Outlines

Outlines

Outlines (by dottxt-ai) makes LLM outputs reliably structured by constraining token sampling to only produce valid JSON, regex patterns, context-free grammars, or Pydantic model instances — with zero retry loops. The constraint is applied at the logit level, so it is mathematically impossible for the model to produce invalid output.

Key Features

• JSON schema enforcement — pass a Pydantic model or JSON schema; get a perfectly valid object every time
• Regex-guided generation — constrain output to any regular expression pattern
• Grammar-based — EBNF context-free grammars for generating valid code, SQL, or custom DSLs
• Choice constraints — force the model to pick from a fixed set of options
• Model-agnostic — works with transformers, llama.cpp (via llama-cpp-python), and vLLM
• No retries — unlike post-hoc JSON parsers, invalid tokens are literally impossible

Quick Start

pip install outlines

import outlines
from pydantic import BaseModel

class Person(BaseModel):
    name: str
    age: int
    city: str

model = outlines.models.transformers("Qwen/Qwen2.5-0.5B-Instruct")
generator = outlines.generate.json(model, Person)
person = generator("Generate a fictional person.")
print(person)  # Person(name='Alice', age=30, city='Paris')

Install via ai-supply

npx ai-supply add outlines-structured-generation

Curated mirror of the open-source Outlines project (Apache-2.0). Install upstream from the repository.