Home

Awesome

<img width="326" alt="Group 311 (2)" src="https://github.com/bananaml/fructose/assets/44653944/8162425c-a485-460f-b816-bcc6be5d2cef">

LLM calls as strongly-typed functions

Fructose is a python package to create a dependable, strongly-typed interface around an LLM call.

Just slap the @ai decorator on a type-annotated function and call it as you would a function. It's lightweight, syntactic sugar.

from fructose import Fructose
ai = Fructose()

@ai
def describe(animals: list[str]) -> str:
  """
  Given a list of animals, use one word that'd describe them all.
  """
  ...

description = describe(["dog", "cat", "parrot", "goldfish"])
print(description) # -> "pets" type: str

The @ai decorator introspects the function and builds a prompt to an LLM to perform the task whenever the function is invoked.

Fructose supports:

Under maintenance pause

We're beyond proud of the excitement for this project following our Hacker News launch, but we've had to direct attention elsewhere. We've paused maintenance on Fructose. If you're looking for a more actively maintained package, check out Instructor or Marvin. If you love Fructose and would like to build on it yourself, you're more than welcome to fork it for your own version, or open an issue requesting to be a maintainer.

Installation

pip3 install fructose

It currently executes the prompt with OpenAI, so you'll need to use your own OpenAI API Key

export OPENAI_API_KEY=sk-abcdefghijklmnopqrstuvwxyz

Features

Complex DataTypes

from fructose import Fructose
from dataclasses import dataclass

ai = Fructose()

@dataclass
class Person:
    name: str
    hobbies: str
    dislikes: str
    obscure_inclinations: str
    age: int
    height: float
    is_human: bool

@ai
def generate_fake_person_data() -> Person:
  """
    Generate fake data for a cliche aspiring author
  """
  ...

person = generate_fake_person_data()
print(person)

Local Function Calling

Fructose @ai functions can choose to call local Python functions. Yes, even other @ai functions.

Pass the functions into the decorator with the uses argument: @ai(uses = [func_1, func_2])

For example, here's a fructose function fetching HackerNews comments using a local function and the requests library:

from fructose import Fructose
import requests
from dataclasses import dataclass

ai = Fructose()

def get(uri: str) -> str:
    """
    GET request to a URI
    """
    return requests.get(uri).text

@dataclass
class Comment:
    username: str
    comment: str

@ai(uses=[get], debug=True)
def get_comments(uri: str) -> list[Comment]:
    """
    Gets all base comments from a hacker news post
    """
    ...
    

result = get_comments("https://news.ycombinator.com/item?id=22963649")

for comment in result:
    print(f"šŸ§‘ {comment.username}: \nšŸ’¬ {comment.comment}\n")

Local function calling currently requires:

And supports arguments of basic types:

Config

Model type

Select your OpenAI model with the model keyword. Defaults to gpt-4-turbo-preview

ai = Fructose(model = "gpt-3.5-turbo")

Custom Clients and Alternative APIs

You can configure your own OpenAI client and use it with Fructose. This allows you to do things like route your calls through proxies or OpenAI-compatible LLM APIs.

from fructose import Fructose
from openai import OpenAI

client = OpenAI(
    base_url="http://my.test.server.example.com:8083",
    http_client=httpx.Client(
        proxies="http://my.test.proxy.example.com",
        transport=httpx.HTTPTransport(local_address="0.0.0.0"),
    ),
)

ai = Fructose(client = client)

Note that Fructose uses OpenAI's json return mode for all calls, and tools API for the uses function-calling. Not all alternative LLM providers support these features, and those that do likely exhibit subtle differences in the API.

You're free to point to alternative APIs using a custom client (above) or the OPENAI_BASE_URL environment variable, but Fructose does not officially support it.

Prompting

Fructose has a lightweight prompt wrapper that "just works" in most cases, but you're free to modify it using the below Flavors and Templates features. Note: we're not satisfied with this specific API, so feel free to give suggestions for alternatives.

Flavors

Flavors are optional flags to change the behavior of the prompt.

decorator level:

ai = Fructose(["random", "chain_of_thought"])

or function level:

@ai(flavors=["random", "chain_of_thought"])
def my_func():
    # ...

Custom System Prompt Templates

You're free to bring your own prompt template, using the Jinja templating language.

To use a custom template on a function level, use the system_template_path argument in the @ai() decorator, with a relative path to your Jinja template file:

@ai(system_template_path="relative/path/to/my_template.jinja")
def my_func():
    # ...

You can also set this on the decorator level, to make it default for all decorated functions.

The template must include the following variables:

For reference, find the default template here

Custom Chain Of Thought Prompt Templates

In the case of the chain_of_thought flavor being used, fructose will first run a chain-of-thought call, using a special system prompt.

Customize it on the function level with the chain_of_thought_template_path argument in the @ai() decorator.

@ai(
    flavors = ["chain_of_thought"], 
    chain_of_thought_template_path="relative/path/to/my_template.jinja"
)
def my_func():
    # ...

You can also set this on the decorator level, to make it default for all decorated functions.

For reference, find the default chain-of-thought template here


Stability

We are in v0, meaning the API is unstable version-to-version. Pin your versions to ensure new builds don't break! Also note... since LLM generations are nondeterminsitic, the calls may break too!

Use Cases

Gaming

Getting creative but strongly typed responses from LLMs is particularly useful in game dev scenarios.

Here's a prototype of an alien creature merging/breeding game: https://twitter.com/entreprenik/status/1758948061202809066

fructose in game dev

Developing and Contributing

From the root of this repo:

Create a virtual env

python3 -m venv venv
. ./venv/bin/activate

Install fructose into your pip environment with:

pip3 install -e .

This installs the fructose package in editable mode. All imports of fructose will run the fructose source at ./src/fructose directly.

Under /examples you'll find different usage examples. And run them like so:

python3 examples/fake_data.py

Run tests with:

python3 -m pytest