LangGraph¶

pip install director-ai[langgraph]

Guardrail Node¶

from director_ai.integrations.langgraph import (
    director_ai_node,
    director_ai_conditional_edge,
)
from langgraph.graph import StateGraph

graph = StateGraph(dict)
graph.add_node("generate", llm_node)
graph.add_node("guardrail", director_ai_node(
    facts={"refund": "within 30 days"},
    on_fail="flag",  # "raise" | "flag" | "rewrite"
))
graph.add_node("retry", retry_node)
graph.add_node("output", output_node)

graph.add_edge("generate", "guardrail")
graph.add_conditional_edges(
    "guardrail",
    director_ai_conditional_edge(
        approved_node="output",
        rejected_node="retry",
    ),
)

State Keys¶

After the guardrail node runs, these keys are added to state:

Key	Type	Description
`director_ai_score`	float	Coherence score
`director_ai_approved`	bool	Pass/fail
`director_ai_h_logical`	float	Logical divergence
`director_ai_h_factual`	float	Factual divergence
`director_ai_rewritten`	bool	True if rewrite mode activated

Failure Modes¶

on_fail="raise" — raises HallucinationError
on_fail="flag" — sets director_ai_approved=False, continues
on_fail="rewrite" — replaces response with KB context