How to add guardrails for LLM-generated Cypher

To add guardrails for LLM-generated Cypher, guard the Neo4j driver with guard() and run the LLM's queries through it. Blocked queries raise GuardError and are never sent.

Install

pip install neo4j-guard

Guard the driver

from neo4j import GraphDatabase
from neo4j_guard import GuardError, guard

driver = guard(GraphDatabase.driver("neo4j://localhost:7687", auth=("neo4j", "password")))

try:
    driver.execute_query("MATCH (p:Person) DETACH DELETE p")
except GuardError as error:
    print(error)  # DELETE is not allowed in read mode

Queries are parsed with Neo4j's Cypher grammar, not matched against keywords. A DELETE inside a string is allowed, and a DELETE inside a CALL { } subquery is blocked.

Return blocked queries to the LLM

Catch GuardError in the tool that runs the LLM's queries and return the message. The message says what was blocked, so the LLM can rewrite the query.

from neo4j import GraphDatabase
from neo4j_guard import GuardError, guard

driver = guard(GraphDatabase.driver("neo4j://localhost:7687", auth=("neo4j", "password")))


def run_cypher(query: str) -> list[dict] | str:
    """Run a Cypher query and return its records, or why it was blocked."""
    try:
        records, _, _ = driver.execute_query(query)
    except GuardError as error:
        return f"Query blocked: {error}"
    return [record.data() for record in records]


print(run_cypher("MATCH (p:Person) DETACH DELETE p"))  # Query blocked: DELETE is not allowed in read mode