Outlines-core

This package provides the core functionality for structured generation, formerly implemented in Outlines, with a focus on performance and portability, it offers a convenient way to:

• build regular expressions from JSON schemas

• construct an Index object by combining a Vocabulary and regular expression to efficiently map tokens from a given vocabulary to state transitions in a finite-state automation

Example

Basic example of how it all fits together.

use outlines_core::prelude::*;

// Define a JSON schema
let schema = r#"{
    "type": "object",
    "properties": {
        "name": { "type": "string" },
        "age": { "type": "integer" }
    },
    "required": ["name", "age"]
}"#;

// Generate a regular expression from it
let regex = json_schema::regex_from_str(&schema, None)?;

// Create `Vocabulary` from pretrained large language model (but manually is also possible)
let vocabulary = Vocabulary::from_pretrained("openai-community/gpt2", None)?;

// Create new `Index` from regex and a given `Vocabulary`
let index = Index::new(&regex, &vocabulary)?;

let initial_state = index.initial_state();
let allowed_tokens = index.allowed_tokens(&initial_state).expect("Some allowed token ids");
let token_id = allowed_tokens.first().expect("First token id");
let next_state = index.next_state(&initial_state, token_id);
let final_states = index.final_states();

Python Bindings

Additionally, project provides interfaces to integrate the crate's functionality with Python.

import json

from outlines_core.json_schema import build_regex_from_schema
from outlines_core.guide import Guide, Index, Vocabulary

schema =  {
  "title": "Foo",
  "type": "object",
  "properties": {"date": {"type": "string", "format": "date"}}
}
regex = build_regex_from_schema(json.dumps(schema))

vocabulary = Vocabulary.from_pretrained("openai-community/gpt2")
index = Index(regex, vocabulary)
guide = Guide(index)

# Get current state of the Guide:
current_state = guide.get_state()

# Get allowed tokens for the current state of the Guide:
allowed_tokens = guide.get_tokens()

# Advance Guide to the next state via some token_id and return allowed tokens for that new state:
next_allowed_tokens = guide.advance(allowed_tokens[-1])

# To check if Guide is finished:
guide.is_finished()

# If it's finished then this assertion holds:
assert guide.get_tokens() == [vocabulary.get_eos_token_id()]

How to contribute?

Setup

Fork the repository on GitHub and clone the fork locally:

git clone [email protected]/YourUserName/outlines-core.git
cd outlines-core

Create a new virtual environment and install the dependencies in editable mode:

python -m venv .venv
source .venv/bin/activate
pip install -e ".[test]"
pre-commit install

Before pushing your code

If working with Python bindings don't forget to build Rust extension before testing, for example, in debug mode:

make build-extension-debug

Run Python tests:

pytest

Run Rust tests:

cargo test

Or alternatively using Makefile for both:

make test

Finally, run the code style checks:

pre-commit run --all-files

Or using Makefile:

make pcc

If necessary you can run benchmarks locally:

make pybench

Join us

• 💡 Have an idea? Come chat with us on Discord
• Found a bug? Open an issue