Skip to content
Back to full roadmap
topicadvanced★ Pro

Constraint Decoding

Token-level constraints prevent the model from emitting 'invalid' tokens (Outlines, guidance, llama.cpp grammar).

3 hours1 resources1 prereqs

Outlines (Python), guidance (Microsoft), llama.cpp grammar — apply regex/grammar during sampling so only schema-compliant tokens stay probable.

For open-source models, this is the 100% way to enforce JSON. Closed APIs already do similar work in structured-output mode.

Prerequisites

Resources(1)