{ "messages": [ { "content": "# INSTRUCTION\nYou are an AI agent responsible for routing, and your job is to select the most suitable agent from the following agents based on user needs and lastest chat history.\n\n# AGENTS / BOTS LIST\n[\n {\n \"condition\": \"人类玩家进行发言。\",\n \"agent_id\": 1,\n \"agent_name\": \"人类玩家发言\"\n },\n {\n \"condition\": \"进行游戏准备。\",\n \"agent_id\": 2,\n \"agent_name\": \"游戏准备\"\n }\n]\n\n#USER REQUEST\nIf the last input from the user is \"开始\" or \"开始游戏\", select the agent with the name \"游戏准备\". \n\nOtherwise, output 0.\n\n# CONSTRAINT\nStrictly according to the above USER REQUEST, select the most suitable agent from the AGENTS LIST.\nYou must only reply the agent_id with nothing else. If you think there is no suitable agent, then output 0.\n**Note: You can only output the value of the agent_id or 0, your output should be pure numbers, do not output other content!**\n\n# PREVIOUS CONVERSATION HISTORY\nuser: 开始游戏\n", "role": "system" } ], "model": "GPT-3.5 (16K)", "model_meta": { "format": "text", "frequency_penalty": 0, "max_tokens": 600, "presence_penalty": 0, "temperature": 0, "top_p": 0 }, "tools": null }
忽略其他参数，我们关注 messages 列表。列表中只有一条消息，它的内容是：
# INSTRUCTION You are an AI agent responsible for routing, and your job is to select the most suitable agent from the following agents based on user needs and lastest chat history.
#USER REQUEST If the last input from the user is "开始" or "开始游戏", select the agent with the name "游戏准备".
Otherwise, output 0.
# CONSTRAINT Strictly according to the above USER REQUEST, select the most suitable agent from the AGENTS LIST. You must only reply the agent_id with nothing else. If you think there is no suitable agent, then output 0. **Note: You can only output the value of the agent_id or 0, your output should be pure numbers, do not output other content!**
这个对话记录干扰判断的问题会被另一个问题加剧，就是 USER REQUEST 这部分内容的位置。提示词模板里也强调了 USER REQUEST 很重要，因为这是用户的直接请求。但是这块内容被放在了整个提示词的中间位置。LLM 都有 recency bias，一般会对靠近末尾的提示词比较敏感，对放在开头的提示词关注也比较多，但放在中间的提示词是比较容易被忽略的。对话记录一长，模型压根就不会关注用户要求了啥。
{ "messages": [ { "content": "# INSTRUCTION\nYou are an AI agent responsible for routing, and your job is to select the most suitable agent from the following agents based on user needs and lastest chat history.\n\n# AGENTS / BOTS LIST\n[\n {\n \"condition\": \"人类玩家进行发言。\",\n \"agent_id\": 1,\n \"agent_name\": \"人类玩家发言\"\n }\n]\n\n#USER REQUEST\nIf the last input from the user is *exactly* \"继续\", select the agent with the name \"人类玩家发言\".\n\nOtherwise, output 0.\n\n# CONSTRAINT\nStrictly according to the above USER REQUEST, select the most suitable agent from the AGENTS LIST.\nYou must only reply the agent_id with nothing else. If you think there is no suitable agent, then output 0.\n**Note: You can only output the value of the agent_id or 0, your output should be pure numbers, do not output other content!**\n\n# PREVIOUS CONVERSATION HISTORY\nuser: 开始游戏\n", "role": "system" } ], "model": "GPT-3.5 (16K)", "model_meta": { "format": "text", "frequency_penalty": 0, "max_tokens": 600, "presence_penalty": 0, "temperature": 0, "top_p": 0 }, "tools": null }
提示词模板与之前是一样的：
# INSTRUCTION You are an AI agent responsible for routing, and your job is to select the most suitable agent from the following agents based on user needs and lastest chat history.
#USER REQUEST If the last input from the user is *exactly* "继续", select the agent with the name "人类玩家发言".
Otherwise, output 0.
# CONSTRAINT Strictly according to the above USER REQUEST, select the most suitable agent from the AGENTS LIST. You must only reply the agent_id with nothing else. If you think there is no suitable agent, then output 0. **Note: You can only output the value of the agent_id or 0, your output should be pure numbers, do not output other content!**
{ "messages": [ { "content": "# INSTRUCTION\nYou are an AI agent responsible for routing, and your job is to select the most suitable agent from the following agents based on user needs and lastest chat history.\n\n# AGENTS / BOTS LIST\n[\n {\n \"condition\": \"人类玩家进行投票。\",\n \"agent_id\": 1,\n \"agent_name\": \"人类玩家投票\"\n },\n {\n \"condition\": \"人类玩家进行发言。\",\n \"agent_id\": 2,\n \"agent_name\": \"人类玩家发言\"\n },\n {\n \"condition\": \"用户说“重新开始”或者“重新开始游戏”\",\n \"agent_id\": 3,\n \"agent_name\": \"游戏准备\"\n }\n]\n\n#USER REQUEST\nYou MUST select the agent with the name \"人类玩家投票\" unconditionally.\n\n# CONSTRAINT\nStrictly according to the above USER REQUEST, select the most suitable agent from the AGENTS LIST.\nYou must only reply the agent_id with nothing else. If you think there is no suitable agent, then output 0.\n**Note: You can only output the value of the agent_id or 0, your output should be pure numbers, do not output other content!**\n\n# PREVIOUS CONVERSATION HISTORY\nuser: 开始游戏\n\nassistant: 你被分配到的词语是：【贾宝玉】。\n \n输入“换一个”，重新生成词语；输入“继续”，进入下一个环节。\n(by agent `游戏准备`)\n\nuser: 继续\n\nassistant: 好的，现在是你发言的环节。请根据你获得的词语进行描述，尽量不要暴露你的词语的具体内容。你可以多次发言，直到你说“完成”为止。请开始你的发言。\n(by agent `人类玩家发言`)\n\nuser: 男的\n\nassistant: 好的，我已经记录你的发言。请继续描述，或者说“完成”结束发言。\n(by agent `人类玩家发言`)\n\nuser: 你没有保存！\n\nassistant: 现在已经保存了你的发言：“男的”。请继续描述，或者说“完成”结束发言。\n(by agent `人类玩家发言`)\n\nuser: 是一名小说中的人物\n\nassistant: 已经保存你的发言：“是一名小说中的人物”。请继续描述，或者说“完成”结束发言。\n(by agent `人类玩家发言`)\n\nuser: 完成\n现在是AI玩家发言时间……\n(by agent `AI玩家发言`)\n1号AI玩家说：既然前面的玩家提到了性别和小说人物，那我就从生活背景这个角度来描述。这个人物的背景是非常富裕的。\n(by agent `AI玩家发言`)\n2号AI玩家说：这位人物是一个与众不同的角色，她非常感性而且才华横溢。\n(by agent `AI玩家发言`)\n3号AI玩家说：这位人物在故事中对爱情有着特殊的追求。\n(by agent `AI玩家发言`)\nAI玩家发言完毕。\n(by agent `AI玩家发言`)\n", "role": "system" } ], "model": "GPT-4o (8K)", "model_meta": { "format": "text", "frequency_penalty": 0, "max_tokens": 600, "presence_penalty": 0, "temperature": 0, "top_p": 0 }, "tools": null }
输入给跳转模型的提示词：
# INSTRUCTION You are an AI agent responsible for routing, and your job is to select the most suitable agent from the following agents based on user needs and lastest chat history.
#USER REQUEST You MUST select the agent with the name "人类玩家投票" unconditionally.
# CONSTRAINT Strictly according to the above USER REQUEST, select the most suitable agent from the AGENTS LIST. You must only reply the agent_id with nothing else. If you think there is no suitable agent, then output 0. **Note: You can only output the value of the agent_id or 0, your output should be pure numbers, do not output other content!**
粤ICP备14082021号
免费POC，
零成本试错
