Skip to content

AskIt: Unified programming interface for programming with LLMs (GPT-3.5, GPT-4, Gemini, Claude, Cohere, Llama 2)

License

Notifications You must be signed in to change notification settings

katsumiok/pyaskit

Repository files navigation

AskIt (pyaskit)

Python CI Maintainability codecov PyPI version arXiv

Introduction

AskIt serves as a dedicated library or domain-specific language designed to streamline the utilization of Large Language Models (LLMs) such as GPT-4, Gemini, Claude, COHERE, and LLama2. It simplifies the complexities of prompt engineering and eradicates the requirement for parsing responses from LLMs, making programming tasks smoother.

Using AskIt, you can deploy LLMs for a multitude of tasks, such as:

  • Natural Language Processing: including translation, paraphrasing, and sentiment analysis.
  • Problem Solving: resolving mathematical problems.
  • Code Generation: creating new codes, and more.

pyaskit can use GPT, Gemini, Claude, COHERE, or LLama2 as a backend. pyaskit operates through the OpenAI API,Gemini API, Claude API, and COHERE API or LLama2 API. Besides Python, AskIt has also been implemented in TypeScript. You can access the TypeScript version, ts-askit.

Key Features

  • Type-Guided Output Control: Get a response in the specified type.
    • No need to specify the output format in the prompt
    • No need to parse the response to extract the desired output
from pyaskit import ask

# Automatically parses the response to an integer
sum = ask(int, "add 1 + 1")
# `sum` is an integer with a value of 2
from typing import TypedDict, List
from pyaskit import ask

# Define a typed dictionary for programming languages
class PL(TypedDict):
    name: str
    year_created: int

# Automatically extracts structured information into a list of dictionaries
langs = ask(List[PL], "List the two oldest programming languages.")
# `langs` holds information on the oldest programming languages in a structured format like
# [{'name': 'Fortran', 'year_created': 1957},
#  {'name': 'Lisp', 'year_created': 1958}]
  • Template-Based Function Definition: Define functions using a prompt template.
from pyaskit import function

@function(codable=False)
def translate(s: str, lang: str) -> str:
    """Translate {{s}} into {{lang}} language."""

s = translate("こんにちは世界。", "English")
# `s` would be "Hello, world."
  • Code Generation: Generate functions from the unified interface.
from pyaskit import function

@function(codable=True)
def get_html(url: str) -> str:
    """Get the webpage from {{url}}."""
    # When `codable` is set to True, the body of the function is automatically coded by an LLM.

html = get_html("https://github.com/katsumiok/pyaskit/blob/main/README.md")
# `html` contains the HTML version of this README.md

Installation

To install AskIt, run this command in your terminal:

pip install pyaskit

or

pip install git+https://github.com/katsumiok/pyaskit.git

Preparation for OpenAI API, Gemini API, Claude API, or COHERE API

Before using AskIt, you need to set your API key as an appropriate environment variable:

  • OpenAI API: OPENAI_API_KEY
  • Gemini API: GOOGLE_API_KEY
  • Claude API: ANTHROPIC_API_KEY
  • COHERE API: CO_API_KEY
  • groq API: GROQ_API_KEY

For example, to use OpenAI API, you need to set your OpenAI API key as an environment variable OPENAI_API_KEY:

export OPENAI_API_KEY=<your OpenAI API key>

<your OpenAI API key> is a string that looks like this: sk-<your key>. You can find your OpenAI API key in the OpenAI dashboard.

You need to specify the model name as an environment variable ASKIT_MODEL:

export ASKIT_MODEL=<model name>

<model name> is the name of the model you want to use. The latest AskIt is tested with gpt-4, gpt-3.5-turbo-16k, gemini-pro, claude-2.1, and cohere-2.0. You can find the list of available models in the OpenAI API documentation, Gemini API documentation, Claude API documentation, and COHERE API documentation. You can also find the available models in the models.py file.

Preparation for Llama 2 API (Experimental)

Before using AskIt with Llama 2, you need to install it. To install Llama 2, run this command in your terminal:

pip install git+https://github.com/facebookresearch/llama.git

You also need to download the tokenizer model and the checkpoint of the model you want to use. Please refer to the Llama 2 documentation for further details.

We provide an example of using AskIt with Llama 2 in the examples directory. To run the example, run this command in your terminal:

torchrun --nproc_per_node 1 examples/use_llama2.py \
    --ckpt_dir llama-2-7b-chat/ \
    --tokenizer_path tokenizer.model \
    --max_seq_len 512 --max_batch_size 6

Getting Started

Here are some basic examples to help you familiarize yourself with AskIt:

Hello World

import pyaskit as ai

s = ai.ask(str, 'Paraphrase "Hello World!"')
print(s)

To utilize AskIt, start by importing the pyaskit module. The ask API, which takes two arguments - the output type and the prompt - produces the LLM's output in the designated format. In this case, the output type is str and the prompt is Paraphrase "Hello World!". A comprehensive explanation of types in AskIt is provided in the Types section. Executing this code will yield a paraphrase of the prompt, such as:

Greetings, Planet!

Defining a Function from a Prompt Template

With the function decorator

The function decorator allows defining a function with a prompt template. The parameters of a defined function can be used as parameters of a prompt template. For example,

from pyaskit import function

@function(codable=False)
def paraphrase(text: str) -> str:
    """Paraphrase {{text}}"""

s = paraphrase('Hello World!')
print(s)

Where {{text}} represents a template parameter and corresponds to the function parameter.

With the define API

The define API allows for prompt parameterization using template syntax:

import pyaskit as ai

paraphrase = ai.define(str, 'Paraphrase {{text}}')
s = paraphrase(text='Hello World!')
# s = paraphrase('Hello World!') # This is also valid
print(s)

In this instance, the define API creates a templated function that instructs the LLM to paraphrase specified text. Invoking the paraphrase function with 'Hello World!' will return a paraphrased version of this text. Running this code might output something like "Greetings, Planet!".

The define API allows for straightforward creation of custom functions to harness the capabilities of large language models for diverse tasks. Further examples can be found in the examples directory.

Generating Functions for Codable Tasks

Certain tasks, such as those requiring real-time data, external resources like network access, file access, or database access, are unsuitable for LLM execution. However, AskIt can handle these tasks by converting the prompt into a Python program in the background.

The subsequent example demonstrates using AskIt to tackle a task necessitating network access:

import pyaskit as ai

get_html = ai.define(str, 'Get the webpage from {{url}}').compile()
html = get_html(url='https://csail.mit.edu')
print(html)

In this scenario, you only need to call compile() on the function returned by the define API. The compile function transforms the prompt into a Python program and returns a function that executes this code, behaving just like a regular Python function.

While the above example does not specify the type of the parameter url, AskIt provides the defun API to do so. The following code demonstrates how to define a function in which the type of the parameter url is specified as str:

import pyaskit as ai

get_html = ai.defun(str, {"url": str}, 'Get the webpage from {{url}}').compile()
html = get_html(url='https://csail.mit.edu')
print(html)

The second argument of the defun API is a dictionary that maps parameter names to their types.

We can the same thing with the following code:

from pyaskit import function

@function(codable=True)
def get_html(url: str) -> str:
    """Get the webpage from {{url}}"""
html = get_html(url='https://csail.mit.edu')
print(html)

Programming by Example (PBE)

Function Definition Using Examples

Language Learning Models (LLMs) offer the advantage of few-shot learning, a capability that AskIt utilizes in programming tasks. AskIt enables you to solve tasks using the Programming by Example (PBE) technique, where you provide examples of the desired input and output.

Let's consider creating a function to add two binary numbers (represented as strings). This function accepts two binary numbers and returns their sum, also in binary form. The following code demonstrates defining such a function using illustrative examples.

from pyaskit import define

training_examples = [
    {"input": {"x": "1", "y": "0"}, "output": "1"},
    {"input": {"x": "1", "y": "1"}, "output": "10"},
    {"input": {"x": "101", "y": "11"}, "output": "1000"},
    {"input": {"x": "1001", "y": "110"}, "output": "1111"},
    {"input": {"x": "1111", "y": "1"}, "output": "10000"},
]

add_binary_numbers = define(str, "Add {{x}} and {{y}}", training_examples=training_examples)
sum_binary = add_binary_numbers(x="101", y="11")
print(sum_binary)  # Output: "1000"

In this example, the define API takes three arguments: the output type, the prompt, and the training examples. Each entry in the training examples list is a dictionary containing an 'input' dictionary (with variable names and values) and an 'output' representing the expected function output given the input. The define API then returns a function that accepts input variables as keyword arguments and outputs the LLM's output in the specified type.

The add_binary_numbers function, which adds two binary numbers, behaves like any regular Python function.

Testing the Generated Function

You can use the compile function to test the generated function using an optional list of test examples.

The following code demonstrates how to test the function defined above with new test examples:

test_examples = [
    {"input": {"x": "0", "y": "1"}, "output": "1"},
    {"input": {"x": "10", "y": "0"}, "output": "10"},
    {"input": {"x": "110", "y": "10"}, "output": "1000"},
]
f = add_binary_numbers.compile(test_examples=test_examples)
sum_binary = f(x="101", y="11")
print(sum_binary)  # Output: "1000"

Here, f is the generated function that operates similarly to add_binary_numbers. By comparing the output of the generated function with the expected output for each test example, AskIt ensures the generated function behaves as expected. If any discrepancy arises, AskIt re-attempts the translation. After multiple unsuccessful translation attempts, AskIt raises an exception.

Specifying Types in AskIt

AskIt offers APIs to designate the output types for Language Learning Models (LLMs). By supplying these types as the first argument to the ask and define APIs, you can manage the LLM's output format. You can also use type hints provided Python.

The following table describes the various types supported by AskIt:

Type Description Type Example Value Example
int Integer t.int 123
float Floating Point Number t.float 1.23
bool Boolean t.bool True
str String t.str "Hello World!"
literal Literal t.literal(123) 123
list List t.list(t.int) [1, 2, 3]
dict Dictionary t.dict({
 'a': t.int,
 'b': t.str
})
{'a': 1, 'b': "abc"}
record Dictionary t.record(t.str, t.int) {'a': 1, 'b': 2}
tuple Tuple t.tuple(t.int, t.str) (1, "abc")
union Union (Multiple Possible Values) t.union(t.literal('yes'), t.literal('no')) "yes" or "no"
t.literal('yes') | t.literal('no') "yes" or "no"
t.literal('yes', 'no') "yes" or "no"
None None None None

Note that each type declaration aids AskIt in parsing and understanding the desired output, ensuring your LLM returns data in the precise format you require.

Prompt Template Usage

The prompt template is a string composed of placeholders for the parameters of the function being defined. Placeholders are denoted by double curly braces {{ and }} and can only contain a variable name. This variable name is then used as a parameter in the defined function.

Function parameters can be defined in two ways: either by keyword arguments or by positional arguments. For keyword arguments, the variable name within the placeholder serves as the keyword argument's name. For positional arguments, the sequence in which placeholders appear defines the order of the positional arguments.

Consider the following example which demonstrates how to define a function, add, that accepts two arguments x and y and returns their sum:

from pyaskit import define
import pyaskit.types as t

add = define(t.int, '{{x}} + {{y}}')
print(add(x=1, y=2))  # keyword arguments
print(add(1, 2))  # positional arguments

In this case, the add function can be invoked using either keyword or positional arguments, with the sum of x and y returned as the output.

Notably, if the same variable name is used multiple times in the prompt template, subsequent uses are mapped to the initial occurrence. Observe this behavior in the following example:

from pyaskit import define
import pyaskit.types as t

add = define(t.int, '{{x}} + {{y}} + {{x}} + {{z}}')
print(add(x=1, y=2, z=3))
print(add(1, 2, 3))

Here, {{x}} appears twice in the prompt template. The second occurrence of {{x}} maps back to the first. Hence, even though {{z}} is the fourth placeholder in the template, it aligns with the third argument of the function.

Contributing

See CONTRIBUTING.md for details on our code of conduct, and the process for submitting pull requests to us.

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Citation

If you use our software in your research, please cite our paper:

@misc{okuda2023askit,
      title={AskIt: Unified Programming Interface for Programming with Large Language Models}, 
      author={Katsumi Okuda and Saman Amarasinghe},
      year={2023},
      eprint={2308.15645},
      archivePrefix={arXiv},
      primaryClass={cs.PL}
}