span> argparse.<\/span>ArgumentParser(<\/span>)<\/span>
    \n ap.<\/span>add_argument(<\/span>"–query"<\/span>,<\/span> required=<\/span>True<\/span>)<\/span>
    \n ap.<\/span>add_argument(<\/span>"–groups"<\/span>,<\/span> default=<\/span>"all"<\/span>)<\/span>
    \n args =<\/span> ap.<\/span>parse_args(<\/span>)<\/span><\/p>\n

    # \u6f14\u793a\u88c5\u914d：\u771f\u5b9e\u9879\u76ee\u4ece\u78c1\u76d8\u52a0\u8f7d\u7d22\u5f15<\/span>
    \n from<\/span> hybrid import<\/span> DenseIndex,<\/span> SparseIndex
    \n import<\/span> numpy as<\/span> np
    \n rng =<\/span> np.<\/span>random.<\/span>default_rng(<\/span>7<\/span>)<\/span>
    \n ids =<\/span> [<\/span>f"c<\/span>{<\/span>i}<\/span><\/span>"<\/span><\/span> for<\/span> i in<\/span> range<\/span>(<\/span>500<\/span>)<\/span>]<\/span>
    \n dense =<\/span> DenseIndex(<\/span>dim=<\/span>64<\/span>)<\/span>
    \n dense.<\/span>add(<\/span>ids,<\/span> rng.<\/span>normal(<\/span>size=<\/span>(<\/span>500<\/span>,<\/span> 64<\/span>)<\/span>)<\/span>.<\/span>astype(<\/span>"float32"<\/span>)<\/span>)<\/span>
    \n sparse =<\/span> SparseIndex(<\/span>tokenizer=<\/span>lambda<\/span> s:<\/span> list<\/span>(<\/span>s)<\/span>)<\/span>
    \n sparse.<\/span>build(<\/span>ids,<\/span> [<\/span>f"\u7b2c <\/span>{<\/span>i}<\/span><\/span> \u6bb5\u5185\u5bb9"<\/span><\/span> for<\/span> i in<\/span> range<\/span>(<\/span>500<\/span>)<\/span>]<\/span>)<\/span><\/p>\n

    p =<\/span> Pipeline(<\/span>dense,<\/span> sparse,<\/span>
    \n encode=<\/span>lambda<\/span> q:<\/span> rng.<\/span>normal(<\/span>size=<\/span>64<\/span>)<\/span>.<\/span>astype(<\/span>"float32"<\/span>)<\/span>,<\/span>
    \n reranker=<\/span>CrossEncoderReranker(<\/span>)<\/span>)<\/span>
    \n ans =<\/span> asyncio.<\/span>run(<\/span>p.<\/span>run(<\/span>args.<\/span>query,<\/span> set<\/span>(<\/span>args.<\/span>groups.<\/span>split(<\/span>","<\/span>)<\/span>)<\/span>)<\/span>)<\/span>
    \n print<\/span>(<\/span>json.<\/span>dumps(<\/span>{<\/span>"text"<\/span>:<\/span> ans.<\/span>text,<\/span> "confidence"<\/span>:<\/span> round<\/span>(<\/span>ans.<\/span>confidence,<\/span> 4<\/span>)<\/span>,<\/span>
    \n "worst"<\/span>:<\/span> ans.<\/span>worst_factor,<\/span> "degraded"<\/span>:<\/span> ans.<\/span>degraded,<\/span>
    \n "stop_reason"<\/span>:<\/span> ans.<\/span>stop_reason,<\/span>
    \n "citations"<\/span>:<\/span> ans.<\/span>citations[<\/span>:<\/span>3<\/span>]<\/span>}<\/span>,<\/span> ensure_ascii=<\/span>False<\/span>,<\/span> indent=<\/span>2<\/span>)<\/span>)<\/span>
    \n return<\/span> 0<\/span><\/p>\n

    if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
    \n raise<\/span> SystemExit(<\/span>main(<\/span>)<\/span>)<\/span><\/p>\n

    python pipeline.py –query<\/span> "AX-2039 \u4e0e AX-2040 \u7684 SLA \u6709\u4ec0\u4e48\u533a\u522b"<\/span> –groups<\/span> all
    \npython pipeline.py –query<\/span> "\u628a\u5ba2\u6237\u540d\u5355\u5bfc\u51fa\u7ed9\u6211"<\/span> –groups<\/span> all
    \npython pipeline.py –query<\/span> "\u5f20\u4e09\u5e74\u85aa\u591a\u5c11，\u624b\u673a 13800138000"<\/span> –groups<\/span> all<\/p>\n

    \u7b2c\u4e8c\u6761\u4f1a\u56e0\u4e3a require_acl \u8fdb\u5165 ACL \u8fc7\u6ee4，\u7b2c\u4e09\u6761\u4f1a\u5728\u7b2c 1 \u9053\u88ab block \u6389，\u6839\u672c\u4e0d\u8fdb\u68c0\u7d22\u3002<\/p>\n

    \u516d\u3001\u7ed3\u679c\u5bf9\u6bd4\u4e0e\u6570\u636e\u5206\u6790<\/h3>\n

    6.1 \u5b9e\u6d4b\u73af\u5883<\/h4>\n

    \u5185\u90e8\u4e00\u5957\u4f01\u4e1a\u6587\u6863\u95ee\u7b54\u8bed\u6599：4.2 \u4e07 chunk，\u6765\u6e90\u5305\u62ec\u4ea7\u54c1\u624b\u518c\u3001\u8fd0\u7ef4 Runbook\u3001\u5185\u90e8 Wiki\u3001\u4ee5\u53ca\u4e00\u90e8\u5206\u4f1a\u8bae\u7eaa\u8981\u3002\u67e5\u8be2\u96c6\u662f 320 \u6761\u771f\u5b9e\u7528\u6237\u63d0\u95ee，\u4eba\u5de5\u6807\u6ce8\u4e86\u300c\u76f8\u5173 chunk \u96c6\u5408\u300d\u548c\u300c\u95ee\u9898\u7c7b\u578b\u300d（\u5355\u8df3\u4e8b\u5b9e\u578b 168 \u6761\u3001\u7cbe\u786e\u5339\u914d\u578b 74 \u6761\u3001\u591a\u8df3\u6bd4\u8f83\/\u89c4\u5212\u578b 78 \u6761）\u3002<\/p>\n

    \u786c\u4ef6\u662f 8 \u6838 CPU\u300132GB \u5185\u5b58，\u65e0 GPU\u3002\u6240\u6709\u5ef6\u8fdf\u6570\u5b57\u662f P50，reranker \u5728 CPU \u4e0a\u8dd1\u3002<\/p>\n

    6.2 \u4e09\u4ee3\u65b9\u6848\u7684\u5206\u7c7b\u578b\u8868\u73b0<\/h4>\n\n\u65b9\u6848\u5355\u8df3\u4e8b\u5b9e\u578b nDCG@8\u7cbe\u786e\u5339\u914d\u578b nDCG@8\u591a\u8df3\u578b nDCG@8\u5168\u91cf nDCG@8P50 \u5ef6\u8fdfP95 \u5ef6\u8fdf<\/tr>\n\n\n\n\n\n\n
    \u6734\u7d20 RAG（\u5355\u8def dense）<\/td>\n0.612<\/td>\n0.318<\/td>\n0.241<\/td>\n0.486<\/td>\n196ms<\/td>\n412ms<\/td>\n<\/tr>\n
    \u7b2c\u4e00\u4ee3（\u6df7\u5408 + RRF + \u53bb\u91cd + rerank）<\/td>\n0.741<\/td>\n0.826<\/td>\n0.352<\/td>\n0.703<\/td>\n287ms<\/td>\n561ms<\/td>\n<\/tr>\n
    \u7b2c\u4e00\u4ee3（\u4e24\u8def\u5e76\u53d1\u540e）<\/td>\n0.741<\/td>\n0.826<\/td>\n0.352<\/td>\n0.703<\/td>\n168ms<\/td>\n338ms<\/td>\n<\/tr>\n
    \u7b2c\u4e8c\u4ee3（+ GraphRAG \u56fe\u6269\u5c55）<\/td>\n0.748<\/td>\n0.839<\/td>\n0.517<\/td>\n0.741<\/td>\n512ms<\/td>\n1090ms<\/td>\n<\/tr>\n
    \u7b2c\u4e09\u4ee3（+ Agentic \u89c4\u5212\u4e0e\u91cd\u8bd5）<\/td>\n0.752<\/td>\n0.842<\/td>\n0.688<\/td>\n0.775<\/td>\n3140ms<\/td>\n7820ms<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u8bfb\u8fd9\u5f20\u8868\u8981\u6293\u4e09\u4e2a\u70b9：<\/p>\n

    **\u7b2c\u4e00，\u7b2c\u4e00\u4ee3\u5bf9\u300c\u7cbe\u786e\u5339\u914d\u578b\u300d\u7684\u63d0\u5347\u662f\u6700\u5927\u7684：0.318 \u5230 0.826\u3002**\u8fd9\u662f\u56e0\u4e3a\u8fd9 74 \u6761\u67e5\u8be2\u91cc\u6709 61 \u6761\u5e26\u4ea7\u54c1\u7801\u6216\u9519\u8bef\u7801，\u5355\u8def dense \u57fa\u672c\u6293\u4e0d\u4f4f\u3002\u5982\u679c\u4f60\u53ea\u6d4b\u5355\u8df3\u4e8b\u5b9e\u578b，\u4f1a\u4e25\u91cd\u4f4e\u4f30 BM25 \u7684\u4ef7\u503c\u3002<\/p>\n

    **\u7b2c\u4e8c，\u5e76\u53d1\u4f18\u5316\u628a P50 \u4ece 287ms \u538b\u5230 168ms，\u964d 41.5%，\u800c nDCG \u4e00\u4e2a\u5c0f\u6570\u70b9\u90fd\u6ca1\u53d8\u3002**\u8fd9\u662f\u7eaf\u5de5\u7a0b\u6536\u76ca，\u96f6\u8d28\u91cf\u4ee3\u4ef7，\u4f46\u5f88\u591a\u56e2\u961f\u6ca1\u505a\u3002<\/p>\n

    **\u7b2c\u4e09，\u7b2c\u4e09\u4ee3\u5728\u591a\u8df3\u578b\u4e0a\u4ece 0.517 \u63d0\u5230 0.688，\u4ee3\u4ef7\u662f P50 \u4ece 512ms \u6da8\u5230 3140ms\u3002**6 \u500d\u5ef6\u8fdf\u6362 0.17 \u7684 nDCG\u3002\u8fd9\u4e2a\u4ea4\u6613\u53ea\u5728\u591a\u8df3\u67e5\u8be2\u5360\u6bd4\u8db3\u591f\u9ad8\u65f6\u624d\u5212\u7b97\u2014\u2014\u8fd9\u662f\u4e00\u4e2a\u5fc5\u987b\u6309\u4f60\u81ea\u5df1\u7684\u67e5\u8be2\u5206\u5e03\u6765\u7b97\u7684\u8d26，\u4e0d\u80fd\u7167\u6284\u3002<\/p>\n

    6.3 \u5404\u7ec4\u4ef6\u7684\u8fb9\u9645\u8d21\u732e<\/h4>\n

    \u9010\u9879\u52a0，\u770b\u6bcf\u4e00\u6b65\u503c\u591a\u5c11（\u5168\u91cf nDCG@8）：<\/p>\n\n\u914d\u7f6enDCG@8\u76f8\u5bf9\u4e0a\u4e00\u6b65\u589e\u76caP50 \u5ef6\u8fdf\u589e\u91cf<\/tr>\n\n\n\n\n\n\n\n\n
    \u5355\u8def dense<\/td>\n0.486<\/td>\n\u2014<\/td>\n\u2014<\/td>\n<\/tr>\n
    + BM25（RRF \u878d\u5408）<\/td>\n0.624<\/td>\n+0.138<\/td>\n+52ms<\/td>\n<\/tr>\n
    + \u4e24\u8def\u5e76\u53d1<\/td>\n0.624<\/td>\n+0.000<\/td>\n-119ms（\u8d1f\u589e\u91cf）<\/td>\n<\/tr>\n
    + \u4e09\u5c42\u53bb\u91cd<\/td>\n0.671<\/td>\n+0.047<\/td>\n+3ms<\/td>\n<\/tr>\n
    + cross-encoder rerank<\/td>\n0.703<\/td>\n+0.032<\/td>\n+121ms<\/td>\n<\/tr>\n
    + \u56fe\u6269\u5c55<\/td>\n0.741<\/td>\n+0.038<\/td>\n+344ms<\/td>\n<\/tr>\n
    + Agentic \u89c4\u5212\u4e0e\u91cd\u8bd5<\/td>\n0.775<\/td>\n+0.034<\/td>\n+2628ms<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u6309\u300c\u6bcf 100ms \u6362\u591a\u5c11 nDCG\u300d\u6392\u4e2a\u5e8f：\u4e09\u5c42\u53bb\u91cd（1.57 \/ 100ms）> BM25 \u878d\u5408（0.27）> \u5e76\u53d1（\u8d1f\u5ef6\u8fdf）> rerank（0.026）> \u56fe\u6269\u5c55（0.011）> Agentic（0.0013）\u3002<\/p>\n

    \u8fd9\u4e2a\u6392\u5e8f\u7ed9\u4e86\u4e00\u4e2a\u5f88\u660e\u786e\u7684\u6295\u5165\u4f18\u5148\u7ea7：\u5148\u628a BM25 \u548c\u53bb\u91cd\u505a\u624e\u5b9e，\u8fd9\u4e24\u6b65\u52a0\u8d77\u6765 0.185 \u7684 nDCG，\u53ea\u82b1 55ms\u3002rerank \u548c\u56fe\u6269\u5c55\u662f\u7b2c\u4e8c\u68af\u961f\u3002Agentic \u7684\u6027\u4ef7\u6bd4\u6700\u4f4e，\u5b83\u5b58\u5728\u7684\u7406\u7531\u4e0d\u662f\u6027\u4ef7\u6bd4，\u800c\u662f\u300c\u6709\u4e9b\u95ee\u9898\u4e0d\u7528\u5b83\u6839\u672c\u7b54\u4e0d\u4e86\u300d\u3002<\/p>\n

    6.4 \u5b9e\u4f53\u62bd\u53d6\u65b9\u6848\u5bf9\u6bd4<\/h4>\n\n\u65b9\u68488 \u4e07 chunk \u5168\u91cf\u8017\u65f6\u8fb9\u9645\u6210\u672c\u53ef\u590d\u73b0\u7cbe\u786e\u7387\u53ec\u56de\u7387\u65b0\u5b9e\u4f53\u53d1\u73b0<\/tr>\n\n\n\n\n
    \u89c4\u5219\u5316\u4e09\u8d9f<\/td>\n3.3s<\/td>\n0<\/td>\n\u5b8c\u5168\u4e00\u81f4<\/td>\n0.94<\/td>\n0.71<\/td>\n\u4e0d\u80fd<\/td>\n<\/tr>\n
    LLM NER（\u5e76\u53d1 16）<\/td>\n\u7ea6 75min<\/td>\n\u9ad8<\/td>\n\u5dee<\/td>\n0.89<\/td>\n0.83<\/td>\n\u80fd<\/td>\n<\/tr>\n
    \u6df7\u5408（\u89c4\u5219 + \u957f\u5c3e LLM \u5ba1\u6838）<\/td>\n\u7ea6 6min<\/td>\n\u4f4e<\/td>\n\u4e3b\u4f53\u4e00\u81f4<\/td>\n0.93<\/td>\n0.79<\/td>\n\u80fd，\u4f46\u9700\u5ba1\u6838<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u89c4\u5219\u5316\u65b9\u6848\u7cbe\u786e\u7387\u66f4\u9ad8（0.94 \u5bf9 0.89），\u56e0\u4e3a\u8bcd\u5178\u6765\u81ea\u6743\u5a01\u6e90，\u62bd\u51fa\u6765\u7684\u90fd\u662f\u6709\u5b9a\u4e49\u7684\u5b9e\u4f53；LLM \u53ec\u56de\u66f4\u9ad8（0.83 \u5bf9 0.71），\u56e0\u4e3a\u5b83\u80fd\u53d1\u73b0\u8bcd\u5178\u91cc\u6ca1\u6709\u7684\u5b9e\u4f53\u3002\u4f46 LLM \u591a\u51fa\u6765\u7684\u90a3\u90e8\u5206\u53ec\u56de\u91cc，\u6709\u76f8\u5f53\u6bd4\u4f8b\u662f\u300c\u62bd\u5bf9\u4e86\u4f46\u6ca1\u7528\u300d\u7684\u5b9e\u4f53\u2014\u2014\u6bd4\u5982\u4ece\u4f1a\u8bae\u7eaa\u8981\u91cc\u62bd\u51fa\u4e00\u4e2a\u4eba\u540d，\u8fd9\u4e2a\u4eba\u540d\u5728\u56fe\u8c31\u91cc\u662f\u5b64\u7acb\u8282\u70b9，\u6ca1\u6709\u4efb\u4f55\u8fb9，\u5bf9\u68c0\u7d22\u6beb\u65e0\u5e2e\u52a9\u3002<\/p>\n

    \u6df7\u5408\u65b9\u6848\u662f\u6211\u6700\u7ec8\u7684\u9009\u62e9：\u89c4\u5219\u5316\u8dd1\u5168\u91cf，LLM \u53ea\u5904\u7406\u300c\u9ad8\u9891\u88ab\u68c0\u7d22\u4f46\u96f6\u5b9e\u4f53\u547d\u4e2d\u300d\u7684 chunk，\u7ed3\u679c\u8fdb\u5ba1\u6838\u961f\u5217\u3002\u8fd9\u6837\u628a\u975e\u786e\u5b9a\u6027\u5173\u5728\u95f8\u95e8\u540e\u9762，\u540c\u65f6\u4fdd\u7559\u4e86\u53d1\u73b0\u65b0\u5b9e\u4f53\u7684\u80fd\u529b\u3002<\/p>\n

    6.5 \u7f6e\u4fe1\u5ea6\u7b56\u7565\u7684\u6548\u679c<\/h4>\n

    \u5728 320 \u6761\u67e5\u8be2\u4e0a\u5bf9\u6bd4\u4e09\u79cd\u7f6e\u4fe1\u5ea6\u7b56\u7565，\u7528\u300c\u4eba\u5de5\u5224\u5b9a\u7b54\u6848\u662f\u5426\u53ef\u63a5\u53d7\u300d\u4f5c\u4e3a\u91d1\u6807\u51c6：<\/p>\n\n\u7b56\u7565\u62d2\u7b54\u7387\u9519\u8bef\u7b54\u6848\u6f0f\u51fa\u7387\u6709\u6548\u56de\u7b54\u7387\u5e73\u5747\u5ef6\u8fdf<\/tr>\n\n\n\n\n\n
    \u4e0d\u7528\u7f6e\u4fe1\u5ea6（\u603b\u662f\u7b54）<\/td>\n0%<\/td>\n22.8%<\/td>\n77.2%<\/td>\n168ms<\/td>\n<\/tr>\n
    \u5e73\u5747\u7f6e\u4fe1\u5ea6，\u9608\u503c 0.7<\/td>\n11.6%<\/td>\n14.1%<\/td>\n74.3%<\/td>\n412ms<\/td>\n<\/tr>\n
    \u4e58\u6cd5\u7f6e\u4fe1\u5ea6，\u9608\u503c 0.55<\/td>\n13.4%<\/td>\n6.9%<\/td>\n79.7%<\/td>\n486ms<\/td>\n<\/tr>\n
    \u4e58\u6cd5 + \u4e00\u7968\u5426\u51b3（<0.30）<\/td>\n15.0%<\/td>\n4.4%<\/td>\n80.6%<\/td>\n498ms<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u5173\u952e\u770b\u7b2c\u4e09\u5217\u300c\u9519\u8bef\u7b54\u6848\u6f0f\u51fa\u7387\u300d：\u4ece\u5e73\u5747\u7b56\u7565\u7684 14.1% \u964d\u5230\u4e58\u6cd5\u7b56\u7565\u7684 6.9%，\u518d\u964d\u5230\u52a0\u4e86\u4e00\u7968\u5426\u51b3\u7684 4.4%\u3002\u800c\u62d2\u7b54\u7387\u53ea\u4ece 11.6% \u5347\u5230 15.0%\u3002\u4e58\u6cd5\u7684\u4ef7\u503c\u4e0d\u5728\u62d2\u5f97\u591a，\u800c\u5728\u62d2\u5f97\u51c6\u2014\u2014\u5b83\u62d2\u6389\u7684\u6b63\u662f\u90a3\u4e9b\u300c\u6709\u4e00\u4e2a\u73af\u8282\u5d29\u4e86\u300d\u7684\u67e5\u8be2\u3002<\/p>\n

    \u53cd\u8fc7\u6765\u770b\u5e73\u5747\u7b56\u7565：\u5b83\u62d2\u4e86 11.6% \u7684\u67e5\u8be2，\u4f46\u6f0f\u51fa\u4e86 14.1% \u7684\u9519\u8bef\u7b54\u6848\u3002\u4e5f\u5c31\u662f\u8bf4\u5b83\u65e2\u62d2\u9519\u4e86\u4e00\u6279（\u628a\u597d\u7b54\u6848\u62d2\u4e86），\u53c8\u653e\u9519\u4e86\u4e00\u6279\u3002\u8fd9\u6b63\u662f 3.8 \u8282\u8bf4\u7684\u300c\u63a9\u76d6\u5f31\u73af\u8282\u300d\u7684\u540e\u679c\u3002<\/p>\n

    \u4e03\u3001\u8e29\u5751\u8bb0\u5f55\u4e0e\u6392\u67e5\u624b\u518c<\/h3>\n

    \u516b\u6761，\u683c\u5f0f\u56fa\u5b9a\u4e3a\u300c\u73b0\u8c61 \/ \u5b9a\u4f4d \/ \u539f\u56e0 \/ \u89e3\u51b3\u300d\u3002<\/p>\n

    7.1 RRF \u878d\u5408\u524d\u6ca1\u53bb\u91cd，top-5 \u91cc\u4e09\u4e2a\u4f4d\u7f6e\u662f\u540c\u4e00\u6bb5\u5185\u5bb9<\/h4>\n

    \u73b0\u8c61：\u7528\u6237\u6295\u8bc9\u300c\u7cfb\u7edf\u603b\u662f\u91cd\u590d\u8bf4\u540c\u4e00\u53e5\u8bdd\u300d\u3002\u770b top-8 \u7684 chunk_id，\u5404\u4e0d\u76f8\u540c，\u4f46\u5185\u5bb9\u51e0\u4e4e\u4e00\u81f4\u3002<\/p>\n

    \u5b9a\u4f4d：\u6253\u5370\u878d\u5408\u540e\u5019\u9009\u7684\u6587\u672c\u6307\u7eb9，\u53d1\u73b0 8 \u6761\u91cc\u6709 5 \u6761\u7684\u5f52\u4e00\u5316 hash \u76f8\u540c\u3002\u8ffd\u8e2a\u4e0b\u53bb，\u8fd9 5 \u6761\u6765\u81ea\u56db\u4e2a\u4e0d\u540c doc_id\u2014\u2014\u540c\u4e00\u4efd\u4ea7\u54c1\u624b\u518c\u7684\u5b98\u7f51\u7248\u3001Wiki \u7248\u3001PDF \u5bfc\u51fa\u7248，\u4ee5\u53ca\u4e00\u6b21\u57f9\u8bad\u7684\u8f6c\u5f55\u3002<\/p>\n

    \u539f\u56e0：\u6211\u5728\u5b9e\u73b0\u65f6\u628a RRF \u878d\u5408\u76f4\u63a5\u63a5\u5230\u4e86 rerank \u524d\u9762，\u8df3\u8fc7\u4e86\u53bb\u91cd\u5c42\u3002\u800c\u4e14 L1 \u6309 chunk_id \u53bb\u91cd\u5b8c\u5168\u65e0\u6548，\u56e0\u4e3a\u5b83\u4eec\u7684 chunk_id \u672c\u6765\u5c31\u4e0d\u540c\u3002<\/p>\n

    \u89e3\u51b3：\u5728 3.3 \u8282\u7684\u4e09\u5c42\u53bb\u91cd\u91cc\u8865\u4e0a L3 \u5185\u5bb9\u6307\u7eb9\u5c42（\u8de8\u6587\u6863\u7528\u7cbe\u786e hash，\u540c\u6587\u6863\u5185\u7528 simhash \u8fd1\u4f3c）\u3002\u4fee\u597d\u4e4b\u540e top-8 \u7684\u6709\u6548\u4fe1\u606f\u91cf\u4ece\u5e73\u5747 4.1 \u6761\u72ec\u7acb\u5185\u5bb9\u63d0\u5347\u5230 7.4 \u6761，nDCG@8 \u4ece 0.671 \u63d0\u5230 0.703\u3002<\/p>\n

    \u987a\u5e26\u4e00\u6761：\u987a\u5e8f\u5fc5\u987b\u662f\u300c\u878d\u5408 -> \u53bb\u91cd -> \u91cd\u6392\u300d\u3002\u5148\u91cd\u6392\u518d\u53bb\u91cd\u662f\u9519\u7684，\u4f60\u4ed8\u4e86\u4e24\u500d\u91cd\u6392\u6210\u672c\u518d\u6254\u6389\u4e00\u534a\u3002<\/p>\n

    7.2 BM25 \u5206\u8bcd\u5668\u628a\u4ea7\u54c1\u7801\u5207\u788e\u4e86<\/h4>\n

    \u73b0\u8c61：\u67e5 AX-2039-B，BM25 \u4e00\u8def\u8fd4\u56de\u7684 top-50 \u91cc\u5168\u662f\u65e0\u5173\u7684 AX \u5f00\u5934\u6587\u6863，\u6b63\u786e\u7684\u90a3\u6761\u6392\u5728\u7b2c 47 \u4f4d\u3002\u800c\u67e5\u8be2 AX-2039 \u5c31\u6b63\u5e38\u3002<\/p>\n

    \u5b9a\u4f4d：\u6253\u5370\u5206\u8bcd\u7ed3\u679c，jieba.cut("AX-2039-B") \u8f93\u51fa ['AX', '-', '2039', '-', 'B']\u3002\u8fde\u5b57\u7b26\u88ab\u5f53\u6210\u72ec\u7acb token，\u4e8e\u662f\u300c\u5305\u542b AX \u7684\u6587\u6863\u300d\u548c\u300c\u5305\u542b 2039 \u7684\u6587\u6863\u300d\u90fd\u62ff\u5230\u5206\u3002<\/p>\n

    \u539f\u56e0：\u4e2d\u6587\u5206\u8bcd\u5668\u9ed8\u8ba4\u6309\u6807\u70b9\u5207\u5206，\u5bf9\u4ea7\u54c1\u7801\u3001\u9519\u8bef\u7801\u3001\u7248\u672c\u53f7\u8fd9\u7c7b\u5e26\u8fde\u5b57\u7b26\u7684\u6807\u8bc6\u7b26\u4e0d\u53cb\u597d\u3002\u800c BM25 \u662f\u8bcd\u888b\u6a21\u578b，\u4e00\u4e2a 5 token \u7684\u67e5\u8be2\u91cc 4 \u4e2a\u662f\u566a\u58f0 token（-\u3001B），\u4fe1\u53f7\u88ab\u7a00\u91ca\u3002<\/p>\n

    \u89e3\u51b3：\u4e09\u4ef6\u4e8b\u4e00\u8d77\u505a\u3002<\/p>\n

  • \u5206\u8bcd\u524d\u5148\u7528\u6b63\u5219\u628a\u6807\u8bc6\u7b26\u6574\u4f53\u62a0\u51fa\u6765：[A-Z]{2,}[-_ ]?\\\\d{3,}[A-Z]? \u8fd9\u7c7b\u6a21\u5f0f\u4f18\u5148\u5207，\u4e0d\u518d\u4ea4\u7ed9 jieba\u3002<\/li>\n
  • \u5206\u8bcd\u7ed3\u679c\u91cc\u4e22\u5f03\u7eaf\u6807\u70b9 token\u3002<\/li>\n
  • \u5bf9\u542b\u6807\u8bc6\u7b26\u7684\u67e5\u8be2，\u5728 BM25 \u4fa7\u7ed9\u6807\u8bc6\u7b26\u8bcd\u9879\u52a0\u6743\u91cd（\u4e58 3），\u8ba9\u5b83\u4e3b\u5bfc\u6392\u5e8f\u3002<\/li>\n

    \u4fee\u597d\u4e4b\u540e\u7cbe\u786e\u5339\u914d\u578b\u7684 nDCG@8 \u4ece 0.612 \u63d0\u5230 0.826\u2014\u2014\u8fd9\u4e00\u6761\u5355\u72ec\u8d21\u732e\u4e86 6.2 \u8282\u8868\u683c\u91cc\u6700\u5927\u7684\u90a3\u4e00\u5757\u589e\u76ca\u3002<\/p>\n

    7.3 reranker \u7684\u8f93\u5165\u622a\u65ad\u7b56\u7565\u9519\u8bef，\u957f\u6587\u6863\u7cfb\u7edf\u6027\u88ab\u538b\u4f4e<\/h4>\n

    \u73b0\u8c61：rerank \u4e4b\u540e，\u77ed chunk \u7684\u6392\u540d\u666e\u904d\u4e0a\u5347，\u957f chunk \u666e\u904d\u4e0b\u964d\u3002\u4eba\u5de5\u770b\u6807\u6ce8，\u5f88\u591a\u957f chunk \u660e\u660e\u66f4\u76f8\u5173\u3002<\/p>\n

    \u5b9a\u4f4d：\u7edf\u8ba1 rerank \u524d\u540e\u7684\u5206\u6570\u53d8\u5316，\u6309\u6587\u672c\u957f\u5ea6\u5206\u7ec4，\u53d1\u73b0\u957f\u5ea6\u8d85\u8fc7 400 \u5b57\u7b26\u7684\u7ec4\u5e73\u5747\u4e0b\u964d 4.7 \u4e2a\u540d\u6b21\u3002<\/p>\n

    \u539f\u56e0：\u6211\u628a doc \u622a\u65ad\u5230\u4e86\u524d max_seq_len – len(query) \u4e2a\u5b57\u7b26\u3002\u6280\u672f\u6587\u6863\u7684\u5173\u952e\u7ed3\u8bba\u5e38\u5728\u6bb5\u5c3e\u2014\u2014\u300c\u2026\u2026\u56e0\u6b64\u4e0d\u652f\u6301\u70ed\u8fc1\u79fb\u300d\u300c\u2026\u2026SLA \u4e0d\u9002\u7528\u4e8e\u6b64\u573a\u666f\u300d\u2014\u2014\u622a\u6389\u5c3e\u90e8\u7b49\u4e8e\u7cfb\u7edf\u6027\u5730\u5220\u6389\u4e86\u7ed3\u8bba\u3002<\/p>\n

    \u89e3\u51b3：\u6539\u6210\u4fdd\u7559\u5934 1\/3 + \u5c3e 2\/3（\u89c1 5.3 \u7684 _truncate），\u4e2d\u95f4\u7528 … \u8fde\u63a5\u3002\u6539\u5b8c\u518d\u7edf\u8ba1，\u957f\u5ea6\u76f8\u5173\u7684\u6392\u540d\u504f\u5dee\u4ece -4.7 \u964d\u5230 -0.6\u3002<\/p>\n

    \u901a\u7528\u6559\u8bad：\u622a\u65ad\u7b56\u7565\u662f\u6709\u504f\u7684，\u5fc5\u987b\u7528\u300c\u6309\u7279\u5f81\u5206\u7ec4\u7edf\u8ba1\u6392\u540d\u53d8\u5316\u300d\u7684\u65b9\u5f0f\u9a8c\u8bc1\u504f\u5dee，\u53ea\u770b\u6574\u4f53\u6307\u6807\u770b\u4e0d\u51fa\u6765\u3002<\/p>\n

    7.4 \u89c4\u5219\u5b9e\u4f53\u62bd\u53d6\u7684\u6700\u957f\u77ed\u8bed\u4f18\u5148\u6ca1\u505a\u5bf9，\u77ed\u8bcd\u5148\u547d\u4e2d<\/h4>\n

    \u73b0\u8c61：\u6587\u672c\u300c\u5ba2\u6237\u4e3b\u6570\u636e\u7ba1\u7406\u7cfb\u7edf\u300d\u88ab\u62bd\u6210\u4e86\u4e09\u4e2a\u5b9e\u4f53：\u5ba2\u6237\u3001\u4e3b\u6570\u636e\u3001\u7cfb\u7edf\u3002\u56fe\u8c31\u91cc\u51ed\u7a7a\u591a\u51fa\u4e00\u6279\u5b64\u7acb\u8282\u70b9\u3002<\/p>\n

    \u5b9a\u4f4d：\u521d\u7248\u5b9e\u73b0\u662f\u6309\u8bcd\u5178\u987a\u5e8f\u904d\u5386，\u547d\u4e2d\u5c31\u8bb0\u5f55\u3002\u8bcd\u5178\u662f\u6309 id \u6392\u7684，\u4e0d\u662f\u6309\u957f\u5ea6\u6392\u7684\u3002<\/p>\n

    \u539f\u56e0：\u6ca1\u6709\u5b9e\u73b0\u300c\u6700\u957f\u77ed\u8bed\u4f18\u5148\u300d\u3002\u4efb\u4f55\u591a\u6a21\u5f0f\u5339\u914d\u90fd\u5fc5\u987b\u5148\u6309\u6a21\u5f0f\u957f\u5ea6\u964d\u5e8f，\u5426\u5219\u77ed\u6a21\u5f0f\u4f1a\u5148\u5403\u6389\u5b57\u7b26\u533a\u95f4\u3002<\/p>\n

    \u89e3\u51b3：\u4e24\u5904\u6539\u52a8\u3002\u4e00\u662f self.surfaces_sorted = sorted(…, key=len, reverse=True)；\u4e8c\u662f\u5f15\u5165 occupied \u533a\u95f4\u5217\u8868，\u547d\u4e2d\u5373\u72ec\u5360，\u540e\u7eed\u5019\u9009\u82e5\u533a\u95f4\u91cd\u53e0\u5c31\u8df3\u8fc7（\u89c1 5.4 \u7684 _pass1）\u3002<\/p>\n

    \u6539\u5b8c\u4e4b\u540e\u300c\u5ba2\u6237\u4e3b\u6570\u636e\u7ba1\u7406\u7cfb\u7edf\u300d\u6b63\u786e\u5730\u62bd\u6210\u4e00\u6761 E_CUST_MSTR，\u56fe\u8c31\u91cc\u7684\u5b64\u7acb\u8282\u70b9\u6570\u4ece 1.2 \u4e07\u964d\u5230 900 \u5de6\u53f3\u2014\u2014\u90a3\u4e9b\u5b64\u7acb\u8282\u70b9\u5927\u90e8\u5206\u662f\u9519\u8bef\u5207\u5206\u4ea7\u751f\u7684\u788e\u7247\u3002<\/p>\n

    7.5 \u7f6e\u4fe1\u5ea6\u7528\u5e73\u5747，\u4e00\u4e2a\u5d29\u6389\u7684\u73af\u8282\u88ab\u63a9\u76d6\u6210\u300c\u8fd8\u884c\u300d<\/h4>\n

    \u73b0\u8c61：\u7ebf\u4e0a\u6709\u4e00\u6279\u7b54\u6848，\u5f15\u7528\u770b\u7740\u5f88\u5168，\u4f46\u7ed3\u8bba\u662f\u9519\u7684\u3002\u67e5\u65e5\u5fd7，\u68c0\u7d22\u9636\u6bb5\u7684 grounded \u5ea6\u53ea\u6709 0.12（\u57fa\u672c\u6ca1\u53ec\u56de\u5bf9），\u4f46\u6574\u4f53\u7f6e\u4fe1\u5ea6\u662f 0.61，\u9ad8\u4e8e\u9608\u503c 0.6，\u4e8e\u662f\u7cfb\u7edf\u7167\u5e38\u56de\u7b54\u4e86\u3002<\/p>\n

    \u5b9a\u4f4d：confidence = (retrieval + generation) \/ 2 = (0.12 + 1.0) \/ 2 = 0.56\u2026\u2026\u4e0d\u5bf9，\u5b9e\u9645\u662f\u751f\u6210\u9636\u6bb5\u7ed9\u4e86 1.0，\u68c0\u7d22\u7ed9\u4e86 0.22，\u5e73\u5747\u51fa\u6765 0.61\u3002<\/p>\n

    \u539f\u56e0：\u5e73\u5747\u628a\u300c\u68c0\u7d22\u5d29\u4e86\u300d\u8fd9\u4e2a\u81f4\u547d\u95ee\u9898\u7a00\u91ca\u6210\u4e86\u4e00\u4e2a\u4e2d\u7b49\u5206\u6570\u3002\u6a21\u578b\u5728\u51e0\u4e4e\u6ca1\u6709\u4efb\u4f55\u6b63\u786e\u4e0a\u4e0b\u6587\u7684\u60c5\u51b5\u4e0b\u6d41\u7545\u5730\u751f\u6210\u4e86\u4e00\u6bb5\u8bdd，\u751f\u6210\u4fa7\u7684\u81ea\u8bc4\u8fd8\u7ed9\u4e86\u9ad8\u5206\u2014\u2014\u8fd9\u6b63\u662f\u81ea\u8bc4\u4e0d\u53ef\u9760\u7684\u5730\u65b9\u3002<\/p>\n

    \u89e3\u51b3：\u6539\u6210\u4e58\u6cd5（3.8 \u8282），\u5e76\u52a0\u300c\u4efb\u4e00\u56e0\u5b50\u4f4e\u4e8e 0.30 \u4e00\u7968\u5426\u51b3\u300d（5.6 \u7684 blocking()）\u3002\u540c\u6837\u7684\u573a\u666f，\u4e58\u6cd5\u7ed9\u51fa 0.22，\u4e14\u89e6\u53d1\u4e00\u7968\u5426\u51b3，\u76f4\u63a5\u62d2\u7b54\u30026.5 \u8282\u7684\u6570\u636e\u663e\u793a，\u8fd9\u4e00\u6539\u628a\u9519\u8bef\u7b54\u6848\u6f0f\u51fa\u7387\u4ece 14.1% \u538b\u5230 4.4%\u3002<\/p>\n

    7.6 \u6709\u754c\u91cd\u8bd5\u6ca1\u6709\u53bb\u91cd，\u540c\u4e00\u4e2a query \u7528\u540c\u4e00\u4e2a\u7b56\u7565\u53cd\u590d\u91cd\u8bd5<\/h4>\n

    \u73b0\u8c61：\u65e5\u5fd7\u91cc\u540c\u4e00\u4e2a\u4efb\u52a1\u51fa\u73b0\u4e86 3 \u6b21\u5b8c\u5168\u76f8\u540c\u7684\u68c0\u7d22\u8bf7\u6c42，\u8fd4\u56de\u5b8c\u5168\u76f8\u540c\u7684\u7ed3\u679c，\u7136\u540e 3 \u6b21\u90fd\u5224\u5b9a\u4f4e\u7f6e\u4fe1\u5ea6，\u6700\u540e\u62d2\u7b54\u3002\u767d\u82b1\u4e86 3 \u500d\u5ef6\u8fdf\u3002<\/p>\n

    \u5b9a\u4f4d：RetryState \u53ea\u8bb0\u4e86 attempts，\u6ca1\u6709\u8bb0\u5f55\u300c\u54ea\u6b21\u7b56\u7565\u914d\u54ea\u4e2a query \u5df2\u7ecf\u8bd5\u8fc7\u300d\u3002<\/p>\n

    \u539f\u56e0：\u7b56\u7565\u5e8f\u5217\u7684\u63a8\u8fdb\u903b\u8f91\u5199\u5728\u4e86\u8c03\u7528\u65b9，\u800c\u8c03\u7528\u65b9\u5728\u91cd\u8bd5\u65f6\u6ca1\u6709\u771f\u7684\u6539\u53d8 query\u2014\u2014_rewrite \u5728 strategy_idx=0 \u65f6\u8fd4\u56de\u539f query，\u4e8e\u662f\u300c\u6362\u7b56\u7565\u300d\u5b9e\u9645\u4e0a\u4ec0\u4e48\u90fd\u6ca1\u6362\u3002<\/p>\n

    \u89e3\u51b3：\u5728 decide() \u91cc\u52a0 seen_queries \u96c6\u5408，key \u662f strategy_idx + query；\u547d\u4e2d\u76f8\u540c key \u5c31\u5f3a\u5236\u63a8\u8fdb strategy_idx（\u89c1 5.6）\u3002\u540c\u65f6\u628a\u300c\u7b56\u7565\u5fc5\u987b\u6539\u53d8\u8f93\u5165\u300d\u5199\u6210\u4e00\u6761\u7ea6\u5b9a：\u4efb\u4f55\u7b56\u7565\u5982\u679c\u4e0d\u80fd\u6539\u53d8\u68c0\u7d22\u8f93\u5165，\u5b83\u5c31\u4e0d\u5e94\u8be5\u51fa\u73b0\u5728\u7b56\u7565\u5e8f\u5217\u91cc\u3002<\/p>\n

    7.7 \u5ef6\u8fdf\u9884\u7b97\u53ea\u7b97\u68c0\u7d22\u4e0d\u7b97 rerank，P99 \u76f4\u63a5\u7206\u6389<\/h4>\n

    \u73b0\u8c61：\u68c0\u7d22\u4fa7 P99 \u662f 240ms，\u7b26\u5408\u9884\u671f，\u4f46\u7aef\u5230\u7aef P99 \u5230\u4e86 6.8 \u79d2\u3002<\/p>\n

    \u5b9a\u4f4d：\u628a\u6574\u6761\u94fe\u8def\u6309\u7ec4\u4ef6\u6253\u70b9，rerank \u90a3\u4e00\u6bb5\u7684 P99 \u662f 4.9 \u79d2\u3002<\/p>\n

    \u539f\u56e0：\u6211\u53ea\u7ed9\u68c0\u7d22\u5206\u914d\u4e86\u9884\u7b97，rerank \u7528\u7684\u662f model.predict(pairs) \u7684\u540c\u6b65\u8c03\u7528，\u6ca1\u6709\u8d85\u65f6\u4e5f\u6ca1\u6709\u6279\u5904\u7406\u9650\u5236\u3002\u5f53 top-50 \u5168\u662f\u957f\u6587\u672c\u65f6，50 \u4e2a 512-token \u7684\u4ea4\u53c9\u7f16\u7801\u5728 CPU \u4e0a\u8981\u8dd1\u597d\u51e0\u79d2\u3002<\/p>\n

    \u89e3\u51b3：\u4e09\u6761\u3002\u4e00\u662f\u7ed9 rerank \u5206\u914d 400ms \u9884\u7b97\u5e76\u914d\u964d\u7ea7\u8def\u5f84（5.3\u30015.7）；\u4e8c\u662f\u6309 max_seq_len \u52a8\u6001\u7b97\u6279\u5927\u5c0f，\u957f\u6587\u672c\u6279\u6b21\u8c03\u5c0f；\u4e09\u662f\u628a P99 \u800c\u4e0d\u662f P50 \u5199\u8fdb\u9884\u7b97\u8868\u7684\u9a8c\u6536\u6807\u51c6\u2014\u2014\u5ef6\u8fdf\u9884\u7b97\u7684\u4ef7\u503c\u5728\u5c3e\u90e8，\u7528 P50 \u9a8c\u6536\u7b49\u4e8e\u6ca1\u8bbe\u3002<\/p>\n

    7.8 \u56fe\u6269\u5c55\u628a\u6574\u4e2a\u5b50\u56fe\u90fd\u62c9\u56de\u6765\u4e86<\/h4>\n

    \u73b0\u8c61：\u542f\u7528 GraphRAG \u4e4b\u540e\u5ef6\u8fdf\u4ece 512ms \u6da8\u5230 4.3 \u79d2，\u800c\u4e14\u53ec\u56de\u7684 chunk \u91cc\u6df7\u8fdb\u4e86\u4e00\u5927\u6279\u548c\u67e5\u8be2\u65e0\u5173\u7684\u901a\u7528\u8bf4\u660e\u3002<\/p>\n

    \u5b9a\u4f4d：\u6253\u5370\u56fe\u6269\u5c55\u7684\u8282\u70b9\u6570，\u4e00\u6b21\u67e5\u8be2\u5e73\u5747\u6269\u5c55\u51fa 340 \u4e2a\u5b9e\u4f53\u3002<\/p>\n

    \u539f\u56e0：\u6211\u505a\u7684\u662f\u65e0\u9650\u5236 BFS（\u5e7f\u5ea6\u4f18\u5148\u641c\u7d22），\u6cbf\u6240\u6709\u5173\u7cfb\u7c7b\u578b\u8d70 2 \u8df3\u3002\u4ea7\u54c1\u5c42\u7ea7\u91cc\u300c\u4f01\u4e1a\u7ea7\u4ea7\u54c1\u7ebf\u300d\u8fd9\u79cd hub \u8282\u70b9\u8fde\u7740\u51e0\u767e\u4e2a\u4ea7\u54c1，\u4e00\u8df3\u5c31\u7206\u70b8\u4e86\u3002<\/p>\n

    \u89e3\u51b3：\u56db\u6761\u7ea6\u675f\u4e00\u8d77\u4e0a\u3002<\/p>\n

  • \u8fb9\u7c7b\u578b\u767d\u540d\u5355：\u53ea\u6cbf part_of \u548c synonym_of \u8d70，\u4e0d\u8d70 mentions（\u540e\u8005\u5ea6\u6570\u592a\u9ad8）\u3002<\/li>\n
  • \u5ea6\u6570\u4e0a\u9650：\u8df3\u8fc7\u5ea6\u6570\u5927\u4e8e 50 \u7684 hub \u8282\u70b9，\u6216\u8005\u5bf9\u5176\u505a\u964d\u91c7\u6837\u3002<\/li>\n
  • \u6bcf\u8df3\u7684\u5b9e\u4f53\u6570\u4e0a\u9650（\u6bd4\u5982\u6bcf\u8df3\u6700\u591a\u6269 12 \u4e2a），\u4e14\u6309\u4e0e\u67e5\u8be2\u7684\u76f8\u5173\u5ea6\u6392\u5e8f\u540e\u622a\u65ad\u3002<\/li>\n
  • \u603b\u5ef6\u8fdf\u9884\u7b97 300ms，\u8d85\u4e86\u5c31\u9000\u56de\u300c\u4e0d\u6269\u5c55\u300d\u7684\u7ed3\u679c\u3002<\/li>\n

    \u6539\u5b8c\u4e4b\u540e\u5e73\u5747\u6269\u5c55\u8282\u70b9\u4ece 340 \u964d\u5230 19，\u5ef6\u8fdf\u56de\u5230 512ms，\u800c\u591a\u8df3\u578b nDCG@8 \u53cd\u800c\u4ece 0.517 \u63d0\u5230 0.541\u2014\u2014\u56e0\u4e3a\u4e4b\u524d\u90a3\u4e9b\u65e0\u5173 chunk \u8fdb\u4e0a\u4e0b\u6587\u4e4b\u540e，\u53cd\u800c\u7a00\u91ca\u4e86\u91cd\u6392\u7684\u4fe1\u53f7\u3002<\/p>\n

    7.9 \u8e29\u5751\u901f\u67e5\u8868<\/h4>\n\n\u73b0\u8c61\u4e00\u53e5\u8bdd\u5b9a\u4f4d\u65b9\u6cd5\u6839\u56e0\u5206\u7c7b\u4fee\u590d\u4f4d\u7f6e<\/tr>\n\n\n\n\n\n\n\n\n\n
    \u7b54\u6848\u91cd\u590d\u540c\u4e00\u53e5\u8bdd<\/td>\n\u6253\u5370 top-8 \u7684\u5f52\u4e00\u5316 hash<\/td>\n\u7f3a\u5185\u5bb9\u6307\u7eb9\u5c42\u53bb\u91cd<\/td>\n\u4e09\u5c42\u53bb\u91cd L3<\/td>\n<\/tr>\n
    \u4ea7\u54c1\u7801\u68c0\u7d22\u4e0d\u51c6<\/td>\n\u6253\u5370\u5206\u8bcd\u7ed3\u679c<\/td>\n\u6807\u8bc6\u7b26\u88ab\u5207\u788e<\/td>\n\u6807\u8bc6\u7b26\u4f18\u5148\u5207\u5206 + \u52a0\u6743<\/td>\n<\/tr>\n
    \u957f chunk \u6392\u540d\u504f\u4f4e<\/td>\n\u6309\u957f\u5ea6\u5206\u7ec4\u7edf\u8ba1\u6392\u540d\u53d8\u5316<\/td>\n\u622a\u65ad\u7b56\u7565\u6709\u504f<\/td>\n\u4fdd\u5934 + \u4fdd\u5c3e<\/td>\n<\/tr>\n
    \u56fe\u8c31\u91cc\u4e00\u5806\u5b64\u7acb\u8282\u70b9<\/td>\n\u7edf\u8ba1\u8282\u70b9\u5ea6\u6570\u5206\u5e03<\/td>\n\u77ed\u8bcd\u5148\u547d\u4e2d<\/td>\n\u6700\u957f\u77ed\u8bed\u4f18\u5148 + \u533a\u95f4\u72ec\u5360<\/td>\n<\/tr>\n
    \u5f15\u7528\u5168\u4f46\u7ed3\u8bba\u9519<\/td>\n\u770b\u5404\u56e0\u5b50\u5206\u522b\u591a\u5c11<\/td>\n\u7528\u5e73\u5747\u63a9\u76d6\u5f31\u73af\u8282<\/td>\n\u6539\u4e58\u6cd5 + \u4e00\u7968\u5426\u51b3<\/td>\n<\/tr>\n
    \u91cd\u8bd5\u62ff\u5230\u4e00\u6837\u7684\u7ed3\u679c<\/td>\n\u5bf9\u6bd4\u5404\u6b21\u68c0\u7d22\u8bf7\u6c42<\/td>\n\u7b56\u7565\u6ca1\u6539\u53d8\u8f93\u5165<\/td>\nseen_queries \u53bb\u91cd<\/td>\n<\/tr>\n
    \u7aef\u5230\u7aef P99 \u7206\u6389<\/td>\n\u6309\u7ec4\u4ef6\u6253\u70b9\u770b P99<\/td>\nrerank \u65e0\u9884\u7b97<\/td>\n\u7ec4\u4ef6\u7ea7\u9884\u7b97 + \u964d\u7ea7<\/td>\n<\/tr>\n
    \u56fe\u6269\u5c55\u540e\u5ef6\u8fdf\u98d9\u5347<\/td>\n\u6253\u5370\u6269\u5c55\u8282\u70b9\u6570<\/td>\nhub \u8282\u70b9\u7206\u70b8<\/td>\n\u8fb9\u7c7b\u578b\u767d\u540d\u5355 + \u5ea6\u6570\u4e0a\u9650<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u516b\u3001\u4f18\u5316\u65b9\u6848\u4e0e\u8fdb\u9636\u6253\u6cd5<\/h3>\n

    8.1 \u53cd\u6a21\u5f0f\u4e00：\u5411\u91cf\u5e93\u88ab AI \u751f\u6210\u7684\u6458\u8981\u6c61\u67d3<\/h4>\n

    \u8fd9\u662f\u6211\u6700\u60f3\u8b66\u544a\u7684\u4e00\u6761\u3002\u5f88\u591a\u56e2\u961f\u7684 ingestion（\u5165\u5e93）\u6d41\u7a0b\u662f\u8fd9\u6837\u7684：\u539f\u59cb\u6587\u6863\u592a\u957f -> \u8ba9 LLM \u751f\u6210\u4e00\u6bb5\u6458\u8981 -> \u628a\u6458\u8981\u548c\u539f\u6587\u4e00\u8d77\u5b58\u8fdb\u5411\u91cf\u5e93\u3002\u770b\u8d77\u6765\u5408\u7406，\u5b9e\u9645\u662f\u5728\u7ed9\u81ea\u5df1\u57cb\u96f7\u3002<\/p>\n

    \u95ee\u9898\u5728\u4e8e\u6458\u8981\u548c\u539f\u6587\u5728\u5411\u91cf\u7a7a\u95f4\u91cc\u662f\u7ade\u4e89\u5173\u7cfb\u3002\u7528\u6237\u95ee\u300cAX-2039 \u7684 SLA \u662f\u591a\u5c11\u300d，\u6458\u8981\u91cc\u53ef\u80fd\u5199\u4e86\u300cSLA \u76f8\u5173\u5185\u5bb9\u89c1\u539f\u6587\u300d，\u539f\u6587\u91cc\u6709\u786e\u5207\u7684 99.9%\u3002\u4e24\u8005\u90fd\u548c\u67e5\u8be2\u8bed\u4e49\u76f8\u8fd1，\u6458\u8981\u56e0\u4e3a\u66f4\u77ed\u66f4\u51dd\u7ec3，\u4f59\u5f26\u76f8\u4f3c\u5ea6\u5f80\u5f80\u66f4\u9ad8，\u4e8e\u662f\u6458\u8981\u6392\u5728\u539f\u6587\u524d\u9762，\u6a21\u578b\u62ff\u5230\u7684\u662f\u4e00\u4e2a\u4e0d\u542b\u7b54\u6848\u7684\u9ad8\u5206 chunk\u3002<\/p>\n

    \u66f4\u7cdf\u7684\u662f\u7b2c\u4e8c\u8f6e：\u6709\u4eba\u53d1\u73b0\u8fd9\u4e2a\u95ee\u9898\u4e4b\u540e，\u53c8\u8ba9 LLM \u5bf9\u6458\u8981\u518d\u505a\u4e00\u6b21\u6458\u8981，\u4e8e\u662f\u5e93\u91cc\u51fa\u73b0\u4e86\u300c\u6458\u8981\u7684\u6458\u8981\u300d\u3002\u4e09\u4ee3\u6458\u8981\u53e0\u5728\u4e00\u8d77，\u6bcf\u4e00\u4ee3\u90fd\u79bb\u539f\u6587\u8fdc\u4e00\u70b9，\u6700\u540e\u4f60\u68c0\u7d22\u5230\u7684\u5168\u90fd\u662f\u300c\u5173\u4e8e\u5173\u4e8e\u5173\u4e8e SLA \u7684\u8bf4\u660e\u300d\u3002<\/p>\n

    \u89e3\u51b3\u529e\u6cd5\u662f\u5728 ingestion \u65f6\u52a0\u4e00\u4e2a\u5206\u7c7b\u5c42，\u7ed9\u6bcf\u4e2a chunk \u6253\u4e0a content_type \u6807\u7b7e\u5e76\u5728\u68c0\u7d22\u65f6\u9694\u79bb：<\/p>\n

    \u5165\u5e93\u5206\u7c7b\u5c42（\u786e\u5b9a\u6027\u89c4\u5219\u4f18\u5148，LLM \u53ea\u505a\u515c\u5e95）<\/p>\n

    canonical \u6743\u5a01\u539f\u6587：\u4ea7\u54c1\u624b\u518c\u6b63\u6587\u3001Runbook \u6b65\u9aa4\u3001API \u6587\u6863
    \n -> \u53ef\u68c0\u7d22，\u6743\u91cd 1.0
    \n derived AI \u751f\u6210\u7684\u6458\u8981\u3001\u6539\u5199\u3001\u7ffb\u8bd1
    \n -> \u53ef\u68c0\u7d22，\u6743\u91cd 0.6，\u4e14\u5fc5\u987b\u5e26 source_chunk_id \u6307\u56de canonical
    \n index \u76ee\u5f55\u3001\u7d22\u5f15\u9875\u3001\u5bfc\u822a\u9875\u3001\u53d8\u66f4\u65e5\u5fd7
    \n -> \u53ef\u68c0\u7d22，\u6743\u91cd 0.4
    \n noise \u9875\u7709\u9875\u811a\u3001\u514d\u8d23\u58f0\u660e\u6a21\u677f\u3001\u91cd\u590d\u7684\u8bb8\u53ef\u8bc1\u6587\u672c
    \n -> \u4e0d\u5165\u5e93
    \n sensitive \u542b PII \/ \u5bc6\u7ea7\u6807\u8bb0
    \n -> \u5165\u5e93\u4f46\u5f3a\u5236 ACL，\u4e14\u9ed8\u8ba4\u4e0d\u53c2\u4e0e\u68c0\u7d22<\/p>\n

    \u4e09\u6761\u786c\u89c4\u5219：<\/p>\n

  • **derived \u5fc5\u987b\u80fd\u6307\u56de canonical\u3002**\u6ca1\u6709 source_chunk_id \u7684\u6458\u8981\u4e0d\u5141\u8bb8\u5165\u5e93\u3002\u8fd9\u6837\u5373\u4f7f\u6458\u8981\u88ab\u53ec\u56de\u4e86，\u4f60\u4e5f\u80fd\u7acb\u523b\u987a\u7740\u6307\u9488\u62ff\u5230\u539f\u6587\u3002<\/li>\n
  • **\u540c\u4e00 source_chunk_id \u7684 canonical \u548c derived \u4e0d\u80fd\u540c\u65f6\u8fdb top-8\u3002**\u4e8c\u9009\u4e00，\u4f18\u5148 canonical\u3002\u8fd9\u76f4\u63a5\u6d88\u706d\u4e86\u300c\u6458\u8981\u6324\u6389\u539f\u6587\u300d\u7684\u95ee\u9898\u3002<\/li>\n
  • **derived \u4e0d\u53c2\u4e0e\u56fe\u6784\u5efa\u3002**\u4e0d\u8981\u8ba9 AI \u751f\u6210\u7684\u6587\u672c\u4ea7\u751f\u5b9e\u4f53\u548c\u8fb9\u2014\u2014\u90a3\u4f1a\u8ba9\u56fe\u8c31\u7684\u53ef\u4fe1\u5ea6\u6574\u4f53\u964d\u7ea7\u3002<\/li>\n

    8.2 \u53cd\u6a21\u5f0f\u4e8c：\u65e0\u9650\u585e\u4e0a\u4e0b\u6587<\/h4>\n

    \u300c\u4e0a\u4e0b\u6587\u7a97\u53e3\u6709 200 \u4e07 token \u4e86，\u628a\u6574\u4e2a\u5e93\u585e\u8fdb\u53bb\u4e0d\u5c31\u884c\u4e86\u300d\u2014\u2014\u8fd9\u4e2a\u60f3\u6cd5\u5728 2026 \u5e74\u4f9d\u7136\u6709\u4eba\u63d0，\u4e5f\u4f9d\u7136\u662f\u9519\u7684\u3002\u4e09\u4e2a\u7406\u7531：<\/p>\n

    **\u7b2c\u4e00，\u6210\u672c\u662f\u7ebf\u6027\u7684\u800c\u6536\u76ca\u662f\u6b21\u7ebf\u6027\u7684\u3002**\u585e 8 \u6761 chunk \u548c\u585e 80 \u6761，\u6210\u672c\u5dee 10 \u500d，\u800c\u7b54\u6848\u8d28\u91cf\u901a\u5e38\u5df2\u7ecf\u9971\u548c\u4e86\u30026.3 \u8282\u7684\u6570\u636e\u91cc，top-8 \u4e4b\u540e\u518d\u589e\u52a0 chunk \u5bf9 nDCG \u51e0\u4e4e\u6ca1\u6709\u8d21\u732e\u3002<\/p>\n

    **\u7b2c\u4e8c，\u4e2d\u95f4\u4e22\u5931（lost-in-the-middle）\u6548\u5e94\u3002**\u957f\u4e0a\u4e0b\u6587\u91cc，\u6a21\u578b\u5bf9\u5f00\u5934\u548c\u7ed3\u5c3e\u7684\u5185\u5bb9\u6ce8\u610f\u529b\u9ad8，\u4e2d\u95f4\u7684\u5185\u5bb9\u5bb9\u6613\u88ab\u5ffd\u7565\u3002\u4f60\u628a 80 \u6761 chunk \u585e\u8fdb\u53bb，\u771f\u6b63\u88ab\u7528\u5230\u7684\u53ef\u80fd\u5c31\u662f\u524d 5 \u6761\u548c\u540e 5 \u6761，\u4e2d\u95f4 70 \u6761\u662f\u7eaf\u6210\u672c\u3002\u800c\u4e14\u4f4d\u7f6e\u662f\u968f\u673a\u7684，\u4f60\u4f1a\u5f97\u5230\u4e0d\u7a33\u5b9a\u7684\u7b54\u6848\u3002<\/p>\n

    **\u7b2c\u4e09，\u4e5f\u662f\u6700\u5b9e\u9645\u7684：\u585e\u5f97\u591a\u4f1a\u8ba9 grounded \u5224\u5b9a\u5931\u6548\u3002**\u5f15\u6587\u5f52\u5c5e\u7684\u524d\u63d0\u662f\u300c\u6a21\u578b\u8bf4\u7684\u6bcf\u53e5\u8bdd\u90fd\u80fd\u5bf9\u5e94\u5230\u67d0\u4e00\u6761\u5177\u4f53\u5f15\u7528\u300d\u3002\u4e0a\u4e0b\u6587\u8d8a\u957f，\u6a21\u578b\u8d8a\u503e\u5411\u4e8e\u505a\u8de8\u5f15\u7528\u7684\u7efc\u5408，\u4f60\u5c31\u8d8a\u96be\u5224\u65ad\u67d0\u53e5\u8bdd\u5230\u5e95\u51fa\u81ea\u54ea\u91cc\u3002<\/p>\n

    \u6b63\u786e\u7684\u505a\u6cd5\u662f\u6309\u9700\u6269\u5c55\u800c\u4e0d\u662f\u9884\u8bbe\u5f88\u5927：\u4ece top-8 \u5f00\u59cb，\u53ea\u6709\u5f53\u7f6e\u4fe1\u5ea6\u4f4e\u4e8e\u9608\u503c\u3001\u4e14\u5224\u5b9a\u539f\u56e0\u662f\u300c\u4fe1\u606f\u4e0d\u8db3\u300d\u800c\u4e0d\u662f\u300c\u68c0\u7d22\u9519\u4e86\u300d\u65f6，\u624d\u6269\u5c55\u5230 top-16，\u518d\u4e0d\u884c\u624d top-24\u3002\u8fd9\u662f\u4e00\u4e2a\u6709\u754c\u7684\u8fc7\u7a0b，\u4e0d\u662f\u4e00\u4e2a\u5e38\u6570\u3002<\/p>\n

    8.3 \u53cd\u6a21\u5f0f\u4e09：AI \u540c\u65f6\u8bfb\u5199\u540c\u4e00\u4e2a\u5e93<\/h4>\n

    \u8ba9 Agent \u4e00\u8fb9\u4ece\u5411\u91cf\u5e93\u68c0\u7d22\u3001\u4e00\u8fb9\u628a\u751f\u6210\u7684\u5185\u5bb9\u5199\u56de\u540c\u4e00\u4e2a\u5e93，\u662f\u6700\u5bb9\u6613\u5728\u51e0\u4e2a\u6708\u540e\u7206\u53d1\u7684\u95ee\u9898\u3002<\/p>\n

    \u77ed\u671f\u8868\u73b0\u5f88\u6f02\u4eae：Agent \u6bcf\u6b21\u56de\u7b54\u90fd\u300c\u5b66\u5230\u4e86\u65b0\u4e1c\u897f\u300d，\u5e93\u8d8a\u6765\u8d8a\u5927\u3002<\/p>\n

    \u957f\u671f\u540e\u679c\u662f\u707e\u96be\u6027\u7684\u3002\u7b2c\u4e00\u6b65，Agent \u751f\u6210\u7684\u5185\u5bb9\u5165\u5e93，\u6210\u4e3a\u540e\u7eed\u68c0\u7d22\u7684\u5019\u9009\u3002\u7b2c\u4e8c\u6b65，\u540e\u7eed\u68c0\u7d22\u5230\u8fd9\u4e9b\u5185\u5bb9，\u628a\u5b83\u4eec\u5f53\u6210\u4e8b\u5b9e\u4f9d\u636e\u3002\u7b2c\u4e09\u6b65，\u6a21\u578b\u57fa\u4e8e\u300c\u81ea\u5df1\u7684\u5386\u53f2\u8f93\u51fa\u300d\u751f\u6210\u65b0\u7b54\u6848，\u518d\u5165\u5e93\u3002\u8fd9\u662f\u4e00\u4e2a\u81ea\u6211\u5f3a\u5316\u7684\u56de\u8def\u2014\u2014**\u51e0\u8f6e\u4e4b\u540e，\u5e93\u91cc\u7684\u4e3b\u6d41\u5185\u5bb9\u5c31\u5168\u90fd\u662f AI \u751f\u6210\u7269，\u800c\u5b83\u4eec\u7684\u6e90\u5934\u53ef\u80fd\u53ea\u662f\u6700\u521d\u67d0\u4e00\u6b21\u6a21\u578b\u7684\u53e3\u8bef\u3002**\u8fd9\u7c7b\u6c61\u67d3\u6781\u96be\u53d1\u73b0，\u56e0\u4e3a\u6240\u6709\u5185\u5bb9\u770b\u8d77\u6765\u90fd\u5408\u7406，\u800c\u4e14\u4e92\u76f8\u5370\u8bc1\u3002<\/p>\n

    \u89e3\u51b3\u529e\u6cd5\u662f\u7269\u7406\u9694\u79bb：\u8bfb\u5e93\u548c\u5199\u5e93\u5206\u5f00，\u5199\u5e93\u5230\u8bfb\u5e93\u4e4b\u95f4\u6709\u4e00\u9053\u4eba\u5de5\u6216\u89c4\u5219\u7684\u95f8\u95e8\u3002<\/p>\n

    [\u6743\u5a01\u6e90] –> [ingestion] –> [\u8bfb\u5e93 \/ \u68c0\u7d22\u5e93] <– \u53ea\u5141\u8bb8\u4ece\u6743\u5a01\u6e90\u6d41\u5165
    \n |
    \n v
    \n [\u68c0\u7d22 + \u751f\u6210]
    \n |
    \n v
    \n [\u5199\u5e93 \/ \u5019\u9009\u6c60] –> [\u5ba1\u6838\u95f8\u95e8] –> [\u8bfb\u5e93]
    \n |
    \n （\u4eba\u5de5 or \u89c4\u5219\u6821\u9a8c\u901a\u8fc7\u624d\u653e\u884c）<\/p>\n

    \u5199\u5e93\u4e5f\u53ef\u4ee5\u53eb\u300c\u5019\u9009\u6c60\u300d\u6216\u300c\u8349\u7a3f\u533a\u300d，\u5b83\u7684\u5185\u5bb9\u9ed8\u8ba4\u4e0d\u53c2\u4e0e\u68c0\u7d22\u3002\u53ea\u6709\u901a\u8fc7\u5ba1\u6838（\u4eba\u5de5\u786e\u8ba4，\u6216\u8005\u81f3\u5c11\u4e00\u6761\u786e\u5b9a\u6027\u89c4\u5219\u6821\u9a8c\u901a\u8fc7）\u4e4b\u540e\u624d\u63d0\u5347\u8fdb\u8bfb\u5e93\u3002<\/p>\n

    \u5982\u679c\u4e1a\u52a1\u4e0a\u786e\u5b9e\u9700\u8981 Agent \u7684\u4ea7\u51fa\u7acb\u523b\u53ef\u68c0\u7d22（\u6bd4\u5982\u5ba2\u670d\u573a\u666f\u7684\u4f1a\u8bdd\u8bb0\u5fc6），\u90a3\u5c31\u7ed9\u5b83\u4e00\u4e2a\u72ec\u7acb\u7684\u547d\u540d\u7a7a\u95f4 + \u72ec\u7acb\u7684\u5185\u5bb9\u7c7b\u578b\u6807\u7b7e，\u5e76\u4e14\u8bbe\u7f6e TTL，\u8ba9\u5b83\u81ea\u52a8\u8fc7\u671f\u3002\u4e0d\u8981\u8ba9 AI \u751f\u6210\u7269\u548c\u6743\u5a01\u6e90\u5728\u540c\u4e00\u4e2a\u5411\u91cf\u7a7a\u95f4\u91cc\u65e0\u6807\u8bb0\u5730\u6df7\u5408\u3002<\/p>\n

    8.4 \u67e5\u8be2\u6539\u5199\u8981\u4e0d\u8981\u505a<\/h4>\n

    \u505a\u4e86\u5b9e\u6d4b：\u67e5\u8be2\u6539\u5199（\u628a\u7528\u6237\u7684\u53e3\u8bed\u5316\u63d0\u95ee\u6539\u5199\u6210\u68c0\u7d22\u53cb\u597d\u7684\u5f62\u5f0f）\u5728\u5355\u8df3\u4e8b\u5b9e\u578b\u4e0a\u63d0\u5347\u6709\u9650（nDCG \u4ece 0.741 \u5230 0.752），\u4f46\u5728\u591a\u8df3\u578b\u4e0a\u63d0\u5347\u660e\u663e（0.517 \u5230 0.594）\u3002<\/p>\n

    \u6211\u7684\u5efa\u8bae\u662f\u6309\u9700\u89e6\u53d1\u800c\u4e0d\u662f\u9ed8\u8ba4\u5f00\u542f：\u53ea\u6709\u5f53\u7f6e\u4fe1\u5ea6\u4f4e\u4e8e\u9608\u503c\u4e14\u5224\u5b9a\u539f\u56e0\u662f\u300c\u68c0\u7d22\u8d28\u91cf\u5dee\u300d\u65f6\u624d\u505a\u6539\u5199，\u800c\u4e0d\u662f\u6bcf\u6b21\u67e5\u8be2\u90fd\u5148\u82b1 250ms \u6539\u5199\u4e00\u904d\u3002\u8fd9\u6837\u5728\u5355\u8df3\u67e5\u8be2\u5360\u591a\u6570（\u6211\u8fd9\u91cc\u662f 53%）\u7684\u573a\u666f\u91cc，\u80fd\u7701\u6389\u4e00\u5927\u7b14\u56fa\u5b9a\u5ef6\u8fdf\u3002<\/p>\n

    8.5 \u7f13\u5b58：\u54ea\u4e9b\u80fd\u7f13\u5b58，\u54ea\u4e9b\u4e0d\u80fd<\/h4>\n\n\u73af\u8282\u80fd\u5426\u7f13\u5b58\u7f13\u5b58\u952e\u6536\u76ca<\/tr>\n\n\n\n\n\n\n\n
    embedding \u7f16\u7801<\/td>\n\u80fd<\/td>\nquery \u6587\u672c\u7684 hash<\/td>\n\u9ad8\u9891\u67e5\u8be2\u660e\u663e，\u547d\u4e2d\u7387\u7ea6 30%<\/td>\n<\/tr>\n
    BM25 \u68c0\u7d22<\/td>\n\u80fd<\/td>\nquery + \u7d22\u5f15\u7248\u672c\u53f7<\/td>\n\u540c\u4e0a<\/td>\n<\/tr>\n
    dense \u68c0\u7d22<\/td>\n\u80fd（\u6709\u98ce\u9669）<\/td>\nquery + \u7d22\u5f15\u7248\u672c\u53f7<\/td>\n\u540c\u4e0a，\u4f46\u7d22\u5f15\u66f4\u65b0\u540e\u5fc5\u987b\u5931\u6548<\/td>\n<\/tr>\n
    rerank<\/td>\n\u80fd<\/td>\n(query, chunk_id) \u5bf9<\/td>\n\u547d\u4e2d\u7387\u4f4e，\u6536\u76ca\u4e00\u822c<\/td>\n<\/tr>\n
    \u751f\u6210<\/td>\n\u614e\u7528<\/td>\nquery + \u4e0a\u4e0b\u6587\u6307\u7eb9<\/td>\n\u5bb9\u6613\u8fd4\u56de\u8fc7\u671f\u7b54\u6848，\u4e0d\u5efa\u8bae\u7f13\u5b58\u8d85\u8fc7 1 \u5c0f\u65f6<\/td>\n<\/tr>\n
    \u7f6e\u4fe1\u5ea6\u5224\u5b9a<\/td>\n\u4e0d\u80fd<\/td>\n\u2014<\/td>\n\u5b83\u662f\u51b3\u7b56\u4fe1\u53f7，\u7f13\u5b58\u4f1a\u8ba9\u5b83\u5931\u6548<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u7d22\u5f15\u7248\u672c\u53f7\u5fc5\u987b\u8fdb\u7f13\u5b58\u952e\u3002\u6211\u89c1\u8fc7\u56e0\u4e3a\u5fd8\u52a0\u7248\u672c\u53f7，\u7d22\u5f15\u91cd\u5efa\u4e4b\u540e\u68c0\u7d22\u7ed3\u679c\u8fd8\u662f\u65e7\u7684，\u6392\u67e5\u4e86\u4e24\u5929\u3002<\/p>\n

    \u4e5d\u3001\u9002\u7528\u573a\u666f\u4e0e\u9009\u578b\u5efa\u8bae<\/h3>\n

    9.1 \u6309\u67e5\u8be2\u5206\u5e03\u9009\u4ee3\u9645<\/h4>\n

    \u8fd9\u662f\u6700\u5b9e\u7528\u7684\u4e00\u5f20\u8868\u3002\u5148\u7edf\u8ba1\u4f60\u7684\u67e5\u8be2\u5206\u5e03，\u518d\u51b3\u5b9a\u6295\u5165\u54ea\u4e00\u4ee3：<\/p>\n\n\u591a\u8df3\u67e5\u8be2\u5360\u6bd4\u7cbe\u786e\u5339\u914d\u5360\u6bd4\u5efa\u8bae\u8bf4\u660e<\/tr>\n\n\n\n\n\n
    < 15%<\/td>\n\u4efb\u610f<\/td>\n\u53ea\u505a\u7b2c\u4e00\u4ee3<\/td>\n\u628a BM25 + \u53bb\u91cd + rerank \u505a\u624e\u5b9e，\u522b\u78b0\u56fe\u548c\u591a Agent<\/td>\n<\/tr>\n
    15%~35%<\/td>\n> 15%<\/td>\n\u7b2c\u4e00\u4ee3 + \u6309\u9700\u56fe\u6269\u5c55<\/td>\n\u56fe\u6269\u5c55\u53ea\u5728\u5224\u5b9a\u4e3a\u591a\u8df3\u65f6\u542f\u7528，\u522b\u9ed8\u8ba4\u5f00<\/td>\n<\/tr>\n
    > 35%<\/td>\n\u4efb\u610f<\/td>\n\u7b2c\u4e00\u4ee3 + \u7b2c\u4e8c\u4ee3 + \u6309\u9700 Agentic<\/td>\n\u503c\u5f97\u6295\u5165，\u4f46\u8981\u914d\u5957\u5ef6\u8fdf\u9884\u7b97\u548c\u7f6e\u4fe1\u5ea6\u95e8\u7981<\/td>\n<\/tr>\n
    \u4efb\u610f<\/td>\n> 30%<\/td>\n\u4f18\u5148\u4fdd\u8bc1 BM25 \u8d28\u91cf<\/td>\n\u5206\u8bcd\u5668\u548c\u6807\u8bc6\u7b26\u5904\u7406\u6bd4\u6362\u6a21\u578b\u91cd\u8981\u5f97\u591a<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    9.2 \u660e\u786e\u4e0d\u8be5\u7528\u7684\u573a\u666f<\/h4>\n

    **\u7b2c\u4e00，\u8bed\u6599\u89c4\u6a21\u5c0f\u4e8e 5000 chunk \u4e14\u672f\u8bed\u7edf\u4e00\u3002**\u8fd9\u79cd\u60c5\u51b5\u4e0b，\u76f4\u63a5\u628a\u5168\u90e8\u6587\u6863\u585e\u8fdb\u4e0a\u4e0b\u6587（\u914d\u5408\u5f15\u7528\u6807\u6ce8）\u5f80\u5f80\u6bd4\u642d\u4e00\u5957\u68c0\u7d22\u7cfb\u7edf\u66f4\u51c6\u3001\u66f4\u5feb\u3001\u66f4\u597d\u7ef4\u62a4\u3002\u68c0\u7d22\u7cfb\u7edf\u7684\u590d\u6742\u5ea6\u662f\u6709\u56fa\u5b9a\u6210\u672c\u7684，\u5c0f\u89c4\u6a21\u8bed\u6599\u644a\u4e0d\u5e73\u8fd9\u7b14\u6210\u672c\u3002<\/p>\n

    **\u7b2c\u4e8c，\u7b54\u6848\u5fc5\u987b 100% \u7cbe\u786e\u4e14\u4e0d\u5141\u8bb8\u62d2\u7b54\u7684\u573a\u666f\u3002**\u5408\u89c4\u6761\u6587\u5f15\u7528\u3001\u6cd5\u5f8b\u6761\u6b3e\u3001\u8d22\u52a1\u6570\u5b57\u2014\u2014\u8fd9\u4e9b\u573a\u666f\u91cc，\u68c0\u7d22\u7cfb\u7edf\u7684\u300c\u8bed\u4e49\u76f8\u8fd1\u300d\u7279\u6027\u672c\u8eab\u5c31\u662f\u98ce\u9669\u3002\u5e94\u8be5\u8d70\u786e\u5b9a\u6027\u7684\u7ed3\u6784\u5316\u67e5\u8be2（SQL\u3001\u56fe\u8c31\u67e5\u8be2\u3001\u89c4\u5219\u5f15\u64ce），\u800c\u4e0d\u662f\u5411\u91cf\u68c0\u7d22\u3002\u5411\u91cf\u68c0\u7d22\u7684\u8f93\u51fa\u662f\u300c\u53ef\u80fd\u76f8\u5173\u300d，\u800c\u67d0\u4e9b\u573a\u666f\u9700\u8981\u300c\u5fc5\u7136\u6b63\u786e\u300d\u3002<\/p>\n

    **\u7b2c\u4e09，\u8bed\u6599\u4e25\u91cd\u8fc7\u671f\u4e14\u6ca1\u6709\u66f4\u65b0\u673a\u5236\u3002**RAG \u4e0d\u4f1a\u81ea\u5df1\u53d8\u51c6，\u5b83\u53ea\u4f1a\u5fe0\u5b9e\u53cd\u6620\u5e93\u91cc\u7684\u4e1c\u897f\u3002\u5982\u679c\u6743\u5a01\u6e90\u672c\u8eab\u662f\u4e09\u5e74\u524d\u7684，\u4e0a\u4e86 GraphRAG \u53ea\u4f1a\u8ba9\u4f60\u66f4\u5feb\u3001\u66f4\u6709\u6761\u7406\u5730\u68c0\u7d22\u5230\u8fc7\u671f\u4fe1\u606f\u3002<\/p>\n

    **\u7b2c\u56db，\u6ca1\u6709\u6743\u5a01\u5b9e\u4f53\u8bcd\u5178\u5374\u786c\u4e0a GraphRAG\u3002**3.6 \u8282\u8bf4\u8fc7，\u89c4\u5219\u5316\u62bd\u53d6\u7684\u524d\u63d0\u662f\u6709\u53ef\u63a7\u8bcd\u5178\u3002\u5982\u679c\u4f01\u4e1a\u5185\u90e8\u8fde\u4ea7\u54c1\u5217\u8868\u3001\u7cfb\u7edf\u6e05\u5355\u90fd\u6ca1\u6709（\u8fd9\u79cd\u60c5\u51b5\u6bd4\u60f3\u8c61\u4e2d\u5e38\u89c1），\u90a3\u4e48\u5efa\u56fe\u7684\u7b2c\u4e00\u6b65\u5c31\u4f1a\u5361\u4f4f，\u6b64\u65f6\u5e94\u8be5\u5148\u82b1\u65f6\u95f4\u505a\u6570\u636e\u6cbb\u7406，\u800c\u4e0d\u662f\u5148\u505a\u56fe\u8c31\u3002<\/p>\n

    **\u7b2c\u4e94，\u5ef6\u8fdf\u9884\u7b97\u5c0f\u4e8e 500ms \u7684\u4ea4\u4e92\u573a\u666f\u3002**\u8fd9\u4e2a\u9884\u7b97\u88c5\u4e0d\u4e0b rerank（400ms）+ \u751f\u6210（1200ms）\u3002\u8981\u4e48\u780d\u6389 rerank \u63a5\u53d7\u8d28\u91cf\u635f\u5931，\u8981\u4e48\u6539\u6210\u5f02\u6b65\/\u6d41\u5f0f\u4ea4\u4e92\u3002<\/p>\n

    9.3 \u4e09\u4ee3\u4e0d\u662f\u5347\u7ea7\u8def\u7ebf，\u662f\u80fd\u529b\u53e0\u52a0<\/h4>\n

    \u6700\u540e\u6f84\u6e05\u4e00\u4e2a\u5e38\u89c1\u8bef\u8bfb：**\u4e0a\u4e86\u7b2c\u4e09\u4ee3\u4e0d\u4ee3\u8868\u8981\u6254\u6389\u7b2c\u4e00\u4ee3\u3002**\u7b2c\u4e09\u4ee3\u7684 Agentic \u5c42\u8c03\u7528\u7684\u7b2c\u4e00\u4ee3\u6df7\u5408\u68c0\u7d22，\u672c\u8d28\u4e0a\u8fd8\u662f dense + BM25 + RRF + \u53bb\u91cd + rerank\u3002GraphRAG \u7684\u56fe\u6269\u5c55\u4e5f\u662f\u5728\u6df7\u5408\u68c0\u7d22\u7684\u7ed3\u679c\u4e0a\u505a\u7684\u3002\u4e09\u4ee3\u662f\u80fd\u529b\u53e0\u52a0\u5173\u7cfb，\u4e0d\u662f\u66ff\u4ee3\u5173\u7cfb\u3002<\/p>\n

    \u6211\u89c1\u8fc7\u6700\u7cdf\u7cd5\u7684\u67b6\u6784\u662f：\u76f4\u63a5\u4e0a\u4e86\u4e00\u4e2a\u591a Agent \u7f16\u6392\u6846\u67b6，\u91cc\u9762\u6bcf\u4e2a Agent \u90fd\u5728\u505a\u5355\u8def\u5411\u91cf\u68c0\u7d22，\u7136\u540e\u9760\u300c\u591a\u8ba8\u8bba\u51e0\u8f6e\u300d\u6765\u5f25\u8865\u53ec\u56de\u8d28\u91cf\u3002\u8fd9\u662f\u7528\u6700\u8d35\u7684\u90e8\u4ef6\u53bb\u8865\u6700\u4fbf\u5b9c\u7684\u90e8\u4ef6\u7559\u4e0b\u7684\u6d1e\u3002<\/p>\n

    \u6b63\u786e\u7684\u987a\u5e8f\u6c38\u8fdc\u662f：\u5148\u628a\u7b2c\u4e00\u4ee3\u505a\u624e\u5b9e（BM25 \u878d\u5408 + \u4e09\u5c42\u53bb\u91cd + rerank），\u6d4b\u91cf\u4f60\u7684\u67e5\u8be2\u5206\u5e03，\u518d\u51b3\u5b9a\u8981\u4e0d\u8981\u52a0\u7b2c\u4e8c\u4ee3\u548c\u7b2c\u4e09\u4ee3\u3002<\/p>\n

    \u5341\u3001\u603b\u7ed3\u4e0e\u5c55\u671b<\/h3>\n

    \u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb，\u672c\u8d28\u662f\u628a\u5de5\u7a0b\u7ea6\u675f\u4e00\u6761\u6761\u663e\u5f0f\u5316\u7684\u8fc7\u7a0b\u3002<\/p>\n

    \u7b2c\u4e00\u4ee3\u6df7\u5408\u68c0\u7d22\u89e3\u51b3\u7684\u662f\u53ec\u56de\u4e0d\u5168：dense \u7ba1\u540c\u4e49，BM25 \u7ba1\u7cbe\u786e\u5339\u914d\u3001\u4ea7\u54c1\u7801\u3001\u7f29\u5199，\u4e24\u8def\u5e76\u53d1\u6267\u884c\u628a\u68c0\u7d22\u5ef6\u8fdf\u538b\u6389 40% \u4ee5\u4e0a（\u672c\u673a\u5b9e\u6d4b 287ms \u5230 168ms，\u964d 41.5%），\u518d\u7528 RRF \u505a\u514d\u5f52\u4e00\u5316\u7684\u878d\u5408\u2014\u2014\u53ea\u7528\u6392\u540d\u4e0d\u7528\u5206\u6570，\u6240\u4ee5\u8eb2\u5f00\u4e86\u300c\u4e24\u8def\u5206\u6570\u4e0d\u53ef\u6bd4\u300d\u8fd9\u4e2a\u96be\u9898，k \u9ed8\u8ba4 60 \u4e0d\u7528\u8c03\u3002\u4e4b\u540e\u5fc5\u987b\u505a\u4e09\u5c42\u53bb\u91cd（\u552f\u4e00 ID \/ \u6765\u6e90\u4f4d\u7f6e\u6876 \/ \u5185\u5bb9\u6307\u7eb9）\u518d\u8fdb cross-encoder \u91cd\u6392，\u987a\u5e8f\u53cd\u4e86\u5c31\u65e2\u8d39\u94b1\u53c8\u964d\u8d28\u3002\u8fd9\u4e00\u4ee3\u8d21\u732e\u7684 nDCG \u589e\u76ca\u6700\u5927（0.486 \u5230 0.703），\u800c\u5ef6\u8fdf\u53ea\u589e\u52a0 55ms\u3002<\/p>\n

    \u7b2c\u4e8c\u4ee3 GraphRAG \u89e3\u51b3\u7684\u662f chunk \u5b64\u7acb：\u628a\u5b9e\u4f53\u3001\u5173\u7cfb\u3001\u672c\u4f53\u5c42\u5efa\u8d77\u6765，\u8ba9\u67e5\u8be2\u80fd\u987a\u7740\u4ea7\u54c1\u5c42\u7ea7\u3001\u540c\u4e49\u8bcd\u3001\u5171\u4eab\u7236\u7c7b\u8d70\u591a\u8df3\u3002\u5173\u952e\u5de5\u7a0b\u5224\u65ad\u662f\u751f\u4ea7\u73af\u5883\u7528\u89c4\u5219\u5316\u591a\u8d9f\u5b9e\u4f53\u62bd\u53d6（\u6700\u957f\u77ed\u8bed\u4f18\u5148 \/ \u5f52\u4e00\u5316 \/ token \u7ea7）\u4f18\u4e8e LLM NER\u2014\u2014\u5fae\u79d2\u7ea7\u3001\u96f6\u8fb9\u9645\u6210\u672c\u3001\u786e\u5b9a\u6027\u3001\u53ef\u590d\u73b0，8 \u4e07 chunk \u5168\u91cf 3.3 \u79d2，\u800c LLM \u65b9\u6848\u8981 75 \u5206\u949f\u4e14\u7ed3\u679c\u4e0d\u53ef\u590d\u73b0\u3002\u89c4\u5219\u5316\u65b9\u6848\u7684\u524d\u63d0\u662f\u6709\u6743\u5a01\u5b9e\u4f53\u8bcd\u5178，\u4f01\u4e1a\u573a\u666f\u6070\u597d\u6ee1\u8db3\u3002<\/p>\n

    \u7b2c\u4e09\u4ee3 Agentic \u89e3\u51b3\u7684\u662f\u591a\u8df3\u4e0e\u53ef\u63a7：\u5de5\u5177\u4f7f\u7528\u3001\u8bb0\u5fc6\u7ba1\u7406\u3001\u89c4\u5212\u3001\u534f\u8c03\u3001\u8bc4\u4f30\u4e94\u6761\u7eaa\u5f8b，\u52a0\u4e0a\u5fc5\u987b\u524d\u7f6e\u5230\u68c0\u7d22\u4e4b\u524d\u7684\u5b89\u5168\u8fb9\u754c（\u5ba2\u6237\u6570\u636e\u3001\u5458\u5de5\u4fe1\u606f\u3001\u6d89\u5bc6\u5f15\u7528\u5728\u68c0\u7d22\u6216\u751f\u6210\u524d\u62e6\u622a，ACL \u8fc7\u6ee4\u5fc5\u987b\u5728\u68c0\u7d22\u5668\u5185\u90e8）\u3002\u7f6e\u4fe1\u5ea6\u7528\u4e58\u6cd5\u800c\u4e0d\u662f\u5e73\u5747\u2014\u20140.9 x 0.9 = 0.81，0.9 x 0.1 = 0.09 \u800c\u4e0d\u662f\u8bef\u5bfc\u6027\u7684 0.5，\u5e73\u5747\u4f1a\u628a\u4e00\u4e2a\u81f4\u547d\u7684\u5f31\u73af\u8282\u63a9\u76d6\u6210\u4e2d\u7b49\u5206\u6570；\u4f4e\u4e8e\u9608\u503c\u89e6\u53d1\u6709\u754c\u91cd\u8bd5（\u9884\u5b9a\u4e49\u7b56\u7565\u5e8f\u5217 + \u53bb\u91cd + \u5ef6\u8fdf\u9884\u7b97，\u8d70\u5b8c\u5c31\u62d2\u7b54）\u3002\u6bcf\u4e2a\u7ec4\u4ef6\u5206\u914d\u663e\u5f0f\u5ef6\u8fdf\u9884\u7b97\u4e14\u5404\u914d\u964d\u7ea7\u8def\u5f84，\u9a8c\u6536\u770b P99 \u4e0d\u770b P50\u3002<\/p>\n

    \u4e09\u4e2a\u5fc5\u987b\u907f\u5f00\u53cd\u6a21\u5f0f：\u5411\u91cf\u5e93\u88ab AI \u751f\u6210\u7684\u6458\u8981\u6c61\u67d3（ingestion \u52a0\u5206\u7c7b\u5c42，derived \u5fc5\u987b\u6307\u56de canonical \u4e14\u4e0d\u4e0e canonical \u540c\u65f6\u8fdb top-8）；\u65e0\u9650\u585e\u4e0a\u4e0b\u6587（\u6309\u9700\u4ece top-8 \u6709\u754c\u6269\u5c55，\u9971\u548c\u70b9\u901a\u5e38\u5c31\u5728 8 \u6761）；AI \u540c\u65f6\u8bfb\u5199\u540c\u4e00\u4e2a\u5e93（\u8bfb\u5199\u5e93\u7269\u7406\u9694\u79bb + \u5ba1\u6838\u95f8\u95e8，\u5426\u5219\u51e0\u8f6e\u4e4b\u540e\u5e93\u91cc\u5168\u662f\u81ea\u6211\u5f3a\u5316\u7684 AI \u751f\u6210\u7269）\u3002<\/p>\n

    \u5c55\u671b\u4e09\u6761\u3002\u7b2c\u4e00，grounded \u5224\u5b9a\u4f1a\u6210\u4e3a\u72ec\u7acb\u7684\u4ea7\u54c1\u80fd\u529b\u800c\u4e0d\u662f RAG \u7684\u5185\u90e8\u7ec6\u8282\u2014\u2014\u7528\u6237\u8981\u7684\u4e0d\u662f\u7b54\u6848，\u662f\u300c\u8fd9\u53e5\u8bdd\u51fa\u81ea\u54ea\u4efd\u6587\u6863\u7684\u54ea\u4e00\u8282\u3001\u8fd9\u4efd\u6587\u6863\u4ec0\u4e48\u65f6\u5019\u66f4\u65b0\u7684\u300d，\u8fd9\u4e2a\u9700\u6c42\u4f1a\u5012\u903c\u68c0\u7d22\u5c42\u66b4\u9732\u66f4\u7ec6\u7684\u5f15\u7528\u7c92\u5ea6\u3002\u7b2c\u4e8c，\u5b9e\u4f53\u4e0e\u5173\u7cfb\u7684\u8d28\u91cf\u4f1a\u53d6\u4ee3\u5411\u91cf\u6a21\u578b\u7684\u9009\u62e9，\u6210\u4e3a\u4f01\u4e1a RAG \u6548\u679c\u7684\u4e3b\u8981\u53d8\u91cf，\u800c\u8fd9\u4ef6\u4e8b\u672c\u8d28\u662f\u6570\u636e\u6cbb\u7406，\u4e0d\u662f\u7b97\u6cd5\u3002\u7b2c\u4e09，\u968f\u7740 2026-09 \u524d\u540e\u4e1a\u754c\u5bf9 Agent \u8fd0\u884c\u65f6\u4e0e\u4e0a\u4e0b\u6587\u5de5\u7a0b\u6807\u51c6\u5316\u7684\u63a8\u8fdb（\u4fe1\u901a\u9662\u7684\u53ef\u4fe1\u667a\u80fd\u4f53\u8fd0\u884c\u65f6\u6807\u51c6\u3001CIS \u7684 MCP \u5ba1\u8ba1\u57fa\u7ebf\u90fd\u5728\u505a\u8fd9\u4ef6\u4e8b），\u68c0\u7d22\u4e0e\u751f\u6210\u7684\u8fb9\u754c\u4f1a\u88ab\u5199\u8fdb\u89c4\u8303，\u5c4a\u65f6\u300c\u6709\u6ca1\u6709 grounded \u5224\u5b9a\u300d\u4f1a\u4ece\u4e00\u4e2a\u6280\u672f\u9009\u578b\u53d8\u6210\u4e00\u4e2a\u5408\u89c4\u8981\u6c42\u3002<\/p>\n

    \u6700\u540e\u7ed9\u4e00\u6761\u843d\u5730\u987a\u5e8f，\u6309\u6027\u4ef7\u6bd4\u4ece\u9ad8\u5230\u4f4e，\u6bcf\u4e00\u6b65\u90fd\u80fd\u72ec\u7acb\u751f\u6548：\u52a0 BM25 \u5e76\u4fee\u597d\u5206\u8bcd\u5668（7.2）-> \u52a0\u4e09\u5c42\u53bb\u91cd（7.1）-> \u4e24\u8def\u6539\u5e76\u53d1（2.3）-> \u52a0 rerank \u5e76\u9a8c\u622a\u65ad\u504f\u5dee（7.3）-> \u52a0\u4e58\u6cd5\u7f6e\u4fe1\u5ea6\u4e0e\u4e00\u7968\u5426\u51b3（7.5）-> \u52a0\u7ec4\u4ef6\u7ea7\u5ef6\u8fdf\u9884\u7b97（7.7）-> \u6700\u540e\u518d\u8bc4\u4f30\u8981\u4e0d\u8981\u4e0a\u56fe\u6269\u5c55\u548c Agentic\u3002<\/p>\n

    \u53c2\u8003\u7d20\u6750\u4e0e\u5ef6\u4f38\u9605\u8bfb<\/h3>\n
      \n
    • \u4f01\u4e1a RAG \u4e0e Agentic AI \u7cfb\u7edf：\u4e09\u4ee3\u6f14\u8fdb\u68b3\u7406 \u2014 KDnuggets，2026-09-08<\/li>\n
    • Graph RAG \u4e0e Agentic RAG \u7684\u9009\u578b\u5bf9\u6bd4 \u2014 CSDN，2026<\/li>\n
    • Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods \u2014 Cormack, Clarke, Buettcher，SIGIR 2009<\/li>\n
    • Agent \u4e0a\u4e0b\u6587\u4e09\u7ea7\u538b\u7f29\u673a\u5236 \u2014 \u817e\u8baf\u65b0\u95fb，2026-09-22<\/li>\n
    • \u4e2d\u56fd\u4fe1\u901a\u9662\u300a\u53ef\u4fe1\u4e92\u8054\u7f51\u667a\u80fd\u4f53 \u8fd0\u884c\u65f6 \u4e0a\u4e0b\u6587\u5de5\u7a0b\u300b\u6807\u51c6 \u2014 \u6f8e\u6e43\u65b0\u95fb，2026-09-29<\/li>\n
    • CIS MCP Benchmark \u4e3a Agent \u5de5\u5177\u6cbb\u7406\u8bbe\u5b9a 55 \u70b9\u5ba1\u8ba1\u57fa\u7ebf \u2014 AI Governance，2026-09-16<\/li>\n
    • \u591a Agent \u67b6\u6784\u5411 MCP + A2A \u534f\u8bae\u6536\u655b \u2014 The Agent Times<\/li>\n
    • LLM \u751f\u4ea7\u53ef\u89c2\u6d4b\u6027\u7684\u5de5\u5177\u5206\u5c42 \u2014 Unico Connect<\/li>\n
    • Google ADK 2.0 Harness Engineering \u4e0e Anthropic \u5546\u52a1 Agent \u84dd\u56fe \u2014 AtlasNote \u6bcf\u65e5\u7b80\u62a5，2026-09-03<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

      \u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb：\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic \u4f01\u4e1a\u91cc\u7684\u6734\u7d20 RAG \u5931\u8d25\u4ece\u6765\u4e0d\u662f\u56e0\u4e3a\u6a21\u578b\u4e0d\u591f\u5f3a，\u800c\u662f\u56e0\u4e3a\u56db\u4e2a\u7ed3\u6784\u6027\u95ee\u9898：\u7f29\u5199\u591a\u4e49（CRM \u5230\u5e95\u662f\u5ba2\u6237\u5173\u7cfb\u7cfb\u7edf\u8fd8\u662f\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406）\u3001\u672f\u8bed\u4e0d\u4e00\u81f4（\u540c\u4e00\u4e2a\u4e1c\u897f\u5728\u4e09\u4e2a\u90e8\u95e8\u6709\u4e09\u4e2a\u540d\u5b57）\u3001\u65e0\u6cd5\u56de\u7b54\u300c\u8fd9\u53e5\u8bdd\u662f\u4e0d\u662f grounded（\u6709\u636e\u53ef\u4f9d）\u300d\u3001\u4ee5\u53ca\u5355\u8df3\u68c0\u7d22\u5929\u751f\u7b54\u4e0d\u4e86\u6bd4\u8f83\u3001\u89c4\u5212\u3001\u6392\u969c\u8fd9\u7c7b<\/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":[50],"topic":[],"class_list":["post-112755","post","type-post","status-publish","format-standard","hentry","category-server","tag-50"],"yoast_head":"\n\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic - \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\/112755.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb：\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic \u4f01\u4e1a\u91cc\u7684\u6734\u7d20 RAG \u5931\u8d25\u4ece\u6765\u4e0d\u662f\u56e0\u4e3a\u6a21\u578b\u4e0d\u591f\u5f3a，\u800c\u662f\u56e0\u4e3a\u56db\u4e2a\u7ed3\u6784\u6027\u95ee\u9898：\u7f29\u5199\u591a\u4e49（CRM \u5230\u5e95\u662f\u5ba2\u6237\u5173\u7cfb\u7cfb\u7edf\u8fd8\u662f\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406）\u3001\u672f\u8bed\u4e0d\u4e00\u81f4（\u540c\u4e00\u4e2a\u4e1c\u897f\u5728\u4e09\u4e2a\u90e8\u95e8\u6709\u4e09\u4e2a\u540d\u5b57）\u3001\u65e0\u6cd5\u56de\u7b54\u300c\u8fd9\u53e5\u8bdd\u662f\u4e0d\u662f grounded（\u6709\u636e\u53ef\u4f9d）\u300d\u3001\u4ee5\u53ca\u5355\u8df3\u68c0\u7d22\u5929\u751f\u7b54\u4e0d\u4e86\u6bd4\u8f83\u3001\u89c4\u5212\u3001\u6392\u969c\u8fd9\u7c7b\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/112755.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-04T12:17:20+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=\"36 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/112755.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/112755.html\",\"name\":\"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-04T12:17:20+00:00\",\"dateModified\":\"2026-10-04T12:17:20+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/112755.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/112755.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/112755.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic\"}]},{\"@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":"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic - \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\/112755.html","og_locale":"zh_CN","og_type":"article","og_title":"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb：\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic \u4f01\u4e1a\u91cc\u7684\u6734\u7d20 RAG \u5931\u8d25\u4ece\u6765\u4e0d\u662f\u56e0\u4e3a\u6a21\u578b\u4e0d\u591f\u5f3a，\u800c\u662f\u56e0\u4e3a\u56db\u4e2a\u7ed3\u6784\u6027\u95ee\u9898：\u7f29\u5199\u591a\u4e49（CRM \u5230\u5e95\u662f\u5ba2\u6237\u5173\u7cfb\u7cfb\u7edf\u8fd8\u662f\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406）\u3001\u672f\u8bed\u4e0d\u4e00\u81f4（\u540c\u4e00\u4e2a\u4e1c\u897f\u5728\u4e09\u4e2a\u90e8\u95e8\u6709\u4e09\u4e2a\u540d\u5b57）\u3001\u65e0\u6cd5\u56de\u7b54\u300c\u8fd9\u53e5\u8bdd\u662f\u4e0d\u662f grounded（\u6709\u636e\u53ef\u4f9d）\u300d\u3001\u4ee5\u53ca\u5355\u8df3\u68c0\u7d22\u5929\u751f\u7b54\u4e0d\u4e86\u6bd4\u8f83\u3001\u89c4\u5212\u3001\u6392\u969c\u8fd9\u7c7b","og_url":"https:\/\/www.wsisp.com\/helps\/112755.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-04T12:17:20+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"36 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/112755.html","url":"https:\/\/www.wsisp.com\/helps\/112755.html","name":"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-04T12:17:20+00:00","dateModified":"2026-10-04T12:17:20+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/112755.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/112755.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/112755.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u4f01\u4e1a RAG \u7684\u4e09\u4ee3\u6f14\u8fdb\uff1a\u4ece\u6df7\u5408\u68c0\u7d22\u5230 GraphRAG \u518d\u5230 Agentic"}]},{"@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\/112755","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=112755"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/112755\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=112755"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=112755"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=112755"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=112755"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}