Guardrails AI¶
Director-AI provides a custom Guardrails AI validator that checks generated text with the same coherence scorer used by the rest of the runtime.
The guardrails extra is a stable documentation anchor. The adapter imports
Guardrails AI only when used, so install guardrails-ai from a package index
that provides a compatible build for your Python runtime.
from guardrails import Guard
from director_ai.integrations.guardrails_ai import attach_guardrails_validator
guard = attach_guardrails_validator(
Guard(),
facts={"sla": "Support SLA is four hours."},
threshold=0.4,
use_nli=False,
)
outcome = guard.parse(
"The support SLA is four hours.",
metadata={
"prompt": "What is the support SLA?",
"tenant_id": "tenant-a",
},
)
Validator Class¶
Use build_guardrails_validator_class() when the validator must be registered
before constructing a Guardrails spec:
from guardrails import Guard
from director_ai.integrations.guardrails_ai import (
build_guardrails_validator_class,
)
DirectorAICoherenceValidator = build_guardrails_validator_class()
guard = Guard().use(
DirectorAICoherenceValidator(
facts={"refunds": "Refunds are available within 30 days."},
threshold=0.5,
)
)
Metadata Contract¶
The validator accepts tenant-safe metadata:
| Field | Purpose |
|---|---|
prompt, query, question, or input |
Original user request used as scorer context |
messages |
Chat message list; the latest user message is used |
tenant_id or tenant |
Optional tenant id passed to the scorer |
Failure messages do not include the raw model output. Score details are returned
under metadata["director_ai"] on the Guardrails validation result.
build_guardrails_validator_class(*, name: str = DEFAULT_VALIDATOR_NAME, data_type: str = 'string') -> type
Return and register a Guardrails AI validator class.
The returned class follows Guardrails AI's custom-validator contract:
subclass Validator, implement _validate(value, metadata), and return
PassResult or FailResult.
build_guardrails_validator(*, facts: dict[str, str] | None = None, store: GroundTruthStore | None = None, threshold: float = 0.5, use_nli: bool | None = None, on_fail: Any = None, name: str = DEFAULT_VALIDATOR_NAME, data_type: str = 'string') -> Any
Build a registered Guardrails AI validator instance.
Attach the Director-AI validator to Guard().use(...) style guards.