Outlines

Rather than repairing malformed model output after the fact, Outlines steers generation so the text can only take shapes you declare. You pass a model call a Python type, a Pydantic class, a JSON Schema, a Literal set of options, a regular expression or a context-free grammar; for locally run models the library compiles that target, using the Rust-based outlines-core package, into an index that blocks invalid tokens at each decoding step. Backends include Hugging Face Transformers, llama.cpp, MLX on Apple silicon, vLLM in offline or server mode, SGLang, TGI, Ollama and LM Studio, while hosted OpenAI and Gemini calls receive the type as the provider's own schema parameter (the Anthropic wrapper only returns plain text). Extras include Jinja-based prompt templates and an Application helper that pairs a template with an output type. It installs from pip on Python 3.10 through 3.13, and hardware needs are those of whichever model you load. Apache-2.0 licensed and fine for commercial use; the maintaining company, .txt, sells separate enterprise tooling. Reach for it when downstream code has to parse every response with no retry loop.