Define end-to-end applications

[rlouf]

In this PR we allow users to define end-to-end functions from a prompt template, a Pydantic model and a model name (that can be downloaded from the Hugging Face hub):

import outlines


@outlines.prompt
def test_template(text: str):
    """{{ text }}"""


class Foo(BaseModel):
    id: int


fn = outlines.Function(test_template, Foo, "gpt2")

These functions can then be executed, and imported by another module:

from .foo import function as foo_fn

result = foo_fn("A prompt")

I also added a from_github classmethod that allows to directly download a function from a github repository:

import outlines

fn = outlines.Function.from_github("outlines-dev/outlines/chain_of_density")

which looks for a variable named fn in the chain_of_density.py file in the outlines-dev/outlines repository on GitHub. It is also possible to specify an alternative name for the function:

import outlines

fn = outlines.Function.from_github("outlines-dev/outlines/chain_of_density", function_name="agent")

Possible improvements

• Loading models only when a generator is called would allow to use other model providers;
• Users might want to change the model that is being used.