Outlines

Outlines is an open-source Python library developed by .txt that guarantees structured outputs from large language models during generation — not after. Unlike post-processing approaches that parse or fix bad outputs, Outlines constrains the LLM's token generation to match your desired output type exactly. It works with any model backend including OpenAI, Ollama, vLLM, Hugging Face Transformers, and llama.cpp.

• Guaranteed structured outputs — constrain LLM generation at the token level so outputs always match your schema, no parsing required
• Multiple output types — supports Literal choices, Python primitives (int, float), Pydantic models, JSON schemas, regular expressions, and context-free grammars
• Function calling — automatically infer the required output structure from a Python function's type signature
• Provider independence — the same model(prompt, output_type) API works across OpenAI, Gemini, vLLM, Ollama, Transformers, and the .txt API
• Prompt templates — Jinja2-based reusable templates to separate complex prompts from application code, with support for few-shot examples
• Batch processing — pass a list of prompts to process multiple inputs in a single call
• Custom types — build complex composite types using Python's type system for domain-specific structured extraction
• Applications — encapsulate templates and output types into reusable application functions
• Easy installation — install via pip install outlines and connect to your preferred model in a few lines of Python
• Active community — Discord, GitHub issues, and open contribution guide; trusted by NVIDIA, Cohere, HuggingFace, and vLLM