<!-- {% raw %} -->
# *AskIt* (pyaskit)
[](https://github.com/katsumiok/pyaskit/actions/workflows/ci.yml)
[](https://codeclimate.com/github/katsumiok/pyaskit/maintainability)
[](https://codecov.io/gh/katsumiok/pyaskit)
[](https://badge.fury.io/py/pyaskit)
[](https://arxiv.org/abs/2308.15645)
## 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](https://beta.openai.com/),[Gemini API](https://ai.google.dev/), [Claude API](https://anthropic.com/), and [COHERE API](https://cohere.ai/) or [LLama2 API](https://github.com/facebookresearch/llama). Besides Python, *AskIt* has also been implemented in TypeScript. You can access the TypeScript version, [ts-askit](https://github.com/katsumiok/ts-askit).
## Key Features
- [x] 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
```python
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
```
```python
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}]
```
- [x] Template-Based Function Definition: Define functions using a prompt template.
```python
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."
```
- [x] Code Generation: Generate functions from the unified interface.
```python
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
```
- [x] Programming by Example (PBE): Define functions using examples.
Refer to the [Programming by Example (PBE)](#programming-by-example-pbe) section for further details.
## Installation
To install *AskIt*, run this command in your terminal:
```bash
pip install pyaskit
```
or
```bash
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`:
```bash
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](https://platform.openai.com/account/api-keys).
You need to specify the model name as an environment variable `ASKIT_MODEL`:
```bash
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](https://platform.openai.com/docs/models), [Gemini API documentation](https://ai.google.dev/), [Claude API documentation](https://anthropic.com/), and [COHERE API documentation](https://cohere.ai/).
You can also find the available models in the [`models.py`](https://github.com/katsumiok/pyaskit/blob/main/pyaskit/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:
```bash
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](examples) directory.
To run the example, run this command in your terminal:
```shell
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
```python
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](#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,
```python
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:
```python
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](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:
```python
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`:
```python
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:
```python
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.
```python
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:
```python
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({` <br> `'a': t.int, `<br> `'b': t.str` <br>`})` |{'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:
```python
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:
```python
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](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](LICENSE.md) file for details
## Citation
If you use our software in your research, please cite our paper:
```bibtex
@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}
}
```
<!-- {% endraw %} -->
Raw data
{
"_id": null,
"home_page": "https://github.com/katsumiok/pyaskit",
"name": "pyaskit",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.7",
"maintainer_email": null,
"keywords": "openai, gpt, gpt-3, gpt-4, google, gemini, claude, cohere, llama, api, wrapper, framework, dsl, llm",
"author": "Katsumi Okuda",
"author_email": "okuda@csail.mit.edu",
"download_url": "https://files.pythonhosted.org/packages/aa/b8/33b58cb7a7e91a56943158e0d990acda3af2faf2f02da6e51276c1f8ae97/pyaskit-1.4.3.tar.gz",
"platform": null,
"description": "<!-- {% raw %} -->\n# *AskIt* (pyaskit)\n\n[](https://github.com/katsumiok/pyaskit/actions/workflows/ci.yml)\n[](https://codeclimate.com/github/katsumiok/pyaskit/maintainability)\n[](https://codecov.io/gh/katsumiok/pyaskit)\n[](https://badge.fury.io/py/pyaskit)\n[](https://arxiv.org/abs/2308.15645)\n\n## Introduction\n\n*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.\n\nUsing *AskIt*, you can deploy LLMs for a multitude of tasks, such as:\n\n- Natural Language Processing: including translation, paraphrasing, and sentiment analysis.\n- Problem Solving: resolving mathematical problems.\n- Code Generation: creating new codes, and more.\n\n*pyaskit* can use GPT, Gemini, Claude, COHERE, or LLama2 as a backend.\n*pyaskit* operates through the [OpenAI API](https://beta.openai.com/),[Gemini API](https://ai.google.dev/), [Claude API](https://anthropic.com/), and [COHERE API](https://cohere.ai/) or [LLama2 API](https://github.com/facebookresearch/llama). Besides Python, *AskIt* has also been implemented in TypeScript. You can access the TypeScript version, [ts-askit](https://github.com/katsumiok/ts-askit).\n\n## Key Features\n\n- [x] Type-Guided Output Control: Get a response in the specified type. \n - No need to specify the output format in the prompt\n - No need to parse the response to extract the desired output\n\n```python\nfrom pyaskit import ask\n\n# Automatically parses the response to an integer\nsum = ask(int, \"add 1 + 1\")\n# `sum` is an integer with a value of 2\n```\n\n```python\nfrom typing import TypedDict, List\nfrom pyaskit import ask\n\n# Define a typed dictionary for programming languages\nclass PL(TypedDict):\n name: str\n year_created: int\n\n# Automatically extracts structured information into a list of dictionaries\nlangs = ask(List[PL], \"List the two oldest programming languages.\")\n# `langs` holds information on the oldest programming languages in a structured format like\n# [{'name': 'Fortran', 'year_created': 1957},\n# {'name': 'Lisp', 'year_created': 1958}]\n```\n\n- [x] Template-Based Function Definition: Define functions using a prompt template.\n\n```python\nfrom pyaskit import function\n\n@function(codable=False)\ndef translate(s: str, lang: str) -> str:\n \"\"\"Translate {{s}} into {{lang}} language.\"\"\"\n\ns = translate(\"\u3053\u3093\u306b\u3061\u306f\u4e16\u754c\u3002\", \"English\")\n# `s` would be \"Hello, world.\"\n```\n\n- [x] Code Generation: Generate functions from the unified interface.\n\n```python\nfrom pyaskit import function\n\n@function(codable=True)\ndef get_html(url: str) -> str:\n \"\"\"Get the webpage from {{url}}.\"\"\"\n # When `codable` is set to True, the body of the function is automatically coded by an LLM.\n\nhtml = get_html(\"https://github.com/katsumiok/pyaskit/blob/main/README.md\")\n# `html` contains the HTML version of this README.md\n```\n\n- [x] Programming by Example (PBE): Define functions using examples.\n Refer to the [Programming by Example (PBE)](#programming-by-example-pbe) section for further details.\n\n## Installation\n\nTo install *AskIt*, run this command in your terminal:\n\n```bash\npip install pyaskit\n```\nor\n```bash\npip install git+https://github.com/katsumiok/pyaskit.git\n```\n\n### Preparation for OpenAI API, Gemini API, Claude API, or COHERE API\n\nBefore using *AskIt*, you need to set your API key as an appropriate environment variable:\n\n- OpenAI API: `OPENAI_API_KEY`\n- Gemini API: `GOOGLE_API_KEY`\n- Claude API: `ANTHROPIC_API_KEY`\n- COHERE API: `CO_API_KEY`\n- groq API: `GROQ_API_KEY`\n\nFor example, to use OpenAI API, you need to set your OpenAI API key as an environment variable `OPENAI_API_KEY`:\n```bash\nexport OPENAI_API_KEY=<your OpenAI API key>\n```\n`<your OpenAI API key>` is a string that looks like this: `sk-<your key>`.\n You can find your OpenAI API key in the [OpenAI dashboard](https://platform.openai.com/account/api-keys).\n\nYou need to specify the model name as an environment variable `ASKIT_MODEL`:\n```bash\nexport ASKIT_MODEL=<model name>\n```\n`<model name>` is the name of the model you want to use.\nThe 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](https://platform.openai.com/docs/models), [Gemini API documentation](https://ai.google.dev/), [Claude API documentation](https://anthropic.com/), and [COHERE API documentation](https://cohere.ai/).\nYou can also find the available models in the [`models.py`](https://github.com/katsumiok/pyaskit/blob/main/pyaskit/models.py) file.\n\n### Preparation for Llama 2 API (Experimental)\n\nBefore using *AskIt* with Llama 2, you need to install it. To install Llama 2, run this command in your terminal:\n```bash\npip install git+https://github.com/facebookresearch/llama.git\n```\nYou 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.\n\nWe provide an example of using *AskIt* with Llama 2 in the [examples](examples) directory.\nTo run the example, run this command in your terminal:\n\n```shell\ntorchrun --nproc_per_node 1 examples/use_llama2.py \\\n --ckpt_dir llama-2-7b-chat/ \\\n --tokenizer_path tokenizer.model \\\n --max_seq_len 512 --max_batch_size 6\n```\n\n## Getting Started\n\nHere are some basic examples to help you familiarize yourself with *AskIt*:\n\n### Hello World\n\n```python\nimport pyaskit as ai\n\ns = ai.ask(str, 'Paraphrase \"Hello World!\"')\nprint(s)\n```\n\nTo 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](#types) section. Executing this code will yield a paraphrase of the prompt, such as:\n```\nGreetings, Planet!\n```\n\n\n### Defining a Function from a Prompt Template\n\n#### With the `function` decorator\n\nThe `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,\n\n```python\nfrom pyaskit import function\n\n@function(codable=False)\ndef paraphrase(text: str) -> str:\n \"\"\"Paraphrase {{text}}\"\"\"\n\ns = paraphrase('Hello World!')\nprint(s)\n```\n\nWhere `{{text}}` represents a template parameter and corresponds to the function parameter.\n\n#### With the `define ` API\n\nThe `define` API allows for prompt parameterization using template syntax:\n\n```python\nimport pyaskit as ai\n\nparaphrase = ai.define(str, 'Paraphrase {{text}}')\ns = paraphrase(text='Hello World!')\n# s = paraphrase('Hello World!') # This is also valid\nprint(s)\n```\nIn 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!\".\n\nThe `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](examples) directory.\n\n### Generating Functions for Codable Tasks\n\nCertain 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.\n\nThe subsequent example demonstrates using *AskIt* to tackle a task necessitating network access:\n\n```python\nimport pyaskit as ai\n\nget_html = ai.define(str, 'Get the webpage from {{url}}').compile()\nhtml = get_html(url='https://csail.mit.edu')\nprint(html)\n```\nIn 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.\n\nWhile 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`:\n\n```python\nimport pyaskit as ai\n\nget_html = ai.defun(str, {\"url\": str}, 'Get the webpage from {{url}}').compile()\nhtml = get_html(url='https://csail.mit.edu')\nprint(html)\n```\nThe second argument of the `defun` API is a dictionary that maps parameter names to their types.\n\nWe can the same thing with the following code:\n```python\nfrom pyaskit import function\n\n@function(codable=True)\ndef get_html(url: str) -> str:\n \"\"\"Get the webpage from {{url}}\"\"\"\nhtml = get_html(url='https://csail.mit.edu')\nprint(html)\n```\n## Programming by Example (PBE)\n\n### Function Definition Using Examples\nLanguage 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.\n\nLet'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.\n\n```python\nfrom pyaskit import define\n\ntraining_examples = [\n {\"input\": {\"x\": \"1\", \"y\": \"0\"}, \"output\": \"1\"},\n {\"input\": {\"x\": \"1\", \"y\": \"1\"}, \"output\": \"10\"},\n {\"input\": {\"x\": \"101\", \"y\": \"11\"}, \"output\": \"1000\"},\n {\"input\": {\"x\": \"1001\", \"y\": \"110\"}, \"output\": \"1111\"},\n {\"input\": {\"x\": \"1111\", \"y\": \"1\"}, \"output\": \"10000\"},\n]\n\nadd_binary_numbers = define(str, \"Add {{x}} and {{y}}\", training_examples=training_examples)\nsum_binary = add_binary_numbers(x=\"101\", y=\"11\")\nprint(sum_binary) # Output: \"1000\"\n```\nIn 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.\n\nThe `add_binary_numbers` function, which adds two binary numbers, behaves like any regular Python function.\n\n### Testing the Generated Function\n\nYou can use the `compile` function to test the generated function using an optional list of test examples.\n\nThe following code demonstrates how to test the function defined above with new test examples:\n\n```python\ntest_examples = [\n {\"input\": {\"x\": \"0\", \"y\": \"1\"}, \"output\": \"1\"},\n {\"input\": {\"x\": \"10\", \"y\": \"0\"}, \"output\": \"10\"},\n {\"input\": {\"x\": \"110\", \"y\": \"10\"}, \"output\": \"1000\"},\n]\nf = add_binary_numbers.compile(test_examples=test_examples)\nsum_binary = f(x=\"101\", y=\"11\")\nprint(sum_binary) # Output: \"1000\"\n```\nHere, `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.\n## Specifying Types in *AskIt*\n\n*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.\n\nThe following table describes the various types supported by *AskIt*:\n\n| Type | Description | Type Example | Value Example |\n| --- | --- | --- | --- |\n| `int` | Integer | `t.int` | 123 |\n| `float` | Floating Point Number | `t.float` | 1.23 |\n| `bool` | Boolean | `t.bool` | True |\n| `str` | String | `t.str` | \"Hello World!\" |\n| `literal` | Literal | `t.literal(123)` | 123 |\n| `list` | List | `t.list(t.int)` | [1, 2, 3] |\n| `dict` | Dictionary | `t.dict({` <br> `'a': t.int, `<br> `'b': t.str` <br>`})` |{'a': 1, 'b': \"abc\"} |\n| `record`| Dictionary | `t.record(t.str, t.int)` | {'a': 1, 'b': 2} | \n| `tuple` | Tuple | `t.tuple(t.int, t.str)` | (1, \"abc\") |\n| `union` | Union (Multiple Possible Values) | `t.union(t.literal('yes'), t.literal('no'))` | \"yes\" or \"no\" |\n| | | `t.literal('yes') \\| t.literal('no')` | \"yes\" or \"no\" |\n| | | `t.literal('yes', 'no')` | \"yes\" or \"no\" |\n| `None` | None | `None` | None |\n\n\nNote that each type declaration aids *AskIt* in parsing and understanding the desired output, ensuring your LLM returns data in the precise format you require.\n\n## Prompt Template Usage\n\nThe 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.\n\nFunction 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.\n\nConsider the following example which demonstrates how to define a function, `add`, that accepts two arguments `x` and `y` and returns their sum:\n```python\nfrom pyaskit import define\nimport pyaskit.types as t\n\nadd = define(t.int, '{{x}} + {{y}}')\nprint(add(x=1, y=2)) # keyword arguments\nprint(add(1, 2)) # positional arguments\n```\nIn 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.\n\nNotably, 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:\n```python\nfrom pyaskit import define\nimport pyaskit.types as t\n\nadd = define(t.int, '{{x}} + {{y}} + {{x}} + {{z}}')\nprint(add(x=1, y=2, z=3))\nprint(add(1, 2, 3))\n```\nHere, `{{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.\n\n## Contributing\n\nSee [CONTRIBUTING.md](CONTRIBUTING.md) for details on our code of conduct, and the process for submitting pull requests to us.\n\n## License\n\nThis project is licensed under the MIT License - see the [LICENSE.md](LICENSE.md) file for details\n\n## Citation\n\nIf you use our software in your research, please cite our paper:\n\n```bibtex\n@misc{okuda2023askit,\n title={AskIt: Unified Programming Interface for Programming with Large Language Models}, \n author={Katsumi Okuda and Saman Amarasinghe},\n year={2023},\n eprint={2308.15645},\n archivePrefix={arXiv},\n primaryClass={cs.PL}\n}\n```\n<!-- {% endraw %} -->\n\n\n",
"bugtrack_url": null,
"license": "MIT",
"summary": "AskIt: Unified programming interface for programming with large language models (GPT-3.5, GPT-4, Gemini, Claude, COHERE, Llama 2)",
"version": "1.4.3",
"project_urls": {
"Homepage": "https://github.com/katsumiok/pyaskit"
},
"split_keywords": [
"openai",
" gpt",
" gpt-3",
" gpt-4",
" google",
" gemini",
" claude",
" cohere",
" llama",
" api",
" wrapper",
" framework",
" dsl",
" llm"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "d55ce8952833f97a0deb483cbcf1fb4a9e642ea082817709dcbba3cb7404a493",
"md5": "615ca0719b86f610e25459cc51fc1cac",
"sha256": "87010851fcac404ec9aacc7672b9f7917a37a01793a3198fe7dfab05ce62f701"
},
"downloads": -1,
"filename": "pyaskit-1.4.3-py3-none-any.whl",
"has_sig": false,
"md5_digest": "615ca0719b86f610e25459cc51fc1cac",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.7",
"size": 27022,
"upload_time": "2024-06-30T19:19:26",
"upload_time_iso_8601": "2024-06-30T19:19:26.766448Z",
"url": "https://files.pythonhosted.org/packages/d5/5c/e8952833f97a0deb483cbcf1fb4a9e642ea082817709dcbba3cb7404a493/pyaskit-1.4.3-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "aab833b58cb7a7e91a56943158e0d990acda3af2faf2f02da6e51276c1f8ae97",
"md5": "ca55e73984fce241082b68ae8447cbc9",
"sha256": "3da3a532c087e59a0279cf8c83d76887f33cb325986c57d5662403678643490a"
},
"downloads": -1,
"filename": "pyaskit-1.4.3.tar.gz",
"has_sig": false,
"md5_digest": "ca55e73984fce241082b68ae8447cbc9",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.7",
"size": 27593,
"upload_time": "2024-06-30T19:19:28",
"upload_time_iso_8601": "2024-06-30T19:19:28.469659Z",
"url": "https://files.pythonhosted.org/packages/aa/b8/33b58cb7a7e91a56943158e0d990acda3af2faf2f02da6e51276c1f8ae97/pyaskit-1.4.3.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-06-30 19:19:28",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "katsumiok",
"github_project": "pyaskit",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "pyaskit"
}
JSON
