TL;DR

This project is a demonstration of a simple conversational simulation environment, powered by LLM agents. It takes inspiration from ChatArena. The main objective here is to enable conversation simulations through LLM agents (e.g., the converstation can be task-oriented or just a chitchat). To accomplish this, the frameworks and concepts originally developed in ChatArena has been leveraged and tailored specifically for this conversational simulation experiment. More information can be found on GitHub.

Set Up Local Virtual Environment

  • Step 1: Install Miniconda on MacOS, see instruction here.
  • Step 2: Create a default conda env with Python 3.10: conda create --name agent_sim python=3.10 -y
  • Step 3: Activate the conda env created above: conda activate agent_sim
  • Step 4: Install required libraries in the conda env: 
    pip3 install -r requirements.txt
  • Step 5: Deactivate the conda env when you are done (note: you need activate this virtual env to work on the app) 
    conda deactivate

Set Up Open AI Secrete for Local Dev

OPENAI_API_KEY will be needed when calling the OpenAI API endpoints (DEFAULT_MODEL used in this project is “gpt-3.5-turbo”), so do recommend exporting OPENAI_API_KEY as an enviroment variable on your local dev machine.

Run the Simulation

You can play with any examples under the examples folder and run the converstation simulation in an interactive mode:

python examples/example_food_qna_zh_CN.py

In this example, “Crystal” is utilizing an AI assistant named “Nova” to brainstorm ideas for a fast and nutritious school lunch for her daughter. “Ella” serves as a “moderator” responsible for evaluating whether the conversations between Nancy and Nova adhere to the following rules for termination:

  • The conversation can conclude if Nova successfully fulfills all of Crystal’s requests, and Crystal expresses satisfaction with Nova’s assistance.
  • The conversation can also end if Nova is unable to fulfill Crystal’s requests.

If Ella believes the conversation should conclude, she will state “«««END_OF_CONVERSATION»»»”; otherwise, she will say “converstation can continue”.

Please refer to the screenshot below for an illustration of the simulation results.example_1

Once simulation ends, one can save the converstation history as a json file. See a sample json file below:

[
    {
        "agent_name": "Crystal",
        "content": "嗨，Nova！我需要一些关于为我女儿准备健康午餐的建议。她现在7岁，我希望能给她提供快速但营养丰富的选项。你有什么建议吗？",
        "turn": 0,
        "timestamp": "1694143996023456000",
        "visible_to": "all",
        "msg_type": "text"
    },
    {
        "agent_name": "Nova",
        "content": "嗨，Crystal！为了给你的女儿提供快速又营养丰富的午餐，我有几个建议。首先，你可以考虑给她提供一份均衡的三明治，里面包含蔬菜、蛋白质和全麦面包。你还可以添加一些水果或酸奶作为伴侣。另外，你可以尝试制作一些健康的小吃，例如蔬菜切片搭配花生酱或者酸奶蘸酱。如果你想让她多摄入蛋白质，可以考虑给她准备一份鸡肉或鱼肉配米饭和蔬菜。最后，记得给她提供一杯牛奶或水来保持水分补充。希望这些建议对你有帮助！如果你还有其他问题，请随时告诉我。",
        "turn": 0,
        "timestamp": "1694144002581689000",
        "visible_to": "all",
        "msg_type": "text"
    },
    {
        "agent_name": "Ella",
        "content": "continue",
        "turn": 1,
        "timestamp": "1694144025822335000",
        "visible_to": "all",
        "msg_type": "text"
    },
    {
        "agent_name": "Crystal",
        "content": "谢谢你的建议，Nova！这些都是很不错的选择。我想问一下，有没有一些简单的健康午餐食谱，可以让我在家做并带给我的女儿？我想确保她吃的东西新鲜又健康。你有什么建议吗？",
        "turn": 2,
        "timestamp": "1694144032890695000",
        "visible_to": "all",
        "msg_type": "text"
    },
    {
        "agent_name": "Nova",
        "content": "当然，我可以给你一些简单的健康午餐食谱。以下是几个建议：\n\n1. 蔬菜寿司卷：用紫菜、糙米、鳗鱼、黄瓜和胡萝卜卷起来做成寿司卷。切成小段，这是一个方便携带的午餐选择。\n\n2. 鸡肉沙拉三明治：将烤鸡胸肉切成薄片，加入蔬菜（例如生菜、番茄和黄瓜），用全麦面包做成三明治。可以配以低脂酸奶或者蔬菜片。\n\n3. 彩虹色蔬菜饼：将胡萝卜、西兰花、洋葱和玉米混合在一起，加入蛋液和面粉，煎成小饼。这是一个快速又健康的午餐选择。\n\n4. 水果沙拉：切碎一些新鲜水果，例如苹果、草莓、蓝莓和香蕉。可以加入一些坚果或者酸奶作为调味品。\n\n这些都是简单而健康的午餐选择，你可以根据你女儿的口味和喜好进行调整。希望这些建议能给你一些灵感！如果你需要更多食谱或有其他问题，请随时告诉我。",
        "turn": 2,
        "timestamp": "1694144054290083000",
        "visible_to": "all",
        "msg_type": "text"
    },
    {
        "agent_name": "Ella",
        "content": "continue",
        "turn": 3,
        "timestamp": "1694144059307035000",
        "visible_to": "all",
        "msg_type": "text"
    },
    {
        "agent_name": "Crystal",
        "content": "谢谢你的建议，Nova！这些食谱听起来很棒，我会尝试制作它们。我还想问一下，有没有一些健康的零食选项，我可以给我的女儿放进她的午餐盒里？我希望能给她一些有营养但又容易携带的零食。你有什么建议吗？",
        "turn": 4,
        "timestamp": "1694144062870150000",
        "visible_to": "all",
        "msg_type": "text"
    },
    {
        "agent_name": "Nova",
        "content": "当然，我可以给你一些健康的零食选项，适合你女儿的午餐盒。以下是几个建议：\n\n1. 蔬菜棒和酸奶蘸酱：将胡萝卜、黄瓜和芹菜切成条状，搭配低脂酸奶或者花生酱作为蘸酱。这是一个轻松携带的营养零食。\n\n2. 水果串：将各种水果（例如葡萄、蓝莓、草莓等）穿在竹签上，制作成水果串。这是一个方便吃又有趣的零食选择。\n\n3. 坚果和干果混合：将坚果（例如杏仁、核桃）和干果（例如葡萄干、无花果）混合在一起，制作成小袋子。这是一个高能量和营养丰富的零食。\n\n4. 酸奶杯：将低脂酸奶加入一些新鲜水果和谷物（例如燕麦片、坚果），放入一个密封杯中。这是一个轻松携带的健康零食。\n\n记住，在选择零食时，尽量选择低糖和低盐的选项，并注意控制食用量。希望这些建议能给你一些灵感！如果你还有其他问题，请随时告诉我。",
        "turn": 4,
        "timestamp": "1694144078840960000",
        "visible_to": "all",
        "msg_type": "text"
    },
    {
        "agent_name": "Ella",
        "content": "<<<<<<END_OF_CONVERSATION>>>>>>",
        "turn": 5,
        "timestamp": "1694144091154109000",
        "visible_to": "all",
        "msg_type": "text"
    }
]

Feel free to create a new example and try to experiment with it further! Have fun ㋡ ~