Skip to main content

The Plugin-Only Mode

What is the plugin-only mode?​

The plugin-only mode is a restricted mode of TaskWeaver that only allows you to use plugins. Compared to the full mode, the plugin-only mode has the following restrictions:

  1. The generated code only contains the calls to the plugins. For example, the following code only calls the ascii_render plugin and does not contain any "free-form" code.

    r1=ascii_render(**{"text":"Hello"})
    r1

  2. Only the plugins with plugin_only: true in the yaml file will be loaded. For example, the following plugin will be loaded in the plugin-only mode:

    name: ascii_render
    plugin_only: true
    ...

    If this field is not specified, the default value is false. For plugins in the plugin-only mode, the argument type can only be str, int, boolean, or float. Other types such as DataFrame are not allowed. Essentially, we consider these plugins only produce "text-like" output that can be directly consumed by the LLM.

To support the plugin-only mode, we developed a new role called PluginOrchestrator. To enable the plugin-only mode, you can add the configuration "session.roles": ["planner", "code_interpreter_plugin_only"] in the project configuration file taskweaver_config.json.

info

Although the plugin-only mode is restricted, it is still useful in some scenarios. For example, you may want to use TaskWeaver to only generate the code to call a certain plugin, and you want to get the response from the plugin directly, without generating any other code for safety reasons.

How is the plugin-only mode implemented?​

The plugin-only mode is implemented based on the function calling mode of LLMs. In this mode, the LLM is trained to generate a JSON object that contains the function name and the arguments. For example, the following JSON object is generated by the LLM:

{
  "function": "ascii_render",
  "arguments": {
    "text": "Hello"
  }
}

With this JSON object, we assemble the code to call the plugin:

r1=ascii_render(**{"text":"Hello"})
r1

Then, we execute the code and get the response from the plugin. Therefore, the code is not directly generated by the LLM in the plugin-only mode.

Which models support the plugin-only mode?​

Currently, the plugin-only mode is only supported by

Likely other models that are compatible with the OpenAI models will also support the plugin-only mode.