Your first Guard

Lesson 1's rules were two if statements in a list. Guardrails calls a single rule a validator, and a set of rules wrapped around one value a Guard.

Installing it

pip install guardrails-ai guardrails-ai-valid-length

Two packages, and the second one matters. Until version 0.11 validators came from a private registry, installed with guardrails hub install and a token you had to sign up for. In 0.11 they became ordinary PyPI packages named guardrails-ai-<name>, installed with pip and imported from the guardrails_ai namespace. The 0.11.0 migration guide is the page that says so, and lesson 6 is the whole story.

One rule, one value

from guardrails import Guard
from guardrails_ai.valid_length import ValidLength

desk = Guard().use(ValidLength(min=1, max=40))
outcome = desk.validate("Order 8821 ships today.")

print(outcome.validation_passed)
print(outcome.validated_output)

Guard() on its own holds no rules. .use() attaches a validator instance to it and hands the Guard back, so the two lines are usually written as one. validate() runs the rules against a value you already have.

ValidLength takes min and max in characters. It is one of ten validators that need no model, and lesson 6 lists the rest.

What comes back

print(repr(outcome.raw_llm_output))
print(outcome.error)
print(outcome.validation_summaries)

validate() returns a ValidationOutcome, and it has the same shape whether a model was involved or not. raw_llm_output is the value that went in, which reads oddly here because no model produced it. Lesson 14 calls the same Guard with a model behind it and the name stops being odd.

validation_summaries is empty because nothing failed. It only ever holds the failures, which is why it is the first thing to print when something goes wrong.