The Most Advanced Control Layer
for Large Language Models

Constrain generation, guarantee structure, and run anywhere.

Structured outputs, regex constraints, and CFGs in pure Python.

Control decoding directly. Avoid chaining and slow retries.

Stateless programs. Compatible with OpenAI, llama.cpp, Transformers.

# Select: User chooses an option
gpt + f"Pick one: {select(['red', 'blue', 'green'])}"

# Regex-constrained generation
lm += gen(regex='\d{2,4}')  # e.g. year or number

# Tool use: auto-invoked
@guidance
def add(lm, a, b):
  lm += f" = {int(a) + int(b)}"
  return lm
lm += gen(tools=[add])

# JSON: structured output
def make_profile():
  return f'''{
  "name": "{gen('name', stop='"')}",
  "age": {gen('age', regex='\d+', stop=',')}
}'''

# Grammar: recursive math
def expression():
  return select([number(), expression() + '+' + expression()])
lm += expression()