LLM Widget

Widgets in Workshop

LLM widget

Config

Most fields are compatible with OpenAI API.

Fields besides widget_id and output_name

Field's Name	JSON Type (Required/Optional)	Description	Example
temperature	number (Optional)	Controls randomness in the LLMConfig model's responses.	0.6
top_p	number (Optional)	Controls diversity via nucleus sampling in LLMConfig.	0.8
max_tokens	number (Optional)	Determines the maximum length of the model’s response in LLMConfig.	150
presence_penalty	number (Optional)	Influences the likelihood of the model to talk about new topics in LLMConfig.	0.6
frequency_penalty	number (Optional)	Controls how often the model makes use of infrequent words in LLMConfig.	0.5
memory	MemoryItem[] (Optional)	Stores information acquired across multiple rounds of dialog in LLMConfig.	[{ role: 'user', content: 'Hello World' }]
need_memory	boolean (Optional)	Determines if memory usage is required in LLMConfig.	true
system_prompt	string	Expression (Required)	Provides system prompt for the user in LLMConfig.
user_prompt	string	Expression (Required)	Provides user prompt for the system in LLMConfig.
function_name	string (Optional)	Specifies the name of the function in LLMFunctionConfig.	'get_current_weather'
function_description	string (Optional)	Provides a description of the function in LLMFunctionConfig.	'Get the current weather in a given location'
function_parameters	FunctionParameter[] (Optional)	Specifies the parameters of the function in LLMFunctionConfig.	[{name: 'city', type: 'string', description: 'The city, e.g. San Francisco, CA'}]
knowledge_base_token	string (Optional)	Specify the knowledge base the widget uses.

FunctionParameter

Field's Name	JSON Type (Required/Optional)	Description	Example
name	string (Required)	Specifies the name of the function parameter in FunctionParameter.	'city'
type	string (Required)	Specifies the type of the function parameter in FunctionParameter.	'list', 'string' or 'number'
description	string (Required)	Provides a description of the function parameter in FunctionParameter.	'The city, e.g. San Francisco, CA'

Function Call

If you want to execute LLM function calls, you may specify function_name, function_description and function_parameters fields.

Function calls can be seen as a formatting wrapper provided by LLM (currently only supports OpenAI series models), allowing parameters to be input via JSON, and outputting a serial JSON.

Reference:

• https://www.promptingguide.ai/applications/function_calling

• https://platform.openai.com/docs/api-reference/chat/create

For LLM function calls, the output of the widget is a JSON object, which includes the LLM returned results and will be automatically output to the state machine output in the form of key-value pairs.

Example

{
  "id": "prompt_widget_template",
  "initial": "home_state",
  "states": {
    "home_state": {
      "inputs": {
        "input_message": {
          "type": "IM",
          "user_input": true
        }
      },
      "tasks": [
        {
          "name": "llm_widget_example_task",
          "module_type": "AnyWidgetModule",
          "module_config": {
            "widget_id": "1744214047475109888",
            "user_prompt": "{{input_message}}", // the text inputted into prompt widget, you can get it from user input or upper state
            "system_prompt": "Act as ...", // Optional field. You can input system prompt of bot.
            "top_p": 0.5, // Optional field. Default value is 0.5
            "temperature": 0.5, // Optional field. Default value is 0.5
            "frequency_penalty": 0, // Optional field. Default value is 0
            "presence_penalty": 0, // Optional field. Default value is 0
            "output_name": "result"
          }
        }
      ],
      "render": {
        "text": "{{result}}", // it's a string produced by prompt widget.
        "buttons": [
          {
            "content": "Chat Again",
            "description": "",
            "on_click": "rerun"
          }
        ]
      },
      "transitions": {
        "rerun": "home_state",
        "CHAT": "home_state"
      }
    }
  }
}