Outlines — Structured Output from Any LLM

What is Outlines?

Outlines is a library for structured text generation from LLMs. Instead of hoping the model outputs valid JSON, Outlines guarantees it through guided generation — it constrains the model's token sampling to only produce tokens that match your schema. Works with any open-source model via transformers, llama.cpp, vLLM, or MLX.

Answer-Ready: Outlines guarantees structured LLM outputs through guided generation. Supports JSON schemas, regex patterns, Pydantic models, and grammars. Works with any open-source model. No retries needed — output is always valid. 10k+ GitHub stars.

Best for: AI engineers needing reliable structured extraction. Works with: HuggingFace transformers, vLLM, llama.cpp, MLX. Setup time: Under 2 minutes.

Core Features

1. JSON Schema Generation

from pydantic import BaseModel

class Character(BaseModel):
    name: str
    age: int
    weapon: str

generator = outlines.generate.json(model, Character)
character = generator("Create an RPG character named Aria")
# Always returns valid Character object

2. Regex-Constrained Generation

# Generate valid email addresses
email_gen = outlines.generate.regex(model, r"[a-z]+@[a-z]+\.[a-z]{2,3}")
email = email_gen("Generate an email for support")

# Generate dates in specific format
date_gen = outlines.generate.regex(model, r"\d{4}-\d{2}-\d{2}")
date = date_gen("When was Python created?")

3. Choice (Classification)

classifier = outlines.generate.choice(model, ["positive", "negative", "neutral"])
sentiment = classifier("This product is amazing!")
# Always one of the three options

4. Grammar-Based Generation

# Generate valid SQL
sql_grammar = outlines.grammars.sql
sql_gen = outlines.generate.cfg(model, sql_grammar)
query = sql_gen("Write a query to find users older than 25")

5. Multiple Backend Support

# HuggingFace Transformers
model = outlines.models.transformers("mistralai/Mistral-7B-v0.1")

# vLLM (fast serving)
model = outlines.models.vllm("mistralai/Mistral-7B-v0.1")

# llama.cpp (CPU/Metal)
model = outlines.models.llamacpp("model.gguf")

# MLX (Apple Silicon)
model = outlines.models.mlxlm("mlx-community/Mistral-7B-v0.1-4bit")

Outlines vs Alternatives

FAQ

Q: Does it work with Claude or GPT-4? A: Outlines is designed for open-source models where you control token sampling. For API models, use Instructor (retry-based) instead.

Q: How does guided generation work? A: Outlines builds a finite-state machine from your schema/regex and masks invalid tokens at each generation step, ensuring only valid tokens are sampled.

Q: Does it slow down generation? A: Minimal overhead. The FSM is precompiled, and token masking is O(1) per step.