Skip to content

Latest commit

 

History

History
78 lines (60 loc) · 3.03 KB

model-prompt.md

File metadata and controls

78 lines (60 loc) · 3.03 KB
Raw

Model Prompt

The Empower functions model has only two roles: "user" and "assistant." This setup ensures compatibility with various chat templates from different foundation models, such as the Llama and Mistal families. We use a special format, described below, to encapsulate different types of information, including function call responses, regular conversation responses from the assistant, regular user messages, and function responses in the user role.

Prompt Format

User Message

First message FIrst user message will include the system prompt, json encoded functions and user message. In the follow format:

To ensure the best performance, we recommend to json encode functions with indent=2 and new lines

In this environment you have access to a set of functions defined in the JSON format you can use to address user's requests, use them if needed.
Functions: [json encoded functions]

<u>[user_message]

Format of the json functions: [{name: string, description: string, parameters: object}] # object should be a valid json schema, see OpenAI format for more details.

Example:

In this environment you have access to a set of functions defined in the JSON format you can use to address user's requests, use them if needed.
Functions: [{
    "name": "get_current_weather",
    "description": "Get the current weather in a given location",
    "parameters": {
        "type": "object",
        "properties": {
            "location": {
                "type": "string",
                "description": "The city and state, e.g. San Francisco, CA",
            },
            "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
        },
        "required": ["location"],
    },
}]

<u>How's the weather in San Francisco?

Rest user messages There're two types of formats for the rest of user messages:

  • Function response: function response should be json encoded with a <r> tag in the beginning. Format of json: [{value: string, tool_call_id: string}] Example: <r>[{"value": "{\"tempeature"\:\"70 fahrenheit\"}", "tool_call_id":"get_current_weather_dsfef123"}]
  • Plain text message from user: just add <u> in the beginning, for example: 
    <u>Can you also help me find the weather in New York City?

Assistant Message

There're two types of assistant messages:

  • Function calls response: function calls response include a set of json encoded function calls with a <f> prepend in the beginning. Format of the json: [{name: string, arguments: string}] Example: 
    <f>[{
  "name": "get_current_weather",
  "arguments": "{\"location\":\"San Francisco\"}"
}]
  • Plain text conversation response: just add <c> in the beginning, for example: 
    <c>The weather in San Francisco is 70 degrees Fahrenheit.

Using empower-functions Library

To make it easier, to use, we have encapsulate the prompting logic into our empower-functions pip package. See the example at prompt.py for details.