A tool that integrates a variety of LLM api.

support LLM interfaces can be found in config.json

At first we need to install the dependencies.

pip install -r requirements.txt

You can init an agent object by using the following code.

from agent_use import ChatAgent
my_agent = ChatAgent()

When class ChatAgent is initialized, it will take the value of llm_name in file config.json as the default LLM api.
If you want to use other LLM-api, you can set the llm_name in file config.json to the name of the LLM api you want, or you can just set llm_name variable like this:

from agent_use import ChatAgent
my_agent = ChatAgent(llm_name='chatglm2-6b')

Make sure that the the LLM api you want to use is surpported by config.json.

After init the agent object, you can chat with your agent by using the following code.

my_agent.prompt_add('Hello')
response = my_agent.prompt_post(T=0.1, maxtokens=200, remember_flag=False)

prompt_post() supports 3 parameters:

If you want specify your agent for roleplay base a simple sentence, you can run the following code after initializing the agent object.

my_agent.init_messages_by_sentence("You are a cat. From now on, the words you say will have a cat's tone words at the end")

At most time, you may need you agent to deal with more complex tasks, so you will want use a series of dialogue to specify your agent for roleplay, in this case, use init_messages_by_json() instead of init_messages_by_sentence().

my_agent.init_messages_by_json('your_dialogue_json_file.json')

The content of the dialogue json file should have the following format, and make sure user and assistant take turns speaking.

{
  "dialogues": [
    {
      "role": "user",
      "content": "Now you are a fearch keywards generation tool based on Ubuntu, and will provide information to the user according to the following rules:\n1. Be able to extract keywords that may be helpful for searching related filenames or directory names based on the user-provided task description and return them in list form.\n2. The keywards can be a part of filename or a path, as long as it can be searched.\n3. If you cannot provide any information, just return the string '[]'.\n4. Ensure that the text you generate starts with the character '[' and ends with the character ']'."
    },
    {
      "role": "assistant",
      "content": "Okay, I understand."
    },
    {
      "role": "user",
      "content": "keywards related to the command {enter dir `tool_rsc`} are:"
    },
    {
      "role": "assistant",
      "content": "['tool_rsc']"
    },
    {
      "role": "user",
      "content": "keywards related to the command {Open my secret folder} are:"
    },
    {
      "role": "assistant",
      "content": "[]"
    },
    {
      "role": "user",
      "content": "keywards related to the command {create a new file `home.json` in my `miniconda3` installation directory} are:"
    },
    {
      "role": "assistant",
      "content": "['miniconda3']"
    }
  ]
}

• reset agent role
  • If you want your agent to reset to the state it just initialized according to the JSON file or text description (which means making the agent forget all dialogs after initialization), you can simply call the init_role() function like this.

    my_agent.init_role()

    If you don't want use init_role() so frequently, I highly recommand you use the parameter remember_flag of function prompt_post() flexibly .