PromptAdapter.md

Prompt Adapter

Core Functions

Set System Prompt as file_path or content

# Set the System Prompt for the Adapter
.set_system_prompt(file_path=, variables=)

Set User Prompt as file_path or content

# Set the User Prompt for the Adapter
.set_user_prompt(content=, variables=)

Add Few Shot examples to the Prompt, choose format_type based on your prompting strategy.

# (Optional) Add few shot examples
.add_few_shot(examples=, format_type="converse"|"append_to_user_prompt"|"append_to_system_prompt")

Adapt the Prompt

# Adapt prompt to Standardized Format
.adapt()

Save the prompt to a directory. Saves a user_prompt.txt, system_prompt.txt and optionally a few_shot.json file

# Save the prompt to the required format, requires directory path
.save("path/to/directory")

Standardized Prompt Format

{
    "user_prompt_component": {
        "variables": ["var_1", "var_2", "var_3"],
        "template": "Task: You are a...... {{var_1}}, {{var_2}}",
        "metadata": {
            "format": "text",
        }
    },
    "system_prompt_component": {
        "variables": ["var_1", "var_2", "var_3"],
        "template": "Task: You are a...... {{var_1}}, {{var_2}}",
        "metadata": {
            "format": "text",
        }
    },
    "few_shot": {
        "examples": [{
            "input": "",
            "output": ""
        }, {
            "input": "",
            "output": ""
        }],
        "format": "converse" | "append_to_user_prompt" | "append_to_system_prompt"
    }
}

Example

Inputs to the Prompt Adapter

# system_prompt.txt
'''
You are a review classifier for ABC. Provide a review....
'''

# user_prompt.txt
'''
Classify the sentiment of the following review:

{{ review }}

Respond only with: positive, neutral, or negative.
'''

Functionality of Prompt Adapter

.set_system_prompt(file_path="system_prompt.txt", variables={})

.set_user_prompt(file_path="user_prompt.txt", variables={"review"})

.adapt()

Standardized Prompt

## Standardized Prompt
{
    "user_prompt_component": {
        "variables": ["review"],
        "template": "Classify the sentiment of the following review....",
        "metadata": {
            "format": "text",
        }
    },
    "system_prompt_component": {
        "variables": [],
        "template": "You are a review classifier for ABC. Provide a review....",
        "metadata": {
            "format": "text",
        }
    }
}

Supported Few-Shot Formats

1. Converse

When Few-Shot format type is converse, we pass the examples as User/Assistant turns when running inference.

When you save the prompt adapter, we save the system_prompt.txt, user_prompt.txt, and few_shot.json in Converse format.

Example:

{
    "user_prompt_component": {
        "variables": ["review"],
        "template": "Classify the sentiment of the following review....",
        "metadata": {
            "format": "text"
        }
    },
    "system_prompt_component": {
        "variables": [],
        "template": "You are a review classifier for ABC. Provide a review....",
        "metadata": {
            "format": "text"
        }
    },
    "few_shot": {
        "examples": [{
            "input": "foo_1",
            "output": "bar_1"
        }, {
            "input": "foo_2",
            "output": "bar_2"
        }],
        "format": "converse"
    }
}

On Saving

prompt_adapter.save("optimized_prompt/")

Saved User Prompt (user_prompt.txt)

Classify the sentiment of the following review....

Saved System Prompt (system_prompt.txt)

You are a review classifier for ABC. Provide a review....

Saved Few Shot Samples (few_shot.json)

[{
    "role": "user",
    "content": [{
        "text": "foo_1"
    }]
}, {
    "role": "assistant",
    "content": [{
        "text": "bar_1"
    }]
},{
    "role": "user",
    "content": [{
        "text": "foo_2"
    }]
}, {
    "role": "assistant",
    "content": [{
        "text": "bar_2"
    }]
}]

2. Append to User Prompt / Append to System Prompt

When Few-Shot format type is append_to_user_prompt or append_to_system_prompt, we append the examples to either the user_prompt or the system prompt based on the specification

When you save the prompt adapter, we save the system_prompt.txt and user_prompt.txt and they will contain the examples

Example:

{
    "user_prompt_component": {
        "variables": ["review"],
        "template": "Classify the sentiment of the following review....",
        "metadata": {
            "format": "text"
        }
    },
    "system_prompt_component": {
        "variables": [],
        "template": "You are a review classifier for ABC. Provide a review....",
        "metadata": {
            "format": "text"
        }
    },
    "few_shot": {
        "examples": [{
            "input": "foo_1",
            "output": "bar_1"
        }, {
            "input": "foo_2",
            "output": "bar_2"
        }],
        "format": "append_to_user_prompt"
    }
}

On Saving

prompt_adapter.save("optimized_prompt/")

Saved User Prompt (user_prompt.txt)

Classify the sentiment of the following review....

**Examples**
Example 1:
Input: foo_1
Output: bar_1

Example 2:
Input: foo_2
Output: bar_2

Saved System Prompt (system_prompt.txt)

You are a review classifier for ABC. Provide a review....