Outlines

Overview

Outlines forces a model's output to match a declared type. For local backends it compiles the type (Pydantic model, regex, choice list, context-free grammar) into a constraint on token sampling, so the first generation is already valid and no retry loop is needed. For hosted backends (OpenAI, Anthropic, Gemini, Ollama, vLLM server) it passes the type to the provider's own structured-output feature through one common interface.

Checked against Outlines 1.3 (October 2026). The 1.x API is model(prompt, output_type). The 0.x calls (outlines.generate.json(...), outlines.models.transformers(...), outlines.generate.regex(...)) were removed; code using them fails on current versions.

Instructions

Installation and model loading

pip install "outlines[transformers]"     # local Hugging Face models (also pulls torch)
pip install "outlines[ollama]"           # or: openai, anthropic, gemini, llamacpp, vllm, mlxlm

import outlines
from transformers import AutoModelForCausalLM, AutoTokenizer

name = "microsoft/Phi-3-mini-4k-instruct"
model = outlines.from_transformers(
    AutoModelForCausalLM.from_pretrained(name),
    AutoTokenizer.from_pretrained(name),
)

Other loaders: outlines.from_ollama(client, "llama3.2"), outlines.from_openai(client, model_name), outlines.from_vllm(openai_client, model_name) for a running vLLM server (OpenAI client pointed at http://127.0.0.1:8000/v1), outlines.from_vllm_offline(llm), outlines.from_llamacpp(llm), plus from_anthropic, from_gemini, from_mistral, from_sglang, from_tgi, from_mlxlm.

Output types

The second argument of the call is the constraint. Always pass max_new_tokens on the Transformers backend: the default of 20 truncates JSON.

from enum import Enum
from typing import Literal
from pydantic import BaseModel, Field
from outlines.types import Regex, Choice, CFG

class ReviewAnalysis(BaseModel):
    sentiment: Literal["positive", "negative", "neutral"]
    score: float = Field(ge=0, le=1)
    topics: list[str]

raw = model("Analyze: 'Great product, slow shipping'", ReviewAnalysis, max_new_tokens=200)
review = ReviewAnalysis.model_validate_json(raw)     # the call returns a JSON string

phone = model("Support phone number:", Regex(r"\(\d{3}\) \d{3}-\d{4}"), max_new_tokens=20)
label = model("Is this spam? 'You won $1000000!!!'", Choice(["spam", "ham", "uncertain"]))
count = model("How many minutes in an hour?", int, max_new_tokens=5)

Also accepted: Literal[...], Enum classes, plain types (int, float, bool, datetime.date), JsonSchema(schema_string), and CFG(lark_grammar_string) for a Lark grammar (not every backend supports CFG).

Batches and reuse

answers = model.batch(["Capital of Lithuania?", "Capital of Latvia?"], max_new_tokens=20)

generator = outlines.Generator(model, ReviewAnalysis)   # compile the constraint once
raw = generator("Analyze: 'Arrived broken'", max_new_tokens=200)

Examples

Example 1: Classify support tickets into fixed labels

Request: "Sort these tickets into billing, bug or feature, nothing else."

tickets = ["I was charged twice in September", "Export button does nothing on Safari"]
labels = Choice(["billing", "bug", "feature"])
print(model.batch([f"Ticket: {t}\nCategory:" for t in tickets], labels, max_new_tokens=8))

Result: ['billing', 'bug']. The output can only be one of the three strings, so no post-processing is needed.

Example 2: Extract a typed record with a local server

Request: "Get name, city and plan from this signup email using our vLLM server."

import openai, outlines
from pydantic import BaseModel

class Signup(BaseModel):
    name: str
    city: str
    plan: str

client = openai.OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="local-vllm-key")
model = outlines.from_vllm(client, "microsoft/Phi-3-mini-4k-instruct")
raw = model("Priya Nair from Pune signed up for the Growth plan.", Signup)
print(Signup.model_validate_json(raw))

Result: Signup(name='Priya Nair', city='Pune', plan='Growth'), validated against the schema.

Guidelines

• Constrained output guarantees the shape, not the truth: a valid JSON object can still hold a wrong answer. Keep prompts specific and check values that matter.
• Complex schemas and large grammars add compile time on the first call; reuse a Generator rather than recompiling.
• Backend support differs (for example CFG and some regex features); the model documentation page for each backend lists what it accepts.
• Python 3.14 is not supported yet by the current release.
• If you already use PydanticAI or a provider's native structured outputs against a hosted model, you may not need Outlines; its strength is local and self-hosted models.
• Never send private text to a hosted backend you have not approved; local backends keep data on the machine.