{"id":1817,"date":"2024-04-17T00:43:28","date_gmt":"2024-04-16T16:43:28","guid":{"rendered":"https:\/\/www.tchepai.com\/?p=1817"},"modified":"2024-04-17T00:43:28","modified_gmt":"2024-04-16T16:43:28","slug":"llmagent","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/llmagent\/","title":{"rendered":"LLM\u5e94\u7528\uff1aAgent\u5bf9\u8bdd(\u9644\u5de5\u5177\u8c03\u7528)\u7684\u601d\u8003"},"content":{"rendered":"

ChatGPT\u548cKimi\u7b49\u95ee\u7b54\u4ea7\u54c1\u90fd\u5728\u7528Agent\u5bf9\u8bdd\uff08\u80fd\u8c03\u7528\u4e0d\u540c\u5de5\u5177\u4e0e\u7528\u6237\u8fdb\u884c\u4ea4\u4e92\uff09\uff0c\u6bd4\u5982Kimi\u7684\u5de5\u5177\u5c31\u6709LLM\u5bf9\u8bdd\u3001\u94fe\u63a5\u5bf9\u8bdd\u3001\u6587\u4ef6\u5bf9\u8bdd\u3001\u8054\u7f51\u5bf9\u8bdd\u3002\u6bd4\u5982ChatGPT\u3001\u6587\u5fc3\u4e00\u8a00\u548c\u8baf\u98de\u661f\u706b\u7b49\uff0c\u8fd8\u6269\u5c55\u6709\u6587\u751f\u56fe\u3001\u4ee3\u7801\u7f16\u5199\u3001\u6570\u5b66\u8ba1\u7b97\u5668\u7b49\u5de5\u5177\u3002<\/p>\n

 <\/p>\n

\"LLM\u5e94\u7528\uff1aAgent\u5bf9\u8bdd(\u9644\u5de5\u5177\u8c03\u7528)\u7684\u601d\u8003-1\"<\/p>\n

 <\/p>\n

\"LLM\u5e94\u7528\uff1aAgent\u5bf9\u8bdd(\u9644\u5de5\u5177\u8c03\u7528)\u7684\u601d\u8003-2\"

ChatGPT4\u7684Agent\u5bf9\u8bdd<\/p><\/div>\n

 <\/p>\n

 <\/p>\n

\u73b0\u5728\u5b9e\u73b0Agent\u5bf9\u8bdd\u80fd\u529b\u7684\u4e3b\u6d41\u6846\u67b6\u662fReAct\uff082022\u5e74\u7531\u666e\u6797\u65af\u987f\u5927\u5b66\u548cGoogle\u63d0\u51fa\uff09\uff0cReAct[1]\u5c31\u662f\u4e00\u79cd\u63d0\u793a\u8bcd\u65b9\u6cd5\uff0c\u878d\u5408\u4e86\u601d\u8003\u548c\u884c\u52a8\u3002\u5b83\u7684\u5386\u53f2\u6f14\u53d8\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n

 <\/p>\n

\"LLM\u5e94\u7528\uff1aAgent\u5bf9\u8bdd(\u9644\u5de5\u5177\u8c03\u7528)\u7684\u601d\u8003-3\"<\/p>\n

 <\/p>\n

\u4e0a\u56fe\u4e2d3\u79cd\u65b9\u6cd5\u5206\u522b\u4e3a\uff1a<\/p>\n

 <\/p>\n

Reason Only\uff1a<\/strong>\u4f7f\u7528Chain-of-thought\u8fdb\u884c\u591a\u6b65\u601d\u8003\uff0c\u5728Prompt\u524d\u52a0\u5165\u95ee\u9898\u8f93\u5165\u524d\u52a0\u5165\u201cLet\u2019s think step by step\u201d\u7684\u63d0\u793a\u8bcd\uff0c\u5f15\u5bfc\u591a\u6b65\u601d\u8003\uff0c\u522b\u76f4\u63a5\u7ed9\u7b54\u6848\uff0c\u4f46\u7f3a\u70b9\u5f88\u660e\u663e\uff1aReason Only\u95ed\u95e8\u5149\u60f3\u4e0d\u505a\uff0c\u4e0d\u4f1a\u51fa\u53bb\u770b\u770b\u4e86\u89e3\u5916\u9762\u7684\u4e16\u754c\u518d\u601d\u8003\uff0c\u56e0\u6b64\u4f1a\u4ea7\u751f\u5e7b\u89c9\uff0c\u671d\u4ee3\u66f4\u8fed\u4e86\u90fd\u4e0d\u77e5\u9053\uff1b<\/p>\n

 <\/p>\n

Act-Only\uff1a<\/strong>\u901a\u8fc7\u5355\u6b65\u884c\u52a8Action\u83b7\u53d6\u89c2\u5bdf\u7ed3\u679cObservation\uff0c\u7f3a\u70b9\uff1a\u6ca1\u60f3\u597d\u5c31\u7acb\u9a6c\u505a\uff0c\u6700\u540e\u5f97\u5230\u7684\u7b54\u6848\u53ef\u80fd\u8ddf\u6211\u60f3\u8981\u7684\u4e0d\u4e00\u5b9a\u80fd\u4fdd\u8bc1\u4e00\u81f4\uff1b<\/p>\n

 <\/p>\n

ReAct\uff1a<\/strong>\u878d\u5408\u4e86\u601d\u8003\u548c\u884c\u52a8\uff0c\u5373\u5148\u601d\u8003\u518d\u884c\u52a8\uff0c\u628a\u884c\u52a8\u540e\u7684\u7ed3\u679c\u8fd4\u56de\u540e\u518d\u601d\u8003\u4e0b\u4e00\u6b65\u8be5\u505a\u4ec0\u4e48\uff0c\u518d\u884c\u52a8\uff0c\u4e0d\u65ad\u91cd\u590d\u8be5\u8fc7\u7a0b\uff0c\u77e5\u9053\u4ea7\u751f\u6700\u7ec8\u7b54\u6848\u3002<\/p>\n

 <\/p>\n

\u6269\u5c55\uff1a2023\u5e74\u8fd8\u51fa\u4e86\u4e00\u4e2a\u81ea\u6211\u53cd\u601d\uff08Reflexion<\/strong>\uff09\u7684\u6846\u67b6\uff0c\u52a0\u5165\u4e86\u53cd\u601d\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u8fd9\u91cc\u4e0d\u505a\u6df1\u5165\u3002\u53e6\u5916\u6709\u5174\u8da3\u7684\u670b\u53cb\u53ef\u4ee5\u770b\u4e0b\u5927\u6dd8\u5b9d\u6280\u672f\u516c\u4f17\u53f7\u7684\u300aAgent\u8c03\u7814–19\u7c7bAgent\u6846\u67b6\u5bf9\u6bd4\u300b\uff0c\u8fd9\u5757\u4e3a\u4e86\u53d1\u8bba\u6587\u771f\u7684\u662f\u8d8a\u73a9\u8d8a\u82b1\u3002<\/p>\n

 <\/p>\n

\"LLM\u5e94\u7528\uff1aAgent\u5bf9\u8bdd(\u9644\u5de5\u5177\u8c03\u7528)\u7684\u601d\u8003-4\"<\/p>\n

 <\/p>\n

\u5c55\u793aReAct\u8bba\u6587\u63d0\u5230\u7684\u4f8b\u5b50\uff0c\u80fd\u66f4\u597d\u8bf4\u660e\u4ee5\u4e0a\u903b\u8f91\u548c\u4f18\u7f3a\u70b9\uff1a<\/p>\n

 <\/p>\n

\"LLM\u5e94\u7528\uff1aAgent\u5bf9\u8bdd(\u9644\u5de5\u5177\u8c03\u7528)\u7684\u601d\u8003-5\"<\/p>\n

 <\/p>\n

ReAct\u7684Prompt\u662f\uff1a<\/p>\n

“””
\nAgent Prompt\u7684\u8f93\u5165\u53d8\u91cf\u89e3\u91ca\uff1a
\n– tools: \u5de5\u5177\u96c6\u8bf4\u660e\uff0c\u683c\u5f0f\u4e3a”{tool.name}: {tool.description}”
\n– toool_names\uff1a\u5de5\u5177\u540d\u79f0\u5217\u8868
\n– history: \u7528\u6237\u548cAgent\u7684\u5bf9\u8bdd\u5386\u53f2\uff08\u6ce8\u610fagent chat\u4e2d\u95f4\u591a\u8f6eReAct\u8fc7\u7a0b\u4e0d\u8ba1\u5165\u5bf9\u8bdd\u5386\u53f2\uff09
\n– input: \u7528\u6237\u63d0\u95ee
\n– agent_scratchpad: \u4e2d\u95f4action\u548cobservation\u7684\u8fc7\u7a0b\uff0c
\n\u4f1a\u88ab\u683c\u5f0f\u6210”\\nObservation: {observation}\\nThought:{action} “\uff0c
\n\u7136\u540e\u4f20\u5165agent_scratchpad\uff08Agent\u7684\u601d\u7ef4\u8bb0\u5f55\uff09
\n“””
\nagent_prompt = “””
\nAnswer the following questions as best you can. If it is in order, you can use some tools appropriately. You have access to the following tools:<\/p>\n

{tools}<\/p>\n

Use the following format:
\nQuestion: the input question you must answer1
\nThought: you should always think about what to do and what tools to use.
\nAction: the action to take, should be one of {toool_names}
\nAction Input: the input to the action
\nObservation: the result of the action
\n… (this Thought\/Action\/Action Input\/Observation can be repeated zero or more times)
\nThought: I now know the final answer
\nFinal Answer: the final answer to the original input question
\nBegin!<\/p>\n

history: {history}<\/p>\n

Question: {input}<\/p>\n

Thought: {agent_scratchpad}”””<\/p><\/blockquote>\n

 <\/p>\n

ReAct\u5e94\u7528\u6d41\u7a0b\u56fe\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n

\"LLM\u5e94\u7528\uff1aAgent\u5bf9\u8bdd(\u9644\u5de5\u5177\u8c03\u7528)\u7684\u601d\u8003-6\"<\/p>\n

 <\/p>\n

 <\/p>\n

\u7f51\u4e0a\u6709\u4e2a\u4f8b\u5b50\u633a\u597d\u7684\uff0c\u8fd9\u91cc\u53c2\u8003\u6765\u505a\u4e3e\u4f8b\u9610\u91ca[2]\uff1a<\/p>\n

 <\/p>\n

\u5047\u8bbe\u6211\u4eec\u6709\uff1a<\/p>\n

 <\/p>\n

\u7528\u6237\u63d0\u95ee\uff08Question\uff09<\/strong>\uff1a\u201d\u76ee\u524d\u5e02\u573a\u4e0a\u73ab\u7470\u82b1\u7684\u5e73\u5747\u4ef7\u683c\u662f\u591a\u5c11\uff1f\u5982\u679c\u6211\u5728\u6b64\u57fa\u7840\u4e0a\u52a0\u4ef715%\u5356\u51fa\uff0c\u5e94\u8be5\u5982\u4f55\u5b9a\u4ef7\uff1f\u201d<\/p>\n

 <\/p>\n

\u5de5\u5177\u5e93\uff08Tools\uff09<\/strong>\uff1a{\u2019bing web search\u2019: \u7528\u5fc5\u5e94Search\u641c\u7d22\u7f51\u7edc\u5f00\u6e90\u4fe1\u606f\u7684\u5de5\u5177\uff1b\u2018llm-math\u2019: \u7528\u5927\u6a21\u578b\u548cPython\u505a\u6570\u5b66\u8fd0\u7b97\u7684\u5de5\u5177}<\/p>\n

 <\/p>\n

\u90a3\u4e48\u7b2c\u4e00\u8f6e\u5bf9\u8bdd\u7684\u8f93\u5165\u4e3a\uff1a<\/p>\n

Answer the following questions as best you can. If it is in order, you can use some tools appropriately. You have access to the following tools:
\nbing-web-search: \u7528\u5fc5\u5e94Search\u641c\u7d22\u7f51\u7edc\u5f00\u6e90\u4fe1\u606f\u7684\u5de5\u5177
\nllm-math: \u7528\u5927\u6a21\u578b\u548cPython\u505a\u6570\u5b66\u8fd0\u7b97\u7684\u5de5\u5177
\nUse the following format:
\nQuestion: the input question you must answer1
\nThought: you should always think about what to do and what tools to use.
\nAction: the action to take, should be one of [bing-web-search, llm-math]
\nAction Input: the input to the action
\nObservation: the result of the action
\n… (this Thought\/Action\/Action Input\/Observation can be repeated zero or more times)
\nThought: I now know the final answer
\nFinal Answer: the final answer to the original input question
\nBegin!
\nhistory:
\nQuestion: \u76ee\u524d\u5e02\u573a\u4e0a\u73ab\u7470\u82b1\u7684\u5e73\u5747\u4ef7\u683c\u662f\u591a\u5c11\uff1f\u5982\u679c\u6211\u5728\u6b64\u57fa\u7840\u4e0a\u52a0\u4ef715%\u5356\u51fa\uff0c\u5e94\u8be5\u5982\u4f55\u5b9a\u4ef7\uff1f
\nThought:<\/p><\/blockquote>\n

 <\/p>\n

\u5f97\u5230\u8f93\u51fa\u540e\u89e3\u6790\uff0c\u83b7\u53d6Thought\u3001Action\u548cAction Input\uff1a<\/p>\n

Thought: \u6211\u5e94\u8be5\u4f7f\u7528\u641c\u7d22\u5de5\u5177\u6765\u67e5\u627e\u7b54\u6848\uff0c\u8fd9\u6837\u6211\u53ef\u4ee5\u5feb\u901f\u5730\u627e\u5230\u6240\u9700\u7684\u4fe1\u606f\u3002
\nAction: bing-web-search
\nAction Input: \u73ab\u7470\u82b1\u5e73\u5747\u4ef7\u683c<\/p><\/blockquote>\n

 <\/p>\n

\u8c03\u7528bing-web-search\u5de5\u5177\uff0c\u8f93\u5165”\u73ab\u7470\u82b1\u5e73\u5747\u4ef7\u683c”\uff0c\u83b7\u53d6\u8fd4\u56de\u5185\u5bb9Observation\uff1a\u201d\u6839\u636e\u7f51\u7edc\u8d44\u6599\u663e\u793a\uff0c\u7f8e\u56fd\u6bcf\u675f\u73ab\u7470\u82b1\u572880.16\u7f8e\u5143\u3002\u201d\u3002\u4e4b\u540e\u5c06\u4ee5\u4e0a\u5185\u5bb9\uff0c\u6574\u7406\u518d\u653e\u5165ReAct\u7684\u63d0\u793a\u8bcd\u6a21\u677f\u4e2d\uff0c\u5f00\u542f\u7b2c\u4e8c\u8f6e\u5bf9\u8bdd\u7684\u8f93\u5165\uff1a<\/p>\n

Answer the following questions as best you can. If it is in order, you can use some tools appropriately. You have access to the following tools:
\nbing-web-search: \u7528\u5fc5\u5e94Search\u641c\u7d22\u7f51\u7edc\u5f00\u6e90\u4fe1\u606f\u7684\u5de5\u5177
\nllm-math: \u7528\u5927\u6a21\u578b\u548cPython\u505a\u6570\u5b66\u8fd0\u7b97\u7684\u5de5\u5177
\nUse the following format:
\nQuestion: the input question you must answer1
\nThought: you should always think about what to do and what tools to use.
\nAction: the action to take, should be one of [bing-web-search, llm-math]
\nAction Input: the input to the action
\nObservation: the result of the action
\n… (this Thought\/Action\/Action Input\/Observation can be repeated zero or more times)
\nThought: I now know the final answer
\nFinal Answer: the final answer to the original input question
\nBegin!
\nhistory:
\nQuestion: \u76ee\u524d\u5e02\u573a\u4e0a\u73ab\u7470\u82b1\u7684\u5e73\u5747\u4ef7\u683c\u662f\u591a\u5c11\uff1f\u5982\u679c\u6211\u5728\u6b64\u57fa\u7840\u4e0a\u52a0\u4ef715%\u5356\u51fa\uff0c\u5e94\u8be5\u5982\u4f55\u5b9a\u4ef7\uff1f
\nThought: \u6211\u5e94\u8be5\u4f7f\u7528\u641c\u7d22\u5de5\u5177\u6765\u67e5\u627e\u7b54\u6848\uff0c\u8fd9\u6837\u6211\u53ef\u4ee5\u5feb\u901f\u5730\u627e\u5230\u6240\u9700\u7684\u4fe1\u606f\u3002
\nAction: bing-web-search
\nAction Input: \u73ab\u7470\u82b1\u5e73\u5747\u4ef7\u683c
\nObservation: \u6839\u636e\u7f51\u7edc\u8d44\u6599\u663e\u793a\uff0c\u7f8e\u56fd\u6bcf\u675f\u73ab\u7470\u82b1\u572880.16\u7f8e\u5143\u3002
\nThought:<\/p><\/blockquote>\n

 <\/p>\n

\u5f97\u5230\u8f93\u51fa\u540e\u89e3\u6790\uff0c\u83b7\u53d6Thought\u3001Action\u548cAction Input\uff1a<\/p>\n

Thought: \u6211\u9700\u8981\u6570\u5b66\u8ba1\u7b97\u5728\u6b64\u57fa\u7840\u4e0a\u52a0\u4ef715%\u7684\u4ef7\u683c\u662f\u591a\u5c11\u3002
\nAction: llm-math
\nAction Input: 80.16*1.15<\/p><\/blockquote>\n

 <\/p>\n

\u8c03\u7528llm-math\u5de5\u5177\uff0c\u8f93\u5165”80.16*1.15″\uff0c\u83b7\u53d6\u8fd4\u56de\u5185\u5bb9Observation\uff1a\u201d92.184\u201d\u3002\u4e4b\u540e\u5c06\u4ee5\u4e0a\u5185\u5bb9\uff0c\u6574\u7406\u518d\u653e\u5165ReAct\u7684\u63d0\u793a\u8bcd\u6a21\u677f\u4e2d\uff0c\u5f00\u542f\u7b2c\u4e09\u8f6e\u5bf9\u8bdd\u7684\u8f93\u5165\uff1a<\/p>\n

Answer the following questions as best you can. If it is in order, you can use some tools appropriately. You have access to the following tools:
\nbing-web-search: \u7528\u5fc5\u5e94Search\u641c\u7d22\u7f51\u7edc\u5f00\u6e90\u4fe1\u606f\u7684\u5de5\u5177
\nllm-math: \u7528\u5927\u6a21\u578b\u548cPython\u505a\u6570\u5b66\u8fd0\u7b97\u7684\u5de5\u5177
\nUse the following format:
\nQuestion: the input question you must answer1
\nThought: you should always think about what to do and what tools to use.
\nAction: the action to take, should be one of [bing-web-search, llm-math]
\nAction Input: the input to the action
\nObservation: the result of the action
\n… (this Thought\/Action\/Action Input\/Observation can be repeated zero or more times)
\nThought: I now know the final answer
\nFinal Answer: the final answer to the original input question
\nBegin!
\nhistory:
\nQuestion: \u76ee\u524d\u5e02\u573a\u4e0a\u73ab\u7470\u82b1\u7684\u5e73\u5747\u4ef7\u683c\u662f\u591a\u5c11\uff1f\u5982\u679c\u6211\u5728\u6b64\u57fa\u7840\u4e0a\u52a0\u4ef715%\u5356\u51fa\uff0c\u5e94\u8be5\u5982\u4f55\u5b9a\u4ef7\uff1f
\nThought: \u6211\u5e94\u8be5\u4f7f\u7528\u641c\u7d22\u5de5\u5177\u6765\u67e5\u627e\u7b54\u6848\uff0c\u8fd9\u6837\u6211\u53ef\u4ee5\u5feb\u901f\u5730\u627e\u5230\u6240\u9700\u7684\u4fe1\u606f\u3002
\nAction: bing-web-search
\nAction Input: \u73ab\u7470\u82b1\u5e73\u5747\u4ef7\u683c
\nObservation: \u6839\u636e\u7f51\u7edc\u8d44\u6599\u663e\u793a\uff0c\u7f8e\u56fd\u6bcf\u675f\u73ab\u7470\u82b1\u572880.16\u7f8e\u5143\u3002
\nThought: \u6211\u9700\u8981\u6570\u636e\u8ba1\u7b97\u5728\u6b64\u57fa\u7840\u4e0a\u52a0\u4ef715\u7684\u4ef7\u683c\u662f\u591a\u5c11\u3002
\nAction: llm-math
\nAction Input: 80.16*1.15
\nObservation: 92.184
\nThought:<\/p><\/blockquote>\n

 <\/p>\n

\u5f97\u5230\u8f93\u51fa\u540e\u89e3\u6790\uff0c\u83b7\u53d6Thought\u3001Action\u548cAction Input\uff1a<\/p>\n

Thought: \u6211\u77e5\u9053\u6700\u7ec8\u7b54\u6848\u4e86\u3002
\nFinal Answer: \u5982\u679c\u8981\u52a0\u4ef715%\u5356\uff0c\u5e94\u8be5\u5b9a\u4ef7\u4e3a92.184\u7f8e\u5143\u3002<\/p><\/blockquote>\n

 <\/p>\n

Langchain\u5df2\u5b9e\u73b0\u4e86ReAct\u7684\u8c03\u7528[3]\uff0c\u5177\u4f53\u7684\u9700\u8981\u4f60\u53bb\u521b\u5efaReAct Agent\u3001AgentExecutor\u548ctools\u3002\u901a\u8fc7ReAct Agent\uff0c\u6211\u4eec\u4fbf\u80fd\u5b9e\u73b0\u524d\u9762\u63d0\u5230\u7684\uff0c\u8fde\u63a5\u4e0d\u540c\u5de5\u5177\uff0c\u6839\u636e\u7528\u6237\u9700\u6c42\u548cLLM\u7684\u601d\u8003\u53bb\u8c03\u7528\u4e0d\u540c\u5de5\u5177\u63d0\u9ad8\u95ee\u7b54\u8868\u73b0\u3002\u5982\u679c\u81ea\u5df1\u91cd\u5199ReAct-based\u7684Agent Chat\uff0c\u7279\u522b\u9700\u8981\u6ce8\u610f\u4e00\u4e9b\u8fb9\u754c\u6761\u4ef6\uff0c\u6bd4\u5982\u9009\u62e9\u5de5\u5177\u4e0d\u5b58\u5728\u3001\u5de5\u5177\u8c03\u7528\u5931\u8d25\uff0c\u5de5\u5177\u8c03\u7528\u9677\u5165\u6b7b\u5faa\u73af\u7b49\u60c5\u51b5\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u4f1a\u6700\u7ec8\u5f71\u54cd\u4ea4\u4e92\u7684\u6548\u679c\u3002\u4e2a\u4eba\u5728\u5b9e\u8df5ReAct Agent Chat\u6548\u679c\u65f6\uff0c\u6709\u51e0\u4e0b\u5fc3\u5f97\uff1a<\/p>\n