LLGuidance is a low-level library for super-fast structured outputs from language models, from the Guidance team, with 893 stars on GitHub.
LLGuidance is the engine underneath reliable structured generation. When an application needs an LLM to return valid JSON matching a schema, a value from a fixed set, or text obeying a formal grammar, LLGuidance computes token masks during decoding so invalid outputs become impossible rather than merely unlikely โ and it does so with very low overhead, which is why its approach has been adopted inside major inference stacks for JSON Schema constrained generation. It reached version 1.0 in June 2025 and is actively maintained by Guidance AI, the team behind the Guidance framework, with close ties to Microsoft’s ecosystem. For most developers it works invisibly inside their serving engine; for engine builders it is the reference implementation to integrate. Its constraint engine handles the details that make this hard in practice, including computing which tokens remain valid as generation streams out token by token, so grammars and schemas are enforced continuously rather than checked after the fact.
LLGuidance is free and open source under the MIT license. There are no fees; it is a library you integrate into your own inference stack or consume through engines that already embed it.
LLGuidance is for inference engine developers, platform teams and advanced LLM builders who need guaranteed-valid structured outputs at production speed. If you serve models and your users ask for strict JSON mode, this is the kind of library that makes it fast enough to offer.