Outlines

Outlines is a Python library for structured generation: instead of fixing bad LLM output after the fact with parsing and regex, it constrains generation itself so outputs are valid by construction. You declare the desired shape — a Literal for multiple choice, a Pydantic model / dataclass / TypedDict for JSON, a regex pattern, or a context-free grammar — and Outlines enforces it at the logits level for local models (OutlinesLogitsProcessor plugs straight into HuggingFace transformers) or via provider structured-output support for APIs. It ships model adapters for transformers, llama.cpp, vLLM, Ollama, OpenAI, Gemini, and Dottxt's own API, plus prompt templates and applications that bundle templates with output types. Built by Dottxt (.txt); the library is Apache-2.0 open source with a commercial API tier alongside it.

1. Install: pip install outlines (extras available, e.g. pip install "outlines[transformers]" which pulls transformers, datasets, and torch).
2. Pick your model adapter: local via outlines.models.transformers("microsoft/Phi-3-mini-4k-instruct", device="cuda") (or pass a pre-loaded HF model + tokenizer), server via vLLM/Ollama adapters, or API via outlines.from_gemini(genai.Client(), "gemini-1.5-flash-latest") / OpenAI.
3. Choose an output type: Literal["Yes", "No"] for multiple choice, a Pydantic class for JSON, a regex string, or a grammar.
4. Generate: result = model("Pizza or burger?", PizzaOrBurger) — the returned text is guaranteed to match the declared structure.
5. (Alternative) Drop OutlinesLogitsProcessor into your own HuggingFace generate() call to apply constraints without the Outlines model wrapper.