Source: Outlines Documentation and README
dottxt-ai/outlines v1 docs: one call (model(prompt, output_type)) for Literal, regex, CFG and Pydantic outputs; logits-processor masking for local models, schema pass-through for API models; outlines_core, llguidance and xgrammar backends.
Contents
- What: official docs (home page, Architecture guide) and GitHub README; latest PyPI release
outlines1.3.3 (2026-08-06),outlines-core0.2.14 - Author/publisher: .txt (dottxt), maintainers of Outlines; Apache-2.0
- Date: living documents; read 2026-11-14
- Original: Docs, Architecture, README
Summary
Outlines makes structure a property of generation rather than something fixed afterwards. You wrap a model (outlines.from_transformers(model, tokenizer), outlines.from_openai(client, "model"), llama.cpp, MLX, vLLM, SGLang, Ollama, Gemini, Anthropic and others) and call it with a prompt and an output type. The output is a string that matches the type; Pydantic users then call Model.model_validate_json on it.
Key claims
- Output types: Python basic types,
Literal[...]multiple choice, JSON Schema / Pydantic models, regex, and context-free grammars. - Two model kinds. Steerable (Transformers, llama.cpp, MLX-LM): Outlines builds a logits processor that sets invalid tokens to −inf each step. Black-box (OpenAI, Anthropic, Gemini, Mistral, vLLM, vLLM offline, SGLang, TGI, Ollama, Dottxt): Outlines converts the output type to the provider’s
response_formatand the provider enforces it. - A
Generatorfactory picksSteerableGeneratoror (Async)BlackBoxGenerator; steerable generators cache the logits processor and reset its state per call. - Type system: Python types become
Terms; CFG and JSON Schema terms go straight to the backend, everything else is converted to a regex. - Backends (steerable only):
outlines_core(Rust FSM, default for regex and JSON Schema),llguidance(the only one supporting CFG),xgrammar; an explicitbackend=argument overrides the default.