Locy Use Case: Fraud Risk Propagation¶

Propagate account risk backward over transfer edges and isolate clean accounts.

This notebook uses schema-first mode (recommended): labels, edge types, and typed properties are defined before ingest.

How To Read This Notebook¶

Step 1 initializes an isolated local database.
Step 2 defines schema (the recommended production path).
Step 3 seeds a minimal graph for this use case.
Step 4 declares Locy rules and query statements.
Steps 5-6 evaluate and inspect command/query outputs.
Step 7 tells you what to look for in the results.

1) Setup¶

Creates a temporary database directory so the example is reproducible and leaves no state behind.

import os
import shutil
import tempfile
from pprint import pprint

import uni_db

DB_DIR = tempfile.mkdtemp(prefix="uni_locy_")
print("DB_DIR:", DB_DIR)

db = uni_db.Uni.open(DB_DIR)
session = db.session()

DB_DIR: /tmp/uni_locy_0j1xeu9v

2) Define Schema (Recommended)¶

Define labels, property types, and edge types before inserting data.

(
    db.schema()
    .label("Account")
        .property("id", "string")
        .property("flagged", "bool")
    .done()
    .edge_type("TRANSFER", ["Account"], ["Account"])
    .done()
    .apply()
)

print('Schema created')

Schema created

3) Seed Graph Data¶

Insert only the entities/relationships needed for this scenario so rule behavior stays easy to inspect.

tx = session.tx()
tx.execute("CREATE (:Account {id: 'A1', flagged: true})")
tx.execute("CREATE (:Account {id: 'A2', flagged: false})")
tx.execute("CREATE (:Account {id: 'A3', flagged: false})")
tx.execute("CREATE (:Account {id: 'A4', flagged: false})")
tx.execute("MATCH (a1:Account {id:'A1'}), (a2:Account {id:'A2'}) CREATE (a1)-[:TRANSFER]->(a2)")
tx.execute("MATCH (a2:Account {id:'A2'}), (a3:Account {id:'A3'}) CREATE (a2)-[:TRANSFER]->(a3)")
tx.execute("MATCH (a4:Account {id:'A4'}), (a3:Account {id:'A3'}) CREATE (a4)-[:TRANSFER]->(a3)")
tx.commit()
print('Seeded graph data')

Seeded graph data

4) Locy Program¶

CREATE RULE defines derived relations. QUERY ... WHERE ... RETURN ... reads from those relations.

program = r'''
CREATE RULE risky_seed AS
MATCH (a:Account)
WHERE a.flagged = true
YIELD KEY a

CREATE RULE risky AS
MATCH (a:Account)
WHERE a IS risky_seed
YIELD KEY a

CREATE RULE risky AS
MATCH (a:Account)-[:TRANSFER]->(b:Account)
WHERE b IS risky
YIELD KEY a

CREATE RULE clean AS
MATCH (a:Account)
WHERE a IS NOT risky
YIELD KEY a

QUERY risky WHERE a.id = a.id RETURN a.id AS risky_account
QUERY clean WHERE a.id = a.id RETURN a.id AS clean_account
'''
print(program)

CREATE RULE risky_seed AS
MATCH (a:Account)
WHERE a.flagged = true
YIELD KEY a

CREATE RULE risky AS
MATCH (a:Account)
WHERE a IS risky_seed
YIELD KEY a

CREATE RULE risky AS
MATCH (a:Account)-[:TRANSFER]->(b:Account)
WHERE b IS risky
YIELD KEY a

CREATE RULE clean AS
MATCH (a:Account)
WHERE a IS NOT risky
YIELD KEY a

QUERY risky WHERE a.id = a.id RETURN a.id AS risky_account
QUERY clean WHERE a.id = a.id RETURN a.id AS clean_account

5) Evaluate Locy Program¶

Run the program, then inspect materialization stats (iterations, strata, and executed queries).

out = session.locy(program)

print("Derived relations:", list(out.derived.keys()))
stats = out.stats
print("Iterations:", stats.total_iterations)
print("Strata:", stats.strata_evaluated)
print("Queries executed:", stats.queries_executed)

Derived relations: ['clean', 'risky', 'risky_seed']
Iterations: 2
Strata: 3
Queries executed: 14

6) Inspect Command Results¶

Each command result can contain rows; this is the easiest way to verify your rule outputs and query projections.

print("Derived relation snapshots:")
for rel_name, rel_rows in out.derived.items():
    print(f"\\n{rel_name}: {len(rel_rows)} row(s)")
    pprint(rel_rows)

if out.command_results:
    print("\\nCommand results:")
for i, cmd in enumerate(out.command_results, start=1):
    print(f"\\nCommand #{i}:", cmd.command_type)
    rows = getattr(cmd, 'rows', None)
    if rows is not None:
        pprint(rows)
if not out.command_results:
    print("\\nNo QUERY/EXPLAIN/ABDUCE command outputs in this program.")

Derived relation snapshots:
\nclean: 3 row(s)
[{'a': Node(id=1, labels=["Account"], properties={'id': 'A2', 'flagged': False})},
 {'a': Node(id=2, labels=["Account"], properties={'flagged': False, 'id': 'A3'})},
 {'a': Node(id=3, labels=["Account"], properties={'flagged': False, 'id': 'A4'})}]
\nrisky: 1 row(s)
[{'a': Node(id=0, labels=["Account"], properties={'flagged': True, 'id': 'A1'})}]
\nrisky_seed: 1 row(s)
[{'a': Node(id=0, labels=["Account"], properties={'flagged': True, 'id': 'A1'})}]
\nCommand results:
\nCommand #1: query
[{'risky_account': 'A1'}]
\nCommand #2: query
[{'clean_account': 'A2'}, {'clean_account': 'A3'}, {'clean_account': 'A4'}]

7) What To Expect¶

Use these checks to validate output after evaluation: - A1 is risky by seed; A2 and A4 become risky by backward propagation through TRANSFER. - A3 should remain in clean because it does not transfer to a risky account. - Two query result blocks should appear: one for risky, one for clean.

8) Cleanup¶

Delete the temporary database directory created in setup.

shutil.rmtree(DB_DIR, ignore_errors=True)
print("Cleaned up", DB_DIR)

Cleaned up /tmp/uni_locy_0j1xeu9v