cosmobiosis/POP

POP: A Prompt Oriented Programming library

POP: Prompt Oriented Programming

Setup

Install from PyPI:

pip install pop-llm

Make sure you have an OpenAI API key in your environment variable: OPENAI_API_KEY.

Saving the key to openai_api_key.txt is recommended if you are using Jupyter Notebook:

echo $OPENAI_API_KEY > openai_api_key.txt

Usage

Getting Started

You can define a "function" using natural language. Simply define a PopFunction with three required parameters:

from pop_llm import PopFunction

# Specify the input and output parameters, and a description
add = PopFunction(["a", "b"], ["sum"], "Add the two numbers")

Then you can call the function using

answer = add(1, 2)["sum"]

If you want to view the raw API response, you can use

raw_answer = add(1, 2)["raw"]

Optionally, you can specify these parameters:

add = PopFunction(
    input_keys=["a", "b"], # The names of the input parameters
    output_keys=["sum"], # The names of the output parameters
    description="Add two numbers.", # The description of the task
    name = "add", # A name for the function, default to "function"
    input_assert=lambda a, b: type(a) == int and type(b) == int, # A function to assert the input
    temperature=0 # Set to 0 for deterministic, 1 for the most randomness
)

Allowing LLM to use external tools

You can provide a list of external tools. Simply define some well documented functions and pass them to PopFunction.

The LLM Agent will automatically use these tools to help it complete the task.

For example, you have a secret function:

def secret_function(a: int, b: int) -> int:
    """
    A secret function
    Args:
        a: The first number
        b: The second number
    Returns:
        int: The result of the secret function
    """
    return a * 2 + b - 1

secret_function_runner = PopFunction(
    input_keys=["a", "b"],
    output_keys=["output"],
    description="Return the result of the secret function on input a and b.",
    name = "secret_function",
    input_assert=lambda a, b: type(a) == int and type(b) == int,
    temperature=0,
    tools=[secret_function]
)

Be creative when defining your tools. This can be a web API, a database call, reading a file, etc.

Note: Your function input should be able to be parsed from JSON. The output should be able to be converted to string without loosing information. (i.e. str(output) is not something like <object at 0x7f9c2c0b4a90>)

For more examples, please refer to the Demo Notebook