Guardrail Node | Vellum | Documentation

from vellum import ChatMessage
from vellum.workflows.nodes import GuardrailNode
class QualityCheckNode(GuardrailNode):
    release_tag = "production"
    metric_definition = "response_quality_metric"  # or UUID("...")
    # `metric_inputs` align with the custom Metric we've created in the Vellum UI
    metric_inputs = {
        "response": "This is the AI response to evaluate",
        "chat_history": [
            ChatMessage(role="USER", text="What's the weather?"),
            ChatMessage(role="ASSISTANT", text="It's sunny today!")
        ],
        "evaluation_criteria": {
            "coherence": True,
            "relevance": True,
            "safety": True
        }
    }

QualityCheckNode.Outputs(
    # `score` is a mandatory field for all Metrics and thus will be available on all Guardrail Nodes
    score=0.92, 
    # `details` is a result of our custom Metric definition
    details={ 
        "coherence_score": 0.95,
        "relevance_score": 0.88,
        "safety_score": 0.93
    }
)

vellum.workflows.nodes.GuardrailNode

Used to execute a Metric Definition and surface a float output representing the score. This node is commonly used to implement quality checks and guardrails in your workflows.

Metrics are defined in the Vellum UI and can be reused across workflows.

Attributes

metric_definition

Union[UUID, str]Required

Either the Metric Definition’s UUID or its name

metric_inputs

EntityInputsInterfaceRequired

The inputs for the Metric

release_tag

strDefaults to LATEST

The release tag to use for the Metric

request_options

Optional[RequestOptions]

The request options to use for the Metric execution

Outputs

score

float

The score output from the metric execution (between 0 and 1)

Additional outputs may be available depending on the metric definition.

1	from vellum import ChatMessage
2	from vellum.workflows.nodes import GuardrailNode
3
4	class QualityCheckNode(GuardrailNode):
5	release_tag = "production"
6	metric_definition = "response_quality_metric" # or UUID("...")
7	# `metric_inputs` align with the custom Metric we've created in the Vellum UI
8	metric_inputs = {
9	"response": "This is the AI response to evaluate",
10	"chat_history": [
11	ChatMessage(role="USER", text="What's the weather?"),
12	ChatMessage(role="ASSISTANT", text="It's sunny today!")
13	],
14	"evaluation_criteria": {
15	"coherence": True,
16	"relevance": True,
17	"safety": True
18	}
19	}

1	QualityCheckNode.Outputs(
2	# `score` is a mandatory field for all Metrics and thus will be available on all Guardrail Nodes
3	score=0.92,
4	# `details` is a result of our custom Metric definition
5	details={
6	"coherence_score": 0.95,
7	"relevance_score": 0.88,
8	"safety_score": 0.93
9	}
10	)