{"id":16457,"date":"2024-12-26T11:27:58","date_gmt":"2024-12-26T03:27:58","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=16457"},"modified":"2025-07-14T07:20:39","modified_gmt":"2025-07-13T23:20:39","slug":"kag","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/kag\/","title":{"rendered":"KAG\uff1a\u77e5\u8bc6\u56fe\u8c31\u4e0e\u5411\u91cf\u6df7\u5408\u68c0\u7d22\u7684\u4e13\u4e1a\u77e5\u8bc6\u5e93\u95ee\u7b54\u6846\u67b6"},"content":{"rendered":"<p>KAG (Knowledge Augmented Generation) \u662f\u4e00\u4e2a\u57fa\u4e8eOpenSPG\u5f15\u64ce\u548c\u5927\u8bed\u8a00\u6a21\u578b(LLMs)\u7684\u903b\u8f91\u5f62\u5f0f\u5f15\u5bfc\u7684\u63a8\u7406\u548c\u68c0\u7d22\u6846\u67b6\u3002\u8be5\u6846\u67b6\u4e13\u95e8\u7528\u4e8e\u6784\u5efa\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u5e93\u7684\u903b\u8f91\u63a8\u7406\u548c\u4e8b\u5b9e\u95ee\u7b54\u89e3\u51b3\u65b9\u6848\uff0c\u80fd\u6709\u6548\u514b\u670d\u4f20\u7edfRAG\uff08\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff09\u5411\u91cf\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u6a21\u578b\u7684\u7f3a\u9677\u3002KAG\u901a\u8fc7\u77e5\u8bc6\u56fe\u8c31\u4e0e\u5411\u91cf\u68c0\u7d22\u7684\u4f18\u52bf\u4e92\u8865\uff0c\u4ece\u56db\u4e2a\u65b9\u9762\u53cc\u5411\u589e\u5f3a\u5927\u8bed\u8a00\u6a21\u578b\u548c\u77e5\u8bc6\u56fe\u8c31\uff1aLLM\u53cb\u597d\u7684\u77e5\u8bc6\u8868\u793a\u3001\u77e5\u8bc6\u56fe\u8c31\u4e0e\u539f\u59cb\u6587\u672c\u7247\u6bb5\u4e4b\u95f4\u7684\u76f8\u4e92\u7d22\u5f15\u3001\u6df7\u5408\u63a8\u7406\u6c42\u89e3\u5668\u3001\u4ee5\u53ca\u53ef\u4fe1\u6027\u8bc4\u4f30\u673a\u5236\u3002\u8be5\u6846\u67b6\u7279\u522b\u9002\u5408\u5904\u7406\u6570\u503c\u8ba1\u7b97\u3001\u65f6\u5e8f\u5173\u7cfb\u548c\u4e13\u5bb6\u89c4\u5219\u7b49\u590d\u6742\u7684\u77e5\u8bc6\u903b\u8f91\u95ee\u9898\uff0c\u4e3a\u4e13\u4e1a\u9886\u57df\u5e94\u7528\u63d0\u4f9b\u4e86\u66f4\u51c6\u786e\u548c\u53ef\u9760\u7684\u95ee\u7b54\u80fd\u529b\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-16462\" title=\"KAG\uff1a\u77e5\u8bc6\u56fe\u8c31\u4e0e\u5411\u91cf\u6df7\u5408\u68c0\u7d22\u7684\u4e13\u4e1a\u77e5\u8bc6\u5e93\u95ee\u7b54\u6846\u67b6-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/10e74b94bef80ac.jpg\" alt=\"KAG\uff1a\u77e5\u8bc6\u56fe\u8c31\u4e0e\u5411\u91cf\u6df7\u5408\u68c0\u7d22\u7684\u4e13\u4e1a\u77e5\u8bc6\u5e93\u95ee\u7b54\u6846\u67b6-1\" width=\"1665\" height=\"799\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/10e74b94bef80ac.jpg 1665w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/10e74b94bef80ac-300x144.jpg 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/10e74b94bef80ac-1024x491.jpg 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/10e74b94bef80ac-768x369.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/10e74b94bef80ac-1536x737.jpg 1536w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/10e74b94bef80ac-18x9.jpg 18w\" sizes=\"auto, (max-width: 1665px) 100vw, 1665px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone wp-image-16466\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/849f866725ff436.png\" alt=\"\" width=\"517\" height=\"1356\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/849f866725ff436.png 1464w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/849f866725ff436-114x300.png 114w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/849f866725ff436-390x1024.png 390w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/849f866725ff436-768x2014.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/849f866725ff436-586x1536.png 586w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/849f866725ff436-781x2048.png 781w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/849f866725ff436-5x12.png 5w\" sizes=\"auto, (max-width: 517px) 100vw, 517px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li>\u652f\u6301\u590d\u6742\u903b\u8f91\u5f62\u5f0f\u7684\u63a8\u7406\u80fd\u529b<\/li>\n<li>\u63d0\u4f9b\u77e5\u8bc6\u56fe\u8c31\u548c\u5411\u91cf\u68c0\u7d22\u7684\u6df7\u5408\u68c0\u7d22\u673a\u5236<\/li>\n<li>\u5b9e\u73b0LLM\u53cb\u597d\u7684\u77e5\u8bc6\u8868\u793a\u8f6c\u6362<\/li>\n<li>\u652f\u6301\u77e5\u8bc6\u7ed3\u6784\u548c\u6587\u672c\u5757\u7684\u53cc\u5411\u7d22\u5f15<\/li>\n<li>\u96c6\u6210LLM\u63a8\u7406\u3001\u77e5\u8bc6\u63a8\u7406\u548c\u6570\u5b66\u903b\u8f91\u63a8\u7406<\/li>\n<li>\u63d0\u4f9b\u53ef\u4fe1\u6027\u8bc4\u4f30\u548c\u9a8c\u8bc1\u673a\u5236<\/li>\n<li>\u652f\u6301\u591a\u8df3\u95ee\u7b54\u548c\u590d\u6742\u67e5\u8be2\u5904\u7406<\/li>\n<li>\u63d0\u4f9b\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u5e93\u7684\u5b9a\u5236\u5316\u89e3\u51b3\u65b9\u6848<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>1. \u73af\u5883\u51c6\u5907<\/h3>\n<p>\u9996\u5148\u9700\u8981\u786e\u4fdd\u60a8\u7684\u7cfb\u7edf\u6ee1\u8db3\u4ee5\u4e0b\u8981\u6c42\uff1a<\/p>\n<ul>\n<li>Python 3.8\u6216\u66f4\u9ad8\u7248\u672c<\/li>\n<li>OpenSPG\u5f15\u64ce\u73af\u5883<\/li>\n<li>\u652f\u6301\u7684\u5927\u8bed\u8a00\u6a21\u578bAPI\u63a5\u53e3<\/li>\n<\/ul>\n<h3>2. \u5b89\u88c5\u6b65\u9aa4<\/h3>\n<ol>\n<li>\u514b\u9686\u9879\u76ee\u4ed3\u5e93\uff1a<\/li>\n<\/ol>\n<pre><code>git clone https:\/\/github.com\/OpenSPG\/KAG.git\r\ncd KAG\r\n<\/code><\/pre>\n<ol start=\"2\">\n<li>\u5b89\u88c5\u4f9d\u8d56\u5305\uff1a<\/li>\n<\/ol>\n<pre><code>pip install -r requirements.txt\r\n<\/code><\/pre>\n<h3>3. \u6846\u67b6\u4f7f\u7528\u6d41\u7a0b<\/h3>\n<h4>3.1 \u77e5\u8bc6\u5e93\u51c6\u5907<\/h4>\n<ul>\n<li>\u5bfc\u5165\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u6570\u636e<\/li>\n<li>\u914d\u7f6e\u77e5\u8bc6\u56fe\u8c31\u6a21\u5f0f<\/li>\n<li>\u5efa\u7acb\u6587\u672c\u7d22\u5f15\u7cfb\u7edf<\/li>\n<\/ul>\n<h4>3.2 \u67e5\u8be2\u5904\u7406<\/h4>\n<ol>\n<li>\u95ee\u9898\u8f93\u5165\uff1a\u7cfb\u7edf\u63a5\u6536\u7528\u6237\u7684\u81ea\u7136\u8bed\u8a00\u95ee\u9898<\/li>\n<li>\u903b\u8f91\u5f62\u5f0f\u8f6c\u6362\uff1a\u5c06\u95ee\u9898\u8f6c\u6362\u4e3a\u6807\u51c6\u5316\u7684\u903b\u8f91\u8868\u8fbe\u5f0f<\/li>\n<li>\u6df7\u5408\u68c0\u7d22\uff1a\n<ul>\n<li>\u8fdb\u884c\u77e5\u8bc6\u56fe\u8c31\u68c0\u7d22<\/li>\n<li>\u6267\u884c\u5411\u91cf\u76f8\u4f3c\u5ea6\u641c\u7d22<\/li>\n<li>\u6574\u5408\u68c0\u7d22\u7ed3\u679c<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h4>3.3 \u63a8\u7406\u8fc7\u7a0b<\/h4>\n<ol>\n<li>\u903b\u8f91\u63a8\u7406\uff1a\u4f7f\u7528\u6df7\u5408\u63a8\u7406\u6c42\u89e3\u5668\u8fdb\u884c\u591a\u6b65\u63a8\u7406<\/li>\n<li>\u77e5\u8bc6\u878d\u5408\uff1a\u7ed3\u5408LLM\u63a8\u7406\u548c\u77e5\u8bc6\u56fe\u8c31\u63a8\u7406\u7ed3\u679c<\/li>\n<li>\u7b54\u6848\u751f\u6210\uff1a\u5f62\u6210\u6700\u7ec8\u7684\u56de\u7b54<\/li>\n<\/ol>\n<h4>3.4 \u53ef\u4fe1\u6027\u4fdd\u969c<\/h4>\n<ul>\n<li>\u7b54\u6848\u9a8c\u8bc1<\/li>\n<li>\u63a8\u7406\u8def\u5f84\u8ffd\u8e2a<\/li>\n<li>\u7f6e\u4fe1\u5ea6\u8bc4\u4f30<\/li>\n<\/ul>\n<h3>4. \u9ad8\u7ea7\u529f\u80fd\u4f7f\u7528<\/h3>\n<h4>4.1 \u81ea\u5b9a\u4e49\u77e5\u8bc6\u8868\u793a<\/h4>\n<p>\u53ef\u4ee5\u6839\u636e\u4e13\u4e1a\u9886\u57df\u9700\u6c42\u81ea\u5b9a\u4e49\u77e5\u8bc6\u8868\u793a\u683c\u5f0f\uff0c\u786e\u4fdd\u4e0eLLM\u7684\u517c\u5bb9\u6027\uff1a<\/p>\n<pre><code># \u793a\u4f8b\u4ee3\u7801\r\nknowledge_config = {\r\n\"domain\": \"your_domain\",\r\n\"schema\": your_schema_definition,\r\n\"representation\": your_custom_representation\r\n}\r\n<\/code><\/pre>\n<h4>4.2 \u63a8\u7406\u89c4\u5219\u914d\u7f6e<\/h4>\n<p>\u53ef\u4ee5\u914d\u7f6e\u4e13\u95e8\u7684\u63a8\u7406\u89c4\u5219\u6765\u5904\u7406\u7279\u5b9a\u9886\u57df\u7684\u903b\u8f91\uff1a<\/p>\n<pre><code># \u793a\u4f8b\u4ee3\u7801\r\nreasoning_rules = {\r\n\"numerical\": numerical_processing_rules,\r\n\"temporal\": temporal_reasoning_rules,\r\n\"domain_specific\": your_domain_rules\r\n}\r\n<\/code><\/pre>\n<h3>5. \u6700\u4f73\u5b9e\u8df5<\/h3>\n<ul>\n<li>\u786e\u4fdd\u77e5\u8bc6\u5e93\u6570\u636e\u8d28\u91cf\u548c\u5b8c\u6574\u6027<\/li>\n<li>\u4f18\u5316\u68c0\u7d22\u7b56\u7565\u4ee5\u63d0\u9ad8\u6548\u7387<\/li>\n<li>\u5b9a\u671f\u66f4\u65b0\u548c\u7ef4\u62a4\u77e5\u8bc6\u5e93<\/li>\n<li>\u76d1\u63a7\u7cfb\u7edf\u6027\u80fd\u548c\u51c6\u786e\u6027<\/li>\n<li>\u6536\u96c6\u7528\u6237\u53cd\u9988\u8fdb\u884c\u6301\u7eed\u6539\u8fdb<\/li>\n<\/ul>\n<h3>6. \u5e38\u89c1\u95ee\u9898\u89e3\u51b3<\/h3>\n<ul>\n<li>\u5982\u9047\u5230\u68c0\u7d22\u6548\u7387\u95ee\u9898\uff0c\u53ef\u4ee5\u9002\u5f53\u8c03\u6574\u7d22\u5f15\u53c2\u6570<\/li>\n<li>\u5bf9\u4e8e\u590d\u6742\u67e5\u8be2\uff0c\u53ef\u4ee5\u4f7f\u7528\u5206\u6b65\u63a8\u7406\u7b56\u7565<\/li>\n<li>\u63a8\u7406\u7ed3\u679c\u4e0d\u51c6\u786e\u65f6\uff0c\u68c0\u67e5\u77e5\u8bc6\u8868\u793a\u548c\u89c4\u5219\u914d\u7f6e<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>KAG\u9879\u76ee\u8bb2\u89e3<\/h2>\n<h3>1. \u5f15\u8a00<\/h3>\n<p>\u524d\u51e0\u5929\u8682\u8681\u6b63\u5f0f\u53d1\u5e03\u4e86\u4e00\u4e2a\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u670d\u52a1\u6846\u67b6, \u53eb\u505a\u77e5\u8bc6\u589e\u5f3a\u751f\u6210(KAG\uff1aKnowledge Augmented Generation)\uff0c\u8be5\u6846\u67b6\u65e8\u5728\u5145\u5206\u5229\u7528\u77e5\u8bc6\u56fe\u8c31\u548c\u5411\u91cf\u68c0\u7d22\u7684\u4f18\u52bf\uff0c\u4ee5\u89e3\u51b3\u73b0\u6709 <a href=\"https:\/\/www.kdjingpai.com\/en\/rag\/\">RAG<\/a> \u6280\u672f\u6808\u7684\u4e00\u4e9b\u6311\u6218\u3002<\/p>\n<p>\u4ece\u8682\u8681\u5bf9\u8fd9\u4e2a\u6846\u67b6\u9884\u70ed\u5f00\u59cb\uff0c\u7b14\u8005\u5c31\u5bf9 KAG \u7684\u4e00\u4e9b\u6838\u5fc3\u529f\u80fd\u6bd4\u8f83\u611f\u5174\u8da3\uff0c\u5c24\u5176\u662f\u903b\u8f91\u7b26\u53f7\u63a8\u7406\u4e0e\u77e5\u8bc6\u5bf9\u9f50\uff0c\u5728\u73b0\u6709\u4e3b\u6d41 RAG \u7cfb\u7edf\u4e2d\uff0c\u8fd9\u4e24\u70b9\u8ba8\u8bba\u8c8c\u4f3c\u8fd8\u4e0d\u7b97\u591a\uff0c\u8d81\u7740\u8fd9\u6b21\u5f00\u6e90\uff0c\u8d76\u7d27\u7814\u7a76\u4e00\u6ce2\u3002<\/p>\n<ul>\n<li>KAG\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2409.13731<\/li>\n<li>KAG\u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/OpenSPG\/KAG<\/li>\n<\/ul>\n<h3>2. \u6846\u67b6\u6982\u8ff0<\/h3>\n<p>\u5177\u4f53\u7814\u8bfb\u4ee3\u7801\u524d\uff0c\u6211\u4eec\u8fd8\u662f\u5148\u7b80\u5355\u4e86\u89e3\u4e0b\u6846\u67b6\u7684\u76ee\u6807\u4e0e\u5b9a\u4f4d\u3002<\/p>\n<h4>2.1 \u89e3\u51b3\u4e86\u4ec0\u4e48\u95ee\u9898( What &amp; Why)?<\/h4>\n<p>\u5176\u5b9e\u770b\u5230 KAG \u8fd9\u4e2a\u6846\u67b6\uff0c\u6211\u76f8\u4fe1\u5f88\u591a\u4eba\u4f30\u8ba1\u8ddf\u6211\u4e00\u6837\uff0c\u60f3\u5230\u7684\u7b2c\u4e00\u4e2a\u95ee\u9898\u5c31\u662f\u4e3a\u4ec0\u4e48\u4e0d\u53eb RAG \u6539\u53eb KAG \u4e86\u3002\u6839\u636e\u76f8\u5173\u6587\u7ae0\u4e0e\u8bba\u6587\uff0cKAG \u6846\u67b6\u4e3b\u8981\u662f\u4e3a\u4e86\u89e3\u51b3\u5f53\u524d\u5927\u6a21\u578b\u5728\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u670d\u52a1\u4e2d\u9762\u4e34\u7684\u4e00\u4e9b\u6311\u6218\uff1a<\/p>\n<ul>\n<li>LLM \u4e0d\u5177\u5907\u4e25\u8c28\u7684\u601d\u8003\u80fd\u529b\uff0c\u63a8\u7406\u80fd\u529b\u7f3a\u5931<\/li>\n<li>\u4e8b\u5b9e\u3001\u903b\u8f91\u3001\u7cbe\u51c6\u6027\u9519\u8bef\uff0c\u65e0\u6cd5\u4f7f\u7528\u9884\u5b9a\u4e49\u7684\u9886\u57df\u77e5\u8bc6\u7ed3\u6784\u6765\u7ea6\u675f\u6a21\u578b\u7684\u884c\u4e3a<\/li>\n<li>\u901a\u7528 RAG \u4e5f\u96be\u4ee5\u89e3\u51b3 LLM \u5e7b\u89c9\u95ee\u9898\uff0c\u5c24\u5176\u662f\u9690\u853d\u7684\u8bef\u5bfc\u6027\u4fe1\u606f<\/li>\n<li>\u4e13\u4e1a\u77e5\u8bc6\u670d\u52a1\u7684\u6311\u6218\u548c\u8981\u6c42\uff0c\u7f3a\u4e4f\u4e25\u683c\u4e14\u53ef\u63a7\u7684\u51b3\u7b56\u8fc7\u7a0b<\/li>\n<\/ul>\n<p>\u56e0\u6b64\uff0c\u8682\u8681\u56e2\u961f\u8ba4\u4e3a\uff0c\u4e00\u4e2a\u4e13\u4e1a\u7684\u77e5\u8bc6\u670d\u52a1\u6846\u67b6\uff0c\u5fc5\u987b\u5177\u5907\u4ee5\u4e0b\u51e0\u4e2a\u7279\u70b9\uff1a<\/p>\n<ul>\n<li>\u5fc5\u987b\u786e\u4fdd\u77e5\u8bc6\u7684\u51c6\u786e\u6027\uff0c\u5305\u62ec\u77e5\u8bc6\u8fb9\u754c\u7684\u5b8c\u6574\u6027\u3001\u77e5\u8bc6\u7ed3\u6784\u548c\u8bed\u4e49\u7684\u6e05\u6670\u6027\uff1b<\/li>\n<li>\u9700\u8981\u5177\u5907\u903b\u8f91\u4e25\u8c28\u6027\u3001\u65f6\u95f4\u654f\u611f\u6027\u548c\u6570\u5b57\u654f\u611f\u6027\uff1b<\/li>\n<li>\u8fd8\u9700\u8981\u5b8c\u5907\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u4ee5\u65b9\u4fbf\u5728\u77e5\u8bc6\u51b3\u7b56\u65f6\u83b7\u53d6\u5b8c\u5907\u7684\u652f\u6301\u4fe1\u606f\uff1b<\/li>\n<\/ul>\n<p>\u8682\u8681\u5b98\u65b9\u5bf9 KAG \u7684\u5b9a\u4f4d\u662f: \u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u589e\u5f3a\u670d\u52a1\u6846\u67b6, \u5177\u4f53\u9488\u5bf9\u5f53\u524d\u5927\u8bed\u8a00\u6a21\u578b\u4e0e\u77e5\u8bc6\u56fe\u8c31\u7ed3\u5408\u5bf9\u4ee5\u4e0b\u4e94\u4e2a\u65b9\u9762\u8fdb\u884c\u4e86\u589e\u5f3a<\/p>\n<ul>\n<li>\u5bf9 LLM \u53cb\u597d\u7684\u77e5\u8bc6\u8868\u793a\u589e\u5f3a<\/li>\n<li>\u77e5\u8bc6\u56fe\u8c31\u4e0e\u539f\u6587\u7247\u6bb5\u4e4b\u95f4\u7684\u4e92\u7d22\u5f15\u7ed3\u6784<\/li>\n<li>\u903b\u8f91\u7b26\u53f7\u5f15\u5bfc\u7684\u6df7\u5408\u63a8\u7406\u5f15\u64ce<\/li>\n<li>\u57fa\u4e8e\u8bed\u4e49\u63a8\u7406\u7684\u77e5\u8bc6\u5bf9\u9f50\u673a\u5236<\/li>\n<li>KAG \u6a21\u578b<\/li>\n<\/ul>\n<p>\u8fd9\u6b21\u5f00\u6e90\u5b8c\u6574\u6db5\u76d6\u4e86\u524d\u9762 4 \u9879\u6838\u5fc3\u7279\u6027\u3002<\/p>\n<p>\u56de\u5230 KAG \u547d\u540d\u7684\u95ee\u9898\u4e0a\uff0c\u4e2a\u4eba\u63a8\u6d4b\u53ef\u80fd\u8fd8\u662f\u4e3a\u4e86\u5f3a\u5316\u77e5\u8bc6\u672c\u4f53\u7684\u6982\u5ff5\u3002\u4ece\u5b98\u65b9\u63cf\u8ff0\u4ee5\u53ca\u5b9e\u9645\u4ee3\u7801\u5b9e\u73b0\u6765\u770b\uff0cKAG \u6846\u67b6\u65e0\u8bba\u5728\u6784\u5efa\u8fd8\u662f\u63a8\u7406\u9636\u6bb5\uff0c\u90fd\u5728\u4e0d\u65ad\u5f3a\u8c03\u4ece\u77e5\u8bc6\u672c\u8eab\u51fa\u53d1\uff0c\u6784\u5efa\u5b8c\u6574\u4e25\u8c28\u7684\u903b\u8f91\u94fe\u8def\uff0c\u4ee5\u5c3d\u53ef\u80fd\u6539\u5584 RAG \u6280\u672f\u6808\u7684\u4e00\u4e9b\u5df2\u77e5\u95ee\u9898\u3002<\/p>\n<h4>2.2 \u5b9e\u73b0\u65b9\u5f0f\u662f\u4ec0\u4e48(How)?<\/h4>\n<p>KAG \u6846\u67b6\u7531\u4e09\u90e8\u5206\u7ec4\u6210\uff1aKAG-Builder\u3001KAG-Solver \u548c KAG-Model\uff1a<\/p>\n<ul>\n<li>KAG-Builder \u7528\u4e8e\u79bb\u7ebf\u7d22\u5f15\uff0c\u4e3b\u8981\u5305\u62ec\u4e0a\u8ff0\u63d0\u5230\u7684\u7279\u6027 1 \u548c 2\uff1a\u77e5\u8bc6\u8868\u793a\u589e\u5f3a\u3001\u4e92\u7d22\u5f15\u7ed3\u6784\u3002<\/li>\n<li>KAG-Solver \u6a21\u5757\u6d89\u53ca\u7279\u6027 3 \u548c 4\uff1a\u903b\u8f91\u7b26\u53f7\u6df7\u5408\u63a8\u7406\u5f15\u64ce\u3001\u77e5\u8bc6\u5bf9\u9f50\u673a\u5236\u3002<\/li>\n<li>KAG-Model \u5219\u5c1d\u8bd5\u6784\u5efa\u4e86\u4e00\u4e2a\u7aef\u5230\u7aef\u7684 KAG \u6a21\u578b\u3002<\/li>\n<\/ul>\n<h3>3. \u6e90\u7801\u89e3\u6790<\/h3>\n<p>\u8fd9\u6b21\u5f00\u6e90\u4e3b\u8981\u6d89\u53ca KAG-Builder \u548c KAG-Solver \u4e24\u4e2a\u6a21\u5757\uff0c\u76f4\u63a5\u5bf9\u5e94\u6e90\u7801\u4e2d\u7684 builder \u4e0e solver \u4e24\u4e2a\u5b50\u76ee\u5f55\u3002<\/p>\n<p>\u5b9e\u9645\u7814\u8bfb\u4ee3\u7801\u8fc7\u7a0b\u4e2d\uff0c\u5efa\u8bae\u5148\u4ece\u00a0<code>examples<\/code>\u00a0\u76ee\u5f55\u5165\u624b\uff0c\u4e86\u89e3\u4e0b\u6574\u4e2a\u6846\u67b6\u7684\u8fd0\u884c\u6d41\u7a0b\uff0c\u7136\u540e\u518d\u6df1\u5165\u5230\u5177\u4f53\u6a21\u5757\u3002\u51e0\u4e2a Demo \u7684\u5165\u53e3\u6587\u4ef6\u8def\u5f84\u90fd\u5dee\u4e0d\u591a\uff0c\u6bd4\u5982\u00a0<code>kag\/examples\/medicine\/builder\/indexer.py<\/code>\u00a0\u4ee5\u53ca\u00a0<code>kag\/examples\/medicine\/solver\/evaForMedicine.py<\/code>\uff0c\u53ef\u4ee5\u6e05\u695a\u770b\u51fa builder \u7ec4\u5408\u4e86\u4e0d\u540c\u7684\u6a21\u5757\uff0c\u800c solver \u7684\u771f\u6b63\u5165\u53e3\u4f4d\u4e8e\u00a0<code>kag\/solver\/logic\/solver_pipeline.py<\/code>\u3002<\/p>\n<h4>3.1 KAG-Builder<\/h4>\n<p>\u5148\u8d34\u4e00\u4e0b\u5b8c\u6574\u76ee\u5f55\u7ed3\u6784<\/p>\n<pre><code>\u276f tree .\r\n.\r\n\u251c\u2500\u2500 __init__.py\r\n\u251c\u2500\u2500 component\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u251c\u2500\u2500 aligner\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u251c\u2500\u2500 kag_post_processor.py\r\n\u2502   \u2502   \u2514\u2500\u2500 spg_post_processor.py\r\n\u2502   \u251c\u2500\u2500 base.py\r\n\u2502   \u251c\u2500\u2500 extractor\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u251c\u2500\u2500 kag_extractor.py\r\n\u2502   \u2502   \u251c\u2500\u2500 spg_extractor.py\r\n\u2502   \u2502   \u2514\u2500\u2500 user_defined_extractor.py\r\n\u2502   \u251c\u2500\u2500 mapping\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u251c\u2500\u2500 relation_mapping.py\r\n\u2502   \u2502   \u251c\u2500\u2500 spg_type_mapping.py\r\n\u2502   \u2502   \u2514\u2500\u2500 spo_mapping.py\r\n\u2502   \u251c\u2500\u2500 reader\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u251c\u2500\u2500 csv_reader.py\r\n\u2502   \u2502   \u251c\u2500\u2500 dataset_reader.py\r\n\u2502   \u2502   \u251c\u2500\u2500 docx_reader.py\r\n\u2502   \u2502   \u251c\u2500\u2500 json_reader.py\r\n\u2502   \u2502   \u251c\u2500\u2500 markdown_reader.py\r\n\u2502   \u2502   \u251c\u2500\u2500 pdf_reader.py\r\n\u2502   \u2502   \u251c\u2500\u2500 txt_reader.py\r\n\u2502   \u2502   \u2514\u2500\u2500 yuque_reader.py\r\n\u2502   \u251c\u2500\u2500 splitter\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u251c\u2500\u2500 base_table_splitter.py\r\n\u2502   \u2502   \u251c\u2500\u2500 length_splitter.py\r\n\u2502   \u2502   \u251c\u2500\u2500 outline_splitter.py\r\n\u2502   \u2502   \u251c\u2500\u2500 pattern_splitter.py\r\n\u2502   \u2502   \u2514\u2500\u2500 semantic_splitter.py\r\n\u2502   \u251c\u2500\u2500 vectorizer\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2514\u2500\u2500 batch_vectorizer.py\r\n\u2502   \u2514\u2500\u2500 writer\r\n\u2502       \u251c\u2500\u2500 __init__.py\r\n\u2502       \u2514\u2500\u2500 kg_writer.py\r\n\u251c\u2500\u2500 default_chain.py\r\n\u251c\u2500\u2500 model\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u251c\u2500\u2500 chunk.py\r\n\u2502   \u251c\u2500\u2500 spg_record.py\r\n\u2502   \u2514\u2500\u2500 sub_graph.py\r\n\u251c\u2500\u2500 <a href=\"https:\/\/www.kdjingpai.com\/en\/openai-tuichushougel\/\">operator<\/a>\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2514\u2500\u2500 base.py\r\n\u2514\u2500\u2500 prompt\r\n\u251c\u2500\u2500 __init__.py\r\n\u251c\u2500\u2500 analyze_table_prompt.py\r\n\u251c\u2500\u2500 default\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u251c\u2500\u2500 ner.py\r\n\u2502   \u251c\u2500\u2500 std.py\r\n\u2502   \u2514\u2500\u2500 triple.py\r\n\u251c\u2500\u2500 medical\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u251c\u2500\u2500 ner.py\r\n\u2502   \u251c\u2500\u2500 std.py\r\n\u2502   \u2514\u2500\u2500 triple.py\r\n\u251c\u2500\u2500 oneke_prompt.py\r\n\u251c\u2500\u2500 outline_prompt.py\r\n\u251c\u2500\u2500 semantic_seg_prompt.py\r\n\u2514\u2500\u2500 spg_prompt.py\r\n<\/code><\/pre>\n<p>Builder \u90e8\u5206\u6d89\u53ca\u7684\u529f\u80fd\u8f83\u591a\uff0c\u8fd9\u91cc\u4ec5\u63a2\u8ba8\u4e00\u4e2a\u6bd4\u8f83\u5173\u952e\u7684\u7ec4\u4ef6\u00a0<code>KAGExtractor<\/code>\u00a0\uff0c\u5176\u57fa\u672c\u6d41\u7a0b\u56fe\u5982\u4e0b\uff1a<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-16459\" title=\"-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/639921df2227ca0.png\" alt=\"-1\" width=\"1440\" height=\"1788\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/639921df2227ca0.png 1440w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/639921df2227ca0-242x300.png 242w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/639921df2227ca0-825x1024.png 825w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/639921df2227ca0-768x954.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/639921df2227ca0-1237x1536.png 1237w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/639921df2227ca0-10x12.png 10w\" sizes=\"auto, (max-width: 1440px) 100vw, 1440px\" \/><\/p>\n<p>\u8fd9\u91cc\u4e3b\u8981\u5728\u505a\u7684\u5c31\u662f\u5229\u7528\u5927\u6a21\u578b\uff0c\u5b9e\u73b0\u4e86\u4ece\u975e\u7ed3\u6784\u5316\u6587\u672c\u5230\u7ed3\u6784\u5316\u77e5\u8bc6\u56fe\u8c31\u7684\u81ea\u52a8\u521b\u5efa\uff0c\u7b80\u5355\u63cf\u8ff0\u4e0b\u5176\u4e2d\u7684\u4e00\u4e9b\u91cd\u8981\u6b65\u9aa4\u3002<\/p>\n<ul>\n<li>\u9996\u5148\u662f\u5b9e\u4f53\u8bc6\u522b\u6a21\u5757\uff0c\u8fd9\u91cc\u4f1a\u9488\u5bf9\u9884\u5b9a\u4e49\u7684\u77e5\u8bc6\u56fe\u8c31\u7c7b\u578b\u5148\u8fdb\u884c\u7279\u5b9a\u5b9e\u4f53\u8bc6\u522b\uff0c\u7136\u540e\u8fdb\u884c\u901a\u7528\u547d\u540d\u5b9e\u4f53\u8bc6\u522b\u3002\u8fd9\u79cd\u53cc\u5c42\u8bc6\u522b\u673a\u5236\u5e94\u8be5\u80fd\u591f\u786e\u4fdd\u65e2\u80fd\u6355\u83b7\u9886\u57df\u7279\u5b9a\u7684\u5b9e\u4f53\uff0c\u53c8\u4e0d\u4f1a\u9057\u6f0f\u901a\u7528\u5b9e\u4f53\u3002<\/li>\n<li>\u56fe\u8c31\u6784\u5efa\u8fc7\u7a0b\u5b9e\u9645\u662f\u7531\u00a0<code>assemble_sub_graph_with_spg_records<\/code>\u00a0\u65b9\u6cd5\u5b8c\u6210\u7684\uff0c\u5b83\u7684\u7279\u6b8a\u4e4b\u5904\u5728\u4e8e\uff0c\u7cfb\u7edf\u4f1a\u5c06\u975e\u57fa\u672c\u7c7b\u578b\u7684\u5c5e\u6027\u8f6c\u6362\u4e3a\u56fe\u4e2d\u7684\u8282\u70b9\u548c\u8fb9\uff0c\u800c\u4e0d\u518d\u662f\u7ee7\u7eed\u4fdd\u7559\u4e3a\u5b9e\u4f53\u7684\u539f\u6709\u5c5e\u6027\u3002\u8fd9\u4e00\u70b9\u7684\u6539\u52a8\u8bf4\u5b9e\u8bdd\u6ca1\u6709\u5f88\u7406\u89e3\uff0c\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u5e94\u8be5\u662f\u7b80\u5316\u4e86\u5b9e\u4f53\u7684\u590d\u6742\u6027\u7684\uff0c\u4f46\u5b9e\u9645\u8fd8\u4e0d\u592a\u6e05\u695a\u8fd9\u79cd\u7b56\u7565\u80fd\u5e26\u6765\u591a\u5927\u7684\u6536\u76ca\uff0c\u6784\u5efa\u7684\u590d\u6742\u5ea6\u80af\u5b9a\u662f\u589e\u52a0\u4e86\u3002<\/li>\n<li>\u5b9e\u4f53\u6807\u51c6\u5316\u7531\u00a0<code>named_entity_standardization<\/code>\u00a0\u548c\u00a0<code>append_official_name<\/code>\u00a0\u4e24\u4e2a\u65b9\u6cd5\u534f\u540c\u5b8c\u6210\u3002\u9996\u5148\u5bf9\u5b9e\u4f53\u540d\u79f0\u8fdb\u884c\u89c4\u8303\u5316\u5904\u7406\uff0c\u7136\u540e\u5c06\u8fd9\u4e9b\u6807\u51c6\u5316\u7684\u540d\u79f0\u4e0e\u539f\u59cb\u5b9e\u4f53\u4fe1\u606f\u8fdb\u884c\u5173\u8054\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u611f\u89c9\u4e0a\u7c7b\u4f3c\u4e8e\u5b9e\u4f53\u53bb\u91cd\uff08Entity Resolution\uff09\u3002<\/li>\n<\/ul>\n<p>\u6574\u4f53\u6765\u8bf4\uff0cBuilder \u6a21\u5757\u7684\u529f\u80fd\u8ddf\u76ee\u524d\u5e38\u89c1\u7684\u56fe\u8c31\u6784\u5efa\u6280\u672f\u6808\u8fd8\u7b97\u6bd4\u8f83\u63a5\u8fd1\uff0c\u76f8\u5173\u6587\u7ae0\u4e0e\u4ee3\u7801\u7406\u89e3\u8d77\u6765\u96be\u5ea6\u4e5f\u4e0d\u592a\u5927\uff0c\u6b64\u5904\u4e0d\u518d\u8d58\u8ff0\u3002<\/p>\n<h4>3.2 KAG-Solver<\/h4>\n<p>Solver \u90e8\u5206\u6d89\u53ca\u5230\u6574\u4e2a\u6846\u67b6\u7684\u5f88\u591a\u6838\u5fc3\u529f\u80fd\u70b9\uff0c\u5c24\u5176\u662f\u903b\u8f91\u7b26\u53f7\u63a8\u7406\u76f8\u5173\u5185\u5bb9\uff0c\u5148\u770b\u4e0b\u6574\u4f53\u7ed3\u6784\uff1a<\/p>\n<pre><code>\u276f tree .\r\n.\r\n\u251c\u2500\u2500 __init__.py\r\n\u251c\u2500\u2500 common\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2514\u2500\u2500 base.py\r\n\u251c\u2500\u2500 implementation\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u251c\u2500\u2500 default_generator.py\r\n\u2502   \u251c\u2500\u2500 default_kg_retrieval.py\r\n\u2502   \u251c\u2500\u2500 default_lf_planner.py\r\n\u2502   \u251c\u2500\u2500 default_memory.py\r\n\u2502   \u251c\u2500\u2500 default_reasoner.py\r\n\u2502   \u251c\u2500\u2500 default_reflector.py\r\n\u2502   \u2514\u2500\u2500 lf_chunk_retriever.py\r\n\u251c\u2500\u2500 logic\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u251c\u2500\u2500 core_modules\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u251c\u2500\u2500 common\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 base_model.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 one_hop_graph.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 schema_utils.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 text_sim_by_vector.py\r\n\u2502   \u2502   \u2502   \u2514\u2500\u2500 utils.py\r\n\u2502   \u2502   \u251c\u2500\u2500 config.py\r\n\u2502   \u2502   \u251c\u2500\u2500 lf_executor.py\r\n\u2502   \u2502   \u251c\u2500\u2500 lf_generator.py\r\n\u2502   \u2502   \u251c\u2500\u2500 lf_solver.py\r\n\u2502   \u2502   \u251c\u2500\u2500 op_executor\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 op_deduce\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 deduce_executor.py\r\n\u2502   \u2502   \u2502   \u2502   \u2514\u2500\u2500 module\r\n\u2502   \u2502   \u2502   \u2502       \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2502       \u251c\u2500\u2500 choice.py\r\n\u2502   \u2502   \u2502   \u2502       \u251c\u2500\u2500 entailment.py\r\n\u2502   \u2502   \u2502   \u2502       \u251c\u2500\u2500 judgement.py\r\n\u2502   \u2502   \u2502   \u2502       \u2514\u2500\u2500 multi_choice.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 op_executor.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 op_math\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2502   \u2514\u2500\u2500 math_executor.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 op_output\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 module\r\n\u2502   \u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2502   \u2502   \u2514\u2500\u2500 get_executor.py\r\n\u2502   \u2502   \u2502   \u2502   \u2514\u2500\u2500 output_executor.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 op_retrieval\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 module\r\n\u2502   \u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 get_spo_executor.py\r\n\u2502   \u2502   \u2502   \u2502   \u2502   \u2514\u2500\u2500 search_s.py\r\n\u2502   \u2502   \u2502   \u2502   \u2514\u2500\u2500 retrieval_executor.py\r\n\u2502   \u2502   \u2502   \u2514\u2500\u2500 op_sort\r\n\u2502   \u2502   \u2502       \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502       \u2514\u2500\u2500 sort_executor.py\r\n\u2502   \u2502   \u251c\u2500\u2500 parser\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2514\u2500\u2500 logic_node_parser.py\r\n\u2502   \u2502   \u251c\u2500\u2500 retriver\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 entity_linker.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 graph_retriver\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2502   \u2502   \u251c\u2500\u2500 dsl_executor.py\r\n\u2502   \u2502   \u2502   \u2502   \u2514\u2500\u2500 dsl_model.py\r\n\u2502   \u2502   \u2502   \u251c\u2500\u2500 retrieval_spo.py\r\n\u2502   \u2502   \u2502   \u2514\u2500\u2500 schema_std.py\r\n\u2502   \u2502   \u2514\u2500\u2500 rule_runner\r\n\u2502   \u2502       \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502       \u2514\u2500\u2500 rule_runner.py\r\n\u2502   \u2514\u2500\u2500 solver_pipeline.py\r\n\u251c\u2500\u2500 main_solver.py\r\n\u251c\u2500\u2500 prompt\r\n\u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u251c\u2500\u2500 default\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u251c\u2500\u2500 deduce_choice.py\r\n\u2502   \u2502   \u251c\u2500\u2500 deduce_entail.py\r\n\u2502   \u2502   \u251c\u2500\u2500 deduce_judge.py\r\n\u2502   \u2502   \u251c\u2500\u2500 deduce_multi_choice.py\r\n\u2502   \u2502   \u251c\u2500\u2500 logic_form_plan.py\r\n\u2502   \u2502   \u251c\u2500\u2500 question_ner.py\r\n\u2502   \u2502   \u251c\u2500\u2500 resp_extractor.py\r\n\u2502   \u2502   \u251c\u2500\u2500 resp_generator.py\r\n\u2502   \u2502   \u251c\u2500\u2500 resp_judge.py\r\n\u2502   \u2502   \u251c\u2500\u2500 resp_reflector.py\r\n\u2502   \u2502   \u251c\u2500\u2500 resp_verifier.py\r\n\u2502   \u2502   \u251c\u2500\u2500 solve_question.py\r\n\u2502   \u2502   \u251c\u2500\u2500 solve_question_without_docs.py\r\n\u2502   \u2502   \u251c\u2500\u2500 solve_question_without_spo.py\r\n\u2502   \u2502   \u2514\u2500\u2500 spo_retrieval.py\r\n\u2502   \u251c\u2500\u2500 lawbench\r\n\u2502   \u2502   \u251c\u2500\u2500 __init__.py\r\n\u2502   \u2502   \u2514\u2500\u2500 logic_form_plan.py\r\n\u2502   \u2514\u2500\u2500 medical\r\n\u2502       \u251c\u2500\u2500 __init__.py\r\n\u2502       \u2514\u2500\u2500 question_ner.py\r\n\u2514\u2500\u2500 tools\r\n\u251c\u2500\u2500 __init__.py\r\n\u2514\u2500\u2500 info_processor.py\r\n<\/code><\/pre>\n<p>\u4e4b\u524d\u63d0\u5230 solver \u7684\u5165\u53e3\u6587\u4ef6\uff0c\u8fd9\u91cc\u8d34\u4e0b\u76f8\u5173\u4ee3\u7801\uff1a<\/p>\n<pre><code>class SolverPipeline:\r\ndef __init__(self, max_run=3, reflector: KagReflectorABC = None, reasoner: KagReasonerABC = None,\r\ngenerator: KAGGeneratorABC = None, **kwargs):\r\n\"\"\"\r\nInitializes the think-and-act loop class.\r\n:param max_run: Maximum number of runs to limit the thinking and acting loop, defaults to 3.\r\n:param reflector: Reflector instance for reflect tasks.\r\n:param reasoner: Reasoner instance for reasoning about tasks.\r\n:param generator: Generator instance for generating actions.\r\n\"\"\"\r\nself.max_run = max_run\r\nself.memory = DefaultMemory(**kwargs)\r\nself.reflector = reflector or DefaultReflector(**kwargs)\r\nself.reasoner = reasoner or DefaultReasoner(**kwargs)\r\nself.generator = generator or DefaultGenerator(**kwargs)\r\nself.trace_log = []\r\ndef run(self, question):\r\n\"\"\"\r\nExecutes the core logic of the problem-solving system.\r\nParameters:\r\n- question (str): The question to be answered.\r\nReturns:\r\n- tuple: answer, trace log\r\n\"\"\"\r\ninstruction = question\r\nif_finished = False\r\nlogger.debug('input instruction:{}'.format(instruction))\r\npresent_instruction = instruction\r\nrun_cnt = 0\r\nwhile not if_finished and run_cnt &lt; self.max_run:\r\nrun_cnt += 1\r\nlogger.debug('present_instruction is:{}'.format(present_instruction))\r\n# Attempt to solve the current instruction and get the answer, supporting facts, and history log\r\nsolved_answer, supporting_fact, history_log = self.reasoner.reason(present_instruction)\r\n# Extract evidence from supporting facts\r\nself.memory.save_memory(solved_answer, supporting_fact, instruction)\r\nhistory_log['present_instruction'] = present_instruction\r\nhistory_log['present_memory'] = self.memory.serialize_memory()\r\nself.trace_log.append(history_log)\r\n# Reflect the current instruction based on the current memory and instruction\r\nif_finished, present_instruction = self.reflector.reflect_query(self.memory, present_instruction)\r\nresponse = self.generator.generate(instruction, self.memory)\r\nreturn response, self.trace_log\r\n<\/code><\/pre>\n<p>\u6574\u4e2a\u00a0<code>SolverPipeline.run()<\/code>\u00a0\u65b9\u6cd5\u4e3b\u8981\u6d89\u53ca 3 \u4e2a\u6a21\u5757\uff1a<code>Reasoner<\/code>,\u00a0<code>Reflector<\/code>\u00a0\u548c\u00a0<code>Generator<\/code>\uff0c\u5176\u6574\u4f53\u903b\u8f91\u8fd8\u662f\u5f88\u6e05\u6670\u7684\uff1a\u5148\u5c1d\u8bd5\u89e3\u7b54\uff0c\u7136\u540e\u53cd\u601d\u95ee\u9898\u662f\u5426\u5df2\u5f97\u5230\u89e3\u51b3\uff0c\u5982\u679c\u6ca1\u6709\u5219\u7ee7\u7eed\u6df1\u5165\u601d\u8003\uff0c\u76f4\u5230\u5f97\u5230\u6ee1\u610f\u7684\u7b54\u6848\u6216\u8fbe\u5230\u6700\u5927\u5c1d\u8bd5\u6b21\u6570\u3002\u57fa\u672c\u7b97\u662f\u6a21\u4eff\u4eba\u7c7b\u89e3\u51b3\u590d\u6742\u95ee\u9898\u7684\u4e00\u822c\u601d\u8003\u65b9\u5f0f\u3002<\/p>\n<p>\u4ee5\u4e0b\u90e8\u5206\u5bf9\u4e0a\u8ff0 3 \u4e2a\u6a21\u5757\u8fdb\u4e00\u6b65\u5206\u6790\u3002<\/p>\n<h4>3.3 Reasoner<\/h4>\n<p>\u63a8\u7406\u6a21\u5757\u53ef\u80fd\u662f\u6574\u4e2a\u6846\u67b6\u6700\u590d\u6742\u7684\u90e8\u5206\u4e86\uff0c\u5176\u5173\u952e\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n<pre><code>class DefaultReasoner(KagReasonerABC):\r\ndef __init__(self, lf_planner: LFPlannerABC = None, lf_solver: LFSolver = None, **kwargs):\r\ndef reason(self, question: str):\r\n\"\"\"\r\nProcesses a given question by planning and executing logical forms to derive an answer.\r\nParameters:\r\n- question (str): The input question to be processed.\r\nReturns:\r\n- solved_answer: The final answer derived from solving the logical forms.\r\n- supporting_fact: Supporting facts gathered during the reasoning process.\r\n- history_log: A dictionary containing the history of QA pairs and re-ranked documents.\r\n\"\"\"\r\n# logic form planing\r\nlf_nodes: List[LFPlanResult] = self.lf_planner.lf_planing(question)\r\n# logic form execution\r\nsolved_answer, sub_qa_pair, recall_docs, history_qa_log = self.lf_solver.solve(question, lf_nodes)\r\n# Generate supporting facts for sub question-answer pair\r\nsupporting_fact = '\\n'.join(sub_qa_pair)\r\n# Retrieve and rank documents\r\nsub_querys = [lf.query for lf in lf_nodes]\r\nif self.lf_solver.chunk_retriever:\r\ndocs = self.lf_solver.chunk_retriever.rerank_docs([question] + sub_querys, recall_docs)\r\nelse:\r\nlogger.info(\"DefaultReasoner not enable chunk retriever\")\r\ndocs = []\r\nhistory_log = {\r\n'history': history_qa_log,\r\n'rerank_docs': docs\r\n}\r\nif len(docs) &gt; 0:\r\n# Append supporting facts for retrieved chunks\r\nsupporting_fact += f\"\\nPassages:{str(docs)}\"\r\nreturn solved_answer, supporting_fact, history_log\r\n<\/code><\/pre>\n<p>\u7531\u6b64\u7ed8\u5236\u51fa\u63a8\u7406\u6a21\u5757\u7684\u6574\u4f53\u6d41\u7a0b\u56fe\uff1a\uff08\u5df2\u7701\u7565\u9519\u8bef\u5904\u7406\u7b49\u903b\u8f91\uff09<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-16458\" title=\"-2\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/f191e05b3251ff9.png\" alt=\"-2\" width=\"678\" height=\"1656\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/f191e05b3251ff9.png 678w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/f191e05b3251ff9-123x300.png 123w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/f191e05b3251ff9-419x1024.png 419w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/f191e05b3251ff9-629x1536.png 629w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/f191e05b3251ff9-5x12.png 5w\" sizes=\"auto, (max-width: 678px) 100vw, 678px\" \/><\/p>\n<p>\u5bb9\u6613\u770b\u51fa\uff0c<code>DefaultReasoner.reason()<\/code>\u00a0\u65b9\u6cd5\u5927\u4f53\u5206\u4e3a\u4e09\u4e2a\u6b65\u9aa4\uff1a<\/p>\n<ul>\n<li>\u903b\u8f91\u5f62\u5f0f\u89c4\u5212 (Logic Form Planning)\uff1a\u4e3b\u8981\u6d89\u53ca\u00a0<code>LFPlanner.lf_planing<\/code><\/li>\n<li>\u903b\u8f91\u5f62\u5f0f\u6267\u884c (Logic Form Execution)\uff1a\u4e3b\u8981\u6d89\u53ca\u00a0<code>LFSolver.solve<\/code><\/li>\n<li>\u6587\u6863\u91cd\u6392\u5e8f (Document Reranking)\uff1a\u4e3b\u8981\u6d89\u53ca\u00a0<code>LFSolver.chunk_retriever.rerank_docs<\/code><\/li>\n<\/ul>\n<p>\u4ee5\u4e0b\u5206\u522b\u5bf9\u4e09\u4e2a\u6b65\u9aa4\u8fdb\u884c\u8be6\u7ec6\u5206\u6790\u3002<\/p>\n<h5>3.3.1\u00a0<strong>\u903b\u8f91\u5f62\u5f0f\u89c4\u5212 (Logic Form Planning)<\/strong><\/h5>\n<p><code>DefaultLFPlanner.lf_planing()<\/code>\u00a0\u65b9\u6cd5\u4e3b\u8981\u7528\u4e8e\u5c06\u67e5\u8be2\u5206\u89e3\u4e3a\u4e00\u7cfb\u5217\u72ec\u7acb\u7684\u903b\u8f91\u5f62\u5f0f\uff08<code>lf_nodes: List[LFPlanResult]<\/code>\uff09\u3002<\/p>\n<pre><code>lf_nodes: List[LFPlanResult] = self.lf_planner.lf_planing(question)\r\n<\/code><\/pre>\n<p>\u5176\u5177\u4f53\u5b9e\u73b0\u903b\u8f91\u53ef\u53c2\u8003\u00a0<code>kag\/solver\/implementation\/default_lf_planner.py<\/code>\uff0c\u4e3b\u8981\u662f\u9488\u5bf9\u00a0<code>llm_output<\/code>\u00a0\u505a\u6b63\u5219\u5316\u89e3\u6790\uff0c\u5982\u679c\u672a\u63d0\u4f9b\u5219\u8c03\u7528 LLM \u751f\u6210\u65b0\u7684\u903b\u8f91\u5f62\u5f0f\u3002<\/p>\n<p>\u8fd9\u91cc\u53ef\u4ee5\u5173\u6ce8\u4e0b\u00a0<code>kag\/solver\/prompt\/default\/logic_form_plan.py<\/code>\u00a0\u4e2d\u6709\u5173\u00a0<code>LogicFormPlanPrompt<\/code>\u00a0\u7684\u8be6\u7ec6\u8bbe\u8ba1\uff0c\u5176\u6838\u5fc3\u5728\u4e8e\u5982\u4f55\u5c06\u590d\u6742\u95ee\u9898\u5206\u89e3\u4e3a\u591a\u4e2a\u5b50\u67e5\u8be2\u4ee5\u53ca\u5bf9\u5e94\u7684\u903b\u8f91\u5f62\u5f0f\u3002<\/p>\n<h5>3.3.2\u00a0<strong>\u903b\u8f91\u5f62\u5f0f\u6267\u884c (Logic Form Execution)<\/strong><\/h5>\n<p><code>LFSolver.solve()<\/code>\u00a0\u65b9\u6cd5\u7528\u4e8e\u6c42\u89e3\u5177\u4f53\u903b\u8f91\u5f62\u5f0f\u95ee\u9898\uff0c\u8fd4\u56de\u7b54\u6848\u3001\u5b50\u95ee\u9898\u7b54\u6848\u5bf9\u3001\u76f8\u5173\u53ec\u56de\u6587\u6863\u548c\u5386\u53f2\u8bb0\u5f55\u7b49\u3002<\/p>\n<pre><code>solved_answer, sub_qa_pair, recall_docs, history_qa_log = self.lf_solver.solve(question, lf_nodes)\r\n<\/code><\/pre>\n<p>\u6df1\u5165<code>kag\/solver\/logic\/core_modules\/lf_solver.py<\/code>\u6e90\u7801\u90e8\u5206\uff0c\u53ef\u4ee5\u53d1\u73b0\u00a0<code>LFSolver<\/code>\u00a0\u7c7b\uff08\u903b\u8f91\u5f62\u5f0f\u6c42\u89e3\u5668\uff09\u662f\u6574\u4e2a\u63a8\u7406\u8fc7\u7a0b\u7684\u6838\u5fc3\u7c7b\uff0c\u8d1f\u8d23\u6267\u884c\u903b\u8f91\u5f62\u5f0f\uff08LF\uff09\u5e76\u751f\u6210\u7b54\u6848\uff1a<\/p>\n<ul>\n<li>\u5176\u4e3b\u8981\u65b9\u6cd5\u662f\u00a0<code>solve<\/code>\uff0c\u63a5\u6536\u4e00\u4e2a\u67e5\u8be2\u548c\u4e00\u7ec4\u903b\u8f91\u5f62\u5f0f\u8282\u70b9\uff08<code>List[LFPlanResult]<\/code>\uff09\u3002<\/li>\n<li>\u4f7f\u7528\u00a0<code>LogicExecutor<\/code>\u00a0\u6765\u6267\u884c\u903b\u8f91\u5f62\u5f0f\uff0c\u751f\u6210\u7b54\u6848\u3001\u77e5\u8bc6\u56fe\u8c31\u8def\u5f84\u548c\u5386\u53f2\u8bb0\u5f55\u3002<\/li>\n<li>\u5904\u7406\u5b50\u67e5\u8be2\u548c\u7b54\u6848\u5bf9\uff0c\u4ee5\u53ca\u76f8\u5173\u6587\u6863\u3002<\/li>\n<li>\u9519\u8bef\u5904\u7406\u548c\u56de\u9000\u7b56\u7565\uff1a\u5982\u679c\u6ca1\u6709\u627e\u5230\u7b54\u6848\u6216\u76f8\u5173\u6587\u6863\uff0c\u4f1a\u5c1d\u8bd5\u4f7f\u7528\u00a0<code>chunk_retriever<\/code>\u00a0\u53ec\u56de\u76f8\u5173\u6587\u6863\u3002<\/li>\n<\/ul>\n<p>\u5176\u4e3b\u8981\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-16460\" title=\"-3\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d4d6ed8623072ff.png\" alt=\"-3\" width=\"853\" height=\"3199\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d4d6ed8623072ff.png 853w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d4d6ed8623072ff-80x300.png 80w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d4d6ed8623072ff-273x1024.png 273w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d4d6ed8623072ff-768x2880.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d4d6ed8623072ff-546x2048.png 546w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d4d6ed8623072ff-3x12.png 3w\" sizes=\"auto, (max-width: 853px) 100vw, 853px\" \/><\/p>\n<p>\u5176\u4e2d\u00a0<code>LogicExecutor<\/code>\u00a0\u662f\u6bd4\u8f83\u5173\u952e\u7684\u4e00\u4e2a\u7c7b\uff0c\u8fd9\u91cc\u8d34\u4e00\u4e0b\u6838\u5fc3\u4ee3\u7801\uff1a<\/p>\n<pre><code>executor = LogicExecutor(\r\nquery, self.project_id, self.schema,\r\nkg_retriever=self.kg_retriever,\r\nchunk_retriever=self.chunk_retriever,\r\nstd_schema=self.std_schema,\r\nel=self.el,\r\ntext_similarity=self.text_similarity,\r\ndsl_runner=DslRunnerOnGraphStore(...),\r\ngenerator=self.generator,\r\nreport_tool=self.report_tool,\r\nreq_id=generate_random_string(10)\r\n)\r\nkg_qa_result, kg_graph, history = executor.execute(lf_nodes, query)\r\n<\/code><\/pre>\n<ol>\n<li>\u6267\u884c\u903b\u8f91<br \/>\n<code>LogicExecutor<\/code>\u00a0\u7c7b\u7684\u76f8\u5173\u4ee3\u7801\u4f4d\u4e8e\u00a0<code>kag\/solver\/logic\/core_modules\/lf_executor.py<\/code>\u3002\u5176\u00a0<code>execute<\/code>\u00a0\u65b9\u6cd5\u7684\u4e3b\u8981\u6267\u884c\u6d41\u7a0b\u5982\u4e0b\u56fe\u6240\u793a:<br \/>\n<img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-16461\" title=\"-4\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36bc068e6a8182e.png\" alt=\"-4\" width=\"1252\" height=\"2994\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36bc068e6a8182e.png 1252w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36bc068e6a8182e-125x300.png 125w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36bc068e6a8182e-428x1024.png 428w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36bc068e6a8182e-768x1837.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36bc068e6a8182e-642x1536.png 642w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36bc068e6a8182e-856x2048.png 856w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/36bc068e6a8182e-5x12.png 5w\" sizes=\"auto, (max-width: 1252px) 100vw, 1252px\" \/><br \/>\n\u8fd9\u4e2a\u6267\u884c\u6d41\u7a0b\u5c55\u793a\u4e86\u4e00\u5957\u53cc\u91cd\u68c0\u7d22\u7b56\u7565: \u4f18\u5148\u4f7f\u7528\u7ed3\u6784\u5316\u7684\u56fe\u8c31\u6570\u636e\u68c0\u7d22\u548c\u63a8\u7406\uff0c\u5f53\u56fe\u8c31\u65e0\u6cd5\u56de\u7b54\u65f6\uff0c\u56de\u9000\u5230\u975e\u7ed3\u6784\u5316\u6587\u672c\u4fe1\u606f\u68c0\u7d22\u3002<br \/>\n\u7cfb\u7edf\u9996\u5148\u5c1d\u8bd5\u901a\u8fc7\u77e5\u8bc6\u56fe\u8c31\u89e3\u7b54\u95ee\u9898\uff0c\u5bf9\u6bcf\u4e2a\u903b\u8f91\u8868\u8fbe\u5f0f\u8282\u70b9\uff0c\u901a\u8fc7\u4e0d\u540c\u7684\u6267\u884c\u5668\uff08\u6d89\u53ca\u00a0<code>deduce<\/code>\u3001<code>math<\/code>\u3001<code>sort<\/code>\u3001<code><a href=\"https:\/\/www.kdjingpai.com\/en\/retrieval\/\">retrieval<\/a><\/code>\u3001<code>output<\/code>\u00a0\u7b49\u64cd\u4f5c\uff09\u8fdb\u884c\u5904\u7406\uff0c\u68c0\u7d22\u8fc7\u7a0b\u4f1a\u6536\u96c6 SPO\uff08\u4e3b\u8c13\u5bbe\uff09\u4e09\u5143\u7ec4\uff0c\u7528\u4e8e\u540e\u7eed\u7684\u7b54\u6848\u751f\u6210\uff1b\u5f53\u56fe\u8c31\u65e0\u6cd5\u63d0\u4f9b\u6ee1\u610f\u7b54\u6848\u65f6\uff08\u8fd4\u56de&#8221;I don&#8217;t know&#8221;\uff09\uff0c\u7cfb\u7edf\u4f1a\u56de\u9000\u5230\u6587\u672c\u5757\u68c0\u7d22\uff1a\u5229\u7528\u4e4b\u524d\u83b7\u53d6\u7684\u547d\u540d\u5b9e\u4f53\uff08NER\uff09\u7ed3\u679c\u4f5c\u4e3a\u68c0\u7d22\u951a\u70b9\uff0c\u7ed3\u5408\u5386\u53f2\u95ee\u7b54\u8bb0\u5f55\u6784\u5efa\u4e0a\u4e0b\u6587\u589e\u5f3a\u7684\u67e5\u8be2\uff0c\u518d\u901a\u8fc7\u00a0<code>chunk_retriever<\/code>\u00a0\u57fa\u4e8e\u68c0\u7d22\u5f97\u5230\u7684\u6587\u6863\u91cd\u65b0\u751f\u6210\u7b54\u6848\u3002<br \/>\n\u6574\u4e2a\u8fc7\u7a0b\u53ef\u4ee5\u770b\u4f5c\u662f\u4e00\u4e2a\u4f18\u96c5\u7684\u964d\u7ea7\u7b56\u7565\uff0c\u901a\u8fc7\u7ed3\u5408\u7ed3\u6784\u5316\u7684\u77e5\u8bc6\u56fe\u8c31\u548c\u975e\u7ed3\u6784\u5316\u7684\u6587\u672c\u6570\u636e\uff0c\u8fd9\u79cd\u6df7\u5408\u68c0\u7d22\u80fd\u591f\u5728\u4fdd\u8bc1\u51c6\u786e\u6027\u7684\u540c\u65f6\uff0c\u5c3d\u53ef\u80fd\u5730\u63d0\u4f9b\u5b8c\u6574\u4e14\u4e0a\u4e0b\u6587\u8fde\u8d2f\u7684\u7b54\u6848\u3002<\/li>\n<li>\u6838\u5fc3\u7ec4\u4ef6<br \/>\n\u9664\u4e86\u4e0a\u8ff0\u5177\u4f53\u7684\u6267\u884c\u903b\u8f91\u5916\uff0c\u6ce8\u610f\u5230\u00a0<code>LogicExecutor<\/code>\u00a0\u521d\u59cb\u5316\u65f6\u9700\u8981\u4f20\u5165\u591a\u4e2a\u7ec4\u4ef6\u3002\u9650\u4e8e\u7bc7\u5e45\uff0c\u8fd9\u91cc\u4ec5\u7b80\u5355\u63cf\u8ff0\u4e0b\u5404\u7ec4\u4ef6\u7684\u6838\u5fc3\u529f\u80fd\uff0c\u5177\u4f53\u5b9e\u73b0\u53ef\u53c2\u8003\u6e90\u7801\u3002<\/p>\n<ul>\n<li>kg_retriever: \u77e5\u8bc6\u56fe\u8c31\u68c0\u7d22\u5668<br \/>\n\u53c2\u8003\u00a0<code>kag\/solver\/implementation\/default_kg_retrieval.py<\/code>\u00a0\u4e2d\u00a0<code>KGRetrieverByLlm(KGRetrieverABC)<\/code>\uff0c\u5b9e\u73b0\u4e86\u5b9e\u4f53\u4e0e\u5173\u7cfb\u7684\u68c0\u7d22\uff0c\u6d89\u53ca\u7cbe\u786e\/\u6a21\u7cca\u3001\u4e00\u8df3\u5b50\u56fe\u7b49\u591a\u79cd\u5339\u914d\u65b9\u5f0f\u3002<\/li>\n<li>chunk_retriever: \u6587\u672c\u5757\u68c0\u7d22\u5668<br \/>\n\u53c2\u8003\u00a0<code>kag\/common\/retriever\/kag_retriever.py<\/code>\u00a0\u4e2d\u00a0<code>DefaultRetriever(ChunkRetrieverABC)<\/code>\uff0c\u8fd9\u91cc\u7684\u4ee3\u7801\u503c\u5f97\u597d\u597d\u7814\u7a76\u4e00\u4e0b\uff0c\u9996\u5148\u5728 Entity \u5904\u7406\u65b9\u9762\u505a\u4e86\u6807\u51c6\u5316\u64cd\u4f5c\uff0c\u6b64\u5916\uff0c\u6b64\u5904\u7684\u68c0\u7d22\u53c2\u8003\u4e86 HippoRAG\uff0c\u91c7\u7528\u4e86 DPR (Dense Passage Retrieval) \u548c PPR (Personalized PageRank) \u76f8\u7ed3\u5408\u7684\u6df7\u5408\u68c0\u7d22\u7b56\u7565\uff0c\u540e\u7eed\u8fd8\u8fdb\u4e00\u6b65\u57fa\u4e8e DPR \u4e0e PPR \u7684 Score \u8fdb\u884c\u4e86\u878d\u5408\uff0c\u5b9e\u73b0\u4e86\u4e24\u79cd\u68c0\u7d22\u7684\u52a8\u6001\u6743\u91cd\u5206\u914d\u3002<\/li>\n<li>entity_linker (el): \u5b9e\u4f53\u94fe\u63a5\u5668<br \/>\n\u53c2\u8003\u00a0<code>kag\/solver\/logic\/core_modules\/retriver\/entity_linker.py<\/code>\u00a0\u4e2d\u00a0<code>DefaultEntityLinker(EntityLinkerBase)<\/code>\uff0c\u8fd9\u91cc\u91c7\u7528\u4e86\u5148\u6784\u5efa\u7279\u5f81\u518d\u5e76\u884c\u5316\u5904\u7406\u5b9e\u4f53\u94fe\u63a5\u7684\u601d\u8def\u3002<\/li>\n<li>dsl_runner: \u56fe\u6570\u636e\u5e93\u67e5\u8be2\u5668<br \/>\n\u53c2\u8003\u00a0<code>kag\/solver\/logic\/core_modules\/retriver\/graph_retriver\/dsl_executor.py<\/code>\u00a0\u4e2d\u00a0<code>DslRunnerOnGraphStore(DslRunner)<\/code>\uff0c\u8d1f\u8d23\u5c06\u7ed3\u6784\u5316\u7684\u67e5\u8be2\u4fe1\u606f\u8f6c\u6362\u4e3a\u5177\u4f53\u7684\u56fe\u6570\u636e\u5e93\u67e5\u8be2\u8bed\u53e5\uff0c\u8fd9\u5757\u4f1a\u6d89\u53ca\u5e95\u5c42\u5177\u4f53\u7684\u56fe\u6570\u636e\u5e93\uff0c\u7ec6\u8282\u76f8\u5bf9\u7e41\u6742\uff0c\u5c31\u4e0d\u8fc7\u591a\u6d89\u53ca\u4e86\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<p>\u901a\u8fc7\u68b3\u7406\u4e0a\u8ff0\u4ee3\u7801\u4e0e\u6d41\u7a0b\u56fe\uff0c\u53ef\u4ee5\u770b\u51fa\uff0c\u6574\u4e2a\u903b\u8f91\u5f62\u5f0f\u6267\u884c (Logic Form Execution) \u73af\u8282\u91c7\u7528\u4e86\u5206\u5c42\u5904\u7406\u67b6\u6784\uff1a<\/p>\n<ul>\n<li>\u9876\u5c42\u00a0<code>LFSolver<\/code>\u00a0\u8d1f\u8d23\u6574\u4f53\u6d41\u7a0b<\/li>\n<li>\u4e2d\u95f4\u5c42\u00a0<code>LogicExecutor<\/code>\u00a0\u8d1f\u8d23\u6267\u884c\u5177\u4f53\u903b\u8f91\u5f62\u5f0f\uff08LF\uff09<\/li>\n<li>\u5e95\u5c42\u00a0<code>DSL Runner<\/code>\u00a0\u8d1f\u8d23\u4e0e\u56fe\u6570\u636e\u5e93\u4ea4\u4e92<\/li>\n<\/ul>\n<h5>3.3.3\u00a0<strong>\u6587\u6863\u91cd\u6392\u5e8f (Document Reranking)<\/strong><\/h5>\n<p>\u5982\u679c\u542f\u7528\u4e86\u00a0<code>chunk_retriever<\/code>\uff0c\u8fd8\u4f1a\u5bf9\u53ec\u56de\u6587\u6863\u8fdb\u884c\u91cd\u6392\u5e8f\u3002<\/p>\n<pre><code>if self.lf_solver.chunk_retriever:\r\ndocs = self.lf_solver.chunk_retriever.rerank_docs(\r\n[question] + sub_querys, recall_docs\r\n)\r\n<\/code><\/pre>\n<h4>3.4 Reflector<\/h4>\n<p><code>Reflector<\/code>\u00a0\u7c7b\u4e3b\u8981\u5b9e\u73b0\u4e86\u00a0<code>_can_answer<\/code>\u00a0\u4e0e\u00a0<code>_refine_query<\/code>\u00a0\u4e24\u4e2a\u65b9\u6cd5\uff0c\u524d\u8005\u7528\u4e8e\u5224\u65ad\u662f\u5426\u53ef\u4ee5\u56de\u7b54\u95ee\u9898\uff0c\u540e\u8005\u7528\u4e8e\u4f18\u5316\u591a\u8df3\u67e5\u8be2\u7684\u4e2d\u95f4\u7ed3\u679c\uff0c\u4ee5\u5f15\u5bfc\u6700\u7ec8\u7b54\u6848\u7684\u751f\u6210\u3002<\/p>\n<p>\u76f8\u5173\u5b9e\u73b0\u53c2\u8003\u00a0<code>kag\/solver\/prompt\/default\/resp_judge.py<\/code>\u00a0\u4e0e\u00a0<code>kag\/solver\/prompt\/default\/resp_reflector.py<\/code>\u00a0\u8fd9\u4e24\u4e2a Prompt \u6587\u4ef6\u66f4\u5bb9\u6613\u7406\u89e3\u3002<\/p>\n<h4>3.5 Generator<\/h4>\n<p>\u4e3b\u8981\u662f\u00a0<code>LFGenerator<\/code>\u00a0\u7c7b\uff0c\u6839\u636e\u4e0d\u540c\u573a\u666f\uff08\u6709\u65e0\u77e5\u8bc6\u56fe\u8c31\u3001\u6709\u65e0\u6587\u6863\u7b49\uff09\u52a8\u6001\u9009\u62e9\u63d0\u793a\u8bcd\u6a21\u677f\uff0c\u5e76\u751f\u6210\u5bf9\u5e94\u95ee\u9898\u7684\u7b54\u6848\u3002<br \/>\n\u76f8\u5173\u5b9e\u73b0\u4f4d\u4e8e\u00a0<code>kag\/solver\/logic\/core_modules\/lf_generator.py<\/code>\uff0c\u4ee3\u7801\u76f8\u5bf9\u76f4\u89c2\uff0c\u4e0d\u518d\u8d58\u8ff0\u3002<\/p>\n<h3>4. \u4e00\u4e9b\u601d\u8003<\/h3>\n<p>\u8682\u8681\u8fd9\u6b21\u5f00\u6e90\u7684 KAG \u6846\u67b6\uff0c\u4e3b\u6253\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u589e\u5f3a\u670d\u52a1\uff0c\u6db5\u76d6\u4e86\u7b26\u53f7\u63a8\u7406\u3001\u77e5\u8bc6\u5bf9\u9f50\u7b49\u4e00\u7cfb\u5217\u521b\u65b0\u70b9\uff0c\u7efc\u5408\u7814\u8bfb\u4e0b\u6765\uff0c\u4e2a\u4eba\u611f\u89c9\u8be5\u6846\u67b6\u5c24\u5176\u9002\u5408\u9700\u8981 \u4e25\u683c\u7ea6\u675f Schema \u7684\u4e13\u4e1a\u77e5\u8bc6\u573a\u666f\uff0c\u65e0\u8bba\u662f\u5728\u7d22\u5f15\u8fd8\u662f\u67e5\u8be2\u9636\u6bb5\uff0c\u6574\u4e2a\u5de5\u4f5c\u6d41\u90fd\u5728\u53cd\u590d\u5f3a\u5316\u4e00\u79cd\u89c2\u70b9\uff1a\u5fc5\u987b\u4ece\u53d7\u7ea6\u675f\u7684\u77e5\u8bc6\u5e93\u51fa\u53d1\uff0c\u53bb\u6784\u5efa\u56fe\u8c31\u6216\u505a\u903b\u8f91\u63a8\u7406\u3002\u8fd9\u79cd\u601d\u8def\u5e94\u8be5\u53ef\u4ee5\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u7f13\u89e3\u5927\u6a21\u578b\u9886\u57df\u77e5\u8bc6\u7f3a\u5931\u4ee5\u53ca\u5e7b\u89c9\u7684\u95ee\u9898\u3002<\/p>\n<p>\u5fae\u8f6f\u7684 GraphRAG \u6846\u67b6\u81ea\u5f00\u6e90\u4ee5\u6765\uff0c\u793e\u533a\u5bf9\u4e8e\u77e5\u8bc6\u56fe\u8c31\u4e0e RAG \u6280\u672f\u6808\u7684\u878d\u5408\u6709\u4e86\u66f4\u591a\u7684\u601d\u8003\uff0c\u6bd4\u5982\u8fd1\u671f\u7684 LightRAG\u3001StructRAG \u7b49\u5de5\u4f5c\uff0c\u90fd\u505a\u4e86\u5f88\u591a\u6709\u76ca\u7684\u63a2\u7d22\u3002KAG \u867d\u7136\u5728\u6280\u672f\u8def\u7ebf\u4e0a\u4e0e GraphRAG \u5b58\u5728\u4e00\u5b9a\u5dee\u5f02\uff0c\u4f46\u4e5f\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u53ef\u4ee5\u770b\u4f5c\u662f GraphRAG \u5728\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u589e\u5f3a\u670d\u52a1\u65b9\u5411\u4e0a\u7684\u4e00\u6b21\u5b9e\u8df5\uff0c\u5c24\u5176\u662f\u8865\u9f50\u4e86\u77e5\u8bc6\u5bf9\u9f50\u4e0e\u63a8\u7406\u65b9\u9762\u7684\u77ed\u677f\u3002\u4ece\u8fd9\u4e2a\u89d2\u5ea6\u6765\u8bf4\uff0c\u6211\u4e2a\u4eba\u5176\u5b9e\u66f4\u613f\u610f\u79f0\u4e4b\u4e3a Knowledge constrained GraphRAG\u3002<\/p>\n<p>\u539f\u751f GraphRAG\uff0c\u4f9d\u636e\u4e0d\u540c\u793e\u533a\u8fdb\u884c\u5206\u5c42\u6458\u8981\uff0c\u4ece\u800c\u53ef\u4ee5\u56de\u7b54\u76f8\u5bf9\u62bd\u8c61\u7684 high level \u7684\u95ee\u9898\uff0c\u4e0d\u8fc7\u4e5f\u6b63\u56e0\u4e3a\u5bf9\u4e8e Query-focused summarization (QFS) \u7684\u8fc7\u5ea6\u5173\u6ce8\uff0c\u5bfc\u81f4\u8be5\u6846\u67b6\u5728\u7ec6\u9897\u7c92\u5ea6\u4e8b\u5b9e\u6027\u95ee\u9898\u4e0a\u53ef\u80fd\u8868\u73b0\u4e0d\u4f73\uff0c\u518d\u8003\u8651\u5230\u6210\u672c\u95ee\u9898\uff0c\u539f\u751f GraphRAG \u5728\u5782\u57df\u843d\u5730\u65b9\u9762\u8fd8\u5b58\u5728\u5f88\u591a\u6311\u6218\uff0c\u800c KAG \u6846\u67b6\u4ece\u56fe\u8c31\u6784\u5efa\u9636\u6bb5\u5c31\u505a\u4e86\u6bd4\u8f83\u591a\u7684\u4f18\u5316\uff0c\u6bd4\u5982\u57fa\u4e8e\u7279\u5b9a Schema \u7684 Entity \u5bf9\u9f50\u4e0e\u6807\u51c6\u5316\u64cd\u4f5c\uff0c\u5728\u67e5\u8be2\u9636\u6bb5\uff0c\u8fd8\u5f15\u5165\u4e86\u57fa\u4e8e\u7b26\u53f7\u903b\u8f91\u7684\u77e5\u8bc6\u56fe\u8c31\u63a8\u7406\u65b9\u6cd5\uff0c\u7b26\u53f7\u63a8\u7406\u867d\u7136\u5728\u56fe\u8c31\u9886\u57df\u7814\u7a76\u5df2\u7ecf\u6bd4\u8f83\u591a\u4e86\uff0c\u4e0d\u8fc7\u771f\u6b63\u5e94\u7528\u5230 RAG \u573a\u666f\u597d\u50cf\u8fd8\u4e0d\u591a\u89c1\u3002RAG \u63a8\u7406\u80fd\u529b\u7684\u5f3a\u5316\u662f\u7b14\u8005\u6bd4\u8f83\u770b\u597d\u7684\u4e00\u4e2a\u7814\u7a76\u65b9\u5411\uff0c\u524d\u6bb5\u65f6\u95f4\u5fae\u8f6f\u603b\u7ed3\u4e86 RAG \u6280\u672f\u6808\u63a8\u7406\u80fd\u529b\u7684 4 \u4e2a\u5c42\u7ea7\uff1a<\/p>\n<ul>\n<li>Level-1 Explicit Facts\uff0c\u663e\u6027\u4e8b\u5b9e<\/li>\n<li>Level-2 Implicit Facts\uff0c\u9690\u6027\u4e8b\u5b9e<\/li>\n<li>Level-3 Interpretable Rationales\uff0c\u53ef\u89e3\u91ca\uff08\u5782\u57df\uff09\u7406\u7531<\/li>\n<li>Level-4 Hidden Rationales\uff0c\u9690\u5f62\uff08\u5782\u57df\uff09\u7406\u7531<\/li>\n<\/ul>\n<p>\u76ee\u524d\u5927\u90e8\u5206 RAG \u6846\u67b6\u7684\u63a8\u7406\u80fd\u529b\u8fd8\u4ec5\u9650\u4e8e Level-1 \u5c42\u7ea7\uff0c\u4e0a\u8ff0\u7684 Level-3 \u4e0e Level-4 \u5c42\u7ea7\u5c31\u5f3a\u8c03\u4e86\u5782\u57df\u63a8\u7406\u7684\u91cd\u8981\u6027\uff0c\u5176\u96be\u70b9\u5728\u4e8e\u5927\u6a21\u578b\u5728\u5782\u57df\u77e5\u8bc6\u7684\u7f3a\u5931\uff0c\u672c\u6b21 KAG \u6846\u67b6\u5728\u67e5\u8be2\u9636\u6bb5\u5f15\u5165\u7b26\u53f7\u63a8\u7406\uff0c\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u53ef\u4ee5\u770b\u4f5c\u662f\u5bf9\u6b64\u65b9\u5411\u7684\u63a2\u7d22\uff0c\u53ef\u9884\u89c1\u7684\u662f RAG \u63a8\u7406\u65b9\u9762\u540e\u7eed\u53ef\u80fd\u4f1a\u6380\u8d77\u4e00\u6ce2\u65b0\u7684\u7814\u7a76\u70ed\u6f6e\uff0c\u6bd4\u5982\u8fdb\u4e00\u6b65\u878d\u5408\u6a21\u578b\u672c\u8eab\u7684\u63a8\u7406\u80fd\u529b\uff0c\u5982 RL \u6216\u8005 CoT \u7b49\uff0c\u73b0\u9636\u6bb5\u5df2\u6709\u7684\u4e00\u4e9b\u5c1d\u8bd5\u5de5\u4f5c\u5728\u573a\u666f\u843d\u5730\u65b9\u9762\u8fd8\u591a\u5c11\u5b58\u5728\u9650\u5236\u3002<\/p>\n<p>\u9664\u4e86\u63a8\u7406\u73af\u8282\uff0cKAG \u5728 Retrieval \u65b9\u9762\u53c2\u8003 <a href=\"https:\/\/www.kdjingpai.com\/en\/hipporag\/\">HippoRAG<\/a> \u91c7\u7528\u4e86 DPR \u4e0e PPR \u6df7\u5408\u68c0\u7d22\u7b56\u7565\uff0cPageRank \u7684\u9ad8\u6548\u4f7f\u7528\uff0c\u4e5f\u8fdb\u4e00\u6b65\u5c55\u793a\u4e86\u77e5\u8bc6\u56fe\u8c31\u76f8\u5bf9\u4f20\u7edf\u5411\u91cf\u68c0\u7d22\u7684\u4f18\u52bf\uff0c\u76f8\u4fe1\u4eca\u540e\u4f1a\u6709\u66f4\u591a\u7684\u56fe\u8c31\u68c0\u7d22\u7b97\u6cd5\u88ab\u96c6\u6210\u5230 RAG \u6280\u672f\u6808\u4e2d\u3002<\/p>\n<p>\u5f53\u7136\uff0cKAG \u6846\u67b6\u76ee\u524d\u4f30\u8ba1\u4ecd\u5904\u4e8e\u65e9\u671f\u5feb\u901f\u8fed\u4ee3\u9636\u6bb5\uff0c\u5728\u529f\u80fd\u5177\u4f53\u5b9e\u73b0\u65b9\u9762\u5e94\u8be5\u8fd8\u662f\u5b58\u5728\u4e00\u5b9a\u7684\u8ba8\u8bba\u7a7a\u95f4\uff0c\u6bd4\u5982\u73b0\u6709\u7684\u903b\u8f91\u5f62\u5f0f\u89c4\u5212 (Logic Form Planning) \u4ee5\u53ca\u903b\u8f91\u5f62\u5f0f\u6267\u884c (Logic Form Execution) \u5728\u8bbe\u8ba1\u5c42\u9762\u662f\u5426\u6709\u5b8c\u5907\u7684\u7406\u8bba\u652f\u6491\uff0c\u5728\u9762\u5bf9\u590d\u6742\u95ee\u9898\u65f6\uff0c\u662f\u5426\u4f1a\u51fa\u73b0\u5206\u89e3\u4e0d\u5145\u5206\u3001\u6267\u884c\u5931\u8d25\u7684\u60c5\u51b5\uff0c\u4e0d\u8fc7\u8fd9\u79cd\u8fb9\u754c\u754c\u5b9a\u4ee5\u53ca\u9c81\u68d2\u6027\u95ee\u9898\u901a\u5e38\u5904\u7406\u8d77\u6765\u90fd\u975e\u5e38\u56f0\u96be\uff0c\u4e5f\u9700\u8981\u5927\u91cf\u7684\u8bd5\u9519\u6210\u672c\uff0c\u5982\u679c\u6574\u4e2a\u63a8\u7406\u94fe\u8def\u8fc7\u4e8e\u590d\u6742\uff0c\u6700\u7ec8\u5931\u8d25\u7387\u53ef\u80fd\u786e\u5b9e\u4f1a\u6bd4\u8f83\u9ad8\uff0c\u6bd5\u7adf\u5404\u79cd\u964d\u7ea7\u56de\u9000\u7b56\u7565\u4e5f\u53ea\u662f\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u7f13\u89e3\u95ee\u9898\u3002\u6b64\u5916\uff0c\u7b14\u8005\u6ce8\u610f\u5230\uff0c\u6846\u67b6\u5e95\u5c42\u7684 GraphStore \u5176\u5b9e\u5df2\u7ecf\u9884\u7559\u4e86\u589e\u91cf\u66f4\u65b0\u63a5\u53e3\uff0c\u4f46\u662f\u4e0a\u5c42\u5e94\u7528\u5e76\u672a\u5c55\u793a\u51fa\u76f8\u5173\u80fd\u529b\uff0c\u8fd9\u4e00\u5757\u4e5f\u662f\u4e2a\u4eba\u4e86\u89e3\u5230\u7684 GraphRAG \u793e\u533a\u547c\u58f0\u6bd4\u8f83\u9ad8\u7684\u4e00\u4e2a\u7279\u6027\u3002<\/p>\n<p>\u7efc\u5408\u6765\u770b\uff0cKAG \u6846\u67b6\u7b97\u662f\u8fd1\u671f\u975e\u5e38\u786c\u6838\u7684\u5de5\u4f5c\uff0c\u5305\u542b\u4e86\u5927\u91cf\u521b\u65b0\u70b9\uff0c\u4ee3\u7801\u65b9\u9762\u4e5f\u786e\u5b9e\u505a\u4e86\u5f88\u591a\u7ec6\u8282\u65b9\u9762\u7684\u6253\u78e8\uff0c\u76f8\u4fe1\u5bf9\u4e8e RAG \u6280\u672f\u6808\u7684\u843d\u5730\u8fdb\u7a0b\u4f1a\u662f\u4e00\u4e2a\u91cd\u8981\u7684\u63a8\u52a8\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>KAG (Knowledge Augmented Generation) \u662f\u4e00\u4e2a\u57fa\u4e8eOpenSPG\u5f15\u64ce\u548c\u5927\u8bed\u8a00\u6a21\u578b(LLMs)\u7684\u903b\u8f91\u5f62\u5f0f\u5f15\u5bfc\u7684\u63a8\u7406\u548c\u68c0\u7d22\u6846\u67b6\u3002\u8be5\u6846\u67b6\u4e13\u95e8\u7528\u4e8e\u6784\u5efa\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u5e93\u7684\u903b\u8f91\u63a8\u7406\u548c\u4e8b\u5b9e\u95ee\u7b54\u89e3\u51b3\u65b9\u6848\uff0c\u80fd\u6709\u6548\u514b\u670d\u4f20\u7edfRAG&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61500,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[425,20,459],"tags":[230,260,243],"class_list":["post-16457","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-professional","category-tool","category-rag-project","tag-aikaiyuanxiangmu","tag-zhishitupu","tag-aizhishikuyukefu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/16457","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=16457"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/16457\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/61500"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=16457"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=16457"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=16457"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}