Neo4j Driver Python Skill

When to Use

• Writing Python code that connects to Neo4j
• Setting up driver, sessions, transactions, or async patterns
• Debugging result handling, serialization, or UNWIND batching
• Reviewing Neo4j driver usage in Python code

When NOT to Use

• Writing/optimizing Cypher → neo4j-cypher-skill
• Driver version upgrades → neo4j-migration-skill
• GraphRAG pipelines (neo4j-graphrag package) → neo4j-graphrag-skill

Installation

bashCopy code
pip install neo4j                  # package name is `neo4j`, NOT `neo4j-driver` (deprecated since v6)
pip install neo4j-rust-ext         # optional: 3–10× faster serialization, same API

Python >=3.10 required for v6.x. Python 3.14 supported [6.1+]. Pandas 3 and PyArrow 23/24 supported [6.2+]. PyArrow 25 and Bolt 6.1 uuid.UUID values supported [6.3+]; driver-created SSL contexts honour SSLKEYLOGFILE [6.3+].

Neo4j 2026.08+ UUID properties require neo4j>=6.3 to round-trip as uuid.UUID.

Environment Variables

Load connection config from environment — never hardcode credentials.

pythonCopy code
import os
from dotenv import load_dotenv   # pip install python-dotenv

load_dotenv(".env")   # reads NEO4J_URI / NEO4J_USERNAME / NEO4J_PASSWORD / NEO4J_DATABASE

URI      = os.getenv("NEO4J_URI",      "neo4j://localhost:7687")
USER     = os.getenv("NEO4J_USERNAME", "neo4j")
PASSWORD = os.getenv("NEO4J_PASSWORD", "")
DATABASE = os.getenv("NEO4J_DATABASE", "neo4j")

.env file format:

Copy code
NEO4J_URI=neo4j+s://xxx.databases.neo4j.io
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=secret
NEO4J_DATABASE=neo4j

Add .env to .gitignore. Without python-dotenv, use export in shell or os.getenv directly.

Driver Lifecycle

Create one Driver per application. Thread-safe, expensive to create. Never create per-request.

pythonCopy code
from neo4j import GraphDatabase

URI  = "neo4j+s://xxx.databases.neo4j.io"   # Aura
AUTH = ("neo4j", "password")

# Context manager — preferred for scripts
with GraphDatabase.driver(URI, auth=AUTH) as driver:
    driver.verify_connectivity()
    # ... work ...

# Long-lived singleton (service / web app)
driver = GraphDatabase.driver(URI, auth=AUTH)
driver.verify_connectivity()
# on shutdown:
driver.close()

URI schemes:

Auth options: ("user", "pass") tuple, basic_auth(), bearer_auth("jwt"), kerberos_auth("b64").

Choosing the Right API

session.run() retry [6.2+]: single immediate retry on DBMS-marked idempotent errors only (currently admission control). Disable with disable_auto_commit_retries=True at driver or session level.

execute_query — Default API

pythonCopy code
from neo4j import GraphDatabase, RoutingControl

# Tuple unpacking — most common
records, summary, keys = driver.execute_query(
    "MATCH (p:Person {name: $name})-[:KNOWS]->(f) RETURN f.name AS name",
    name="Alice",
    routing_=RoutingControl.READ,   # route reads to replicas
    database_="neo4j",              # always specify — saves a round-trip
)
for record in records:
    print(record["name"])
print(summary.result_available_after, "ms")

# Write — check counters
summary = driver.execute_query(
    "CREATE (p:Person {name: $name, age: $age})",
    name="Bob", age=30,
    database_="neo4j",
).summary
print(summary.counters.nodes_created)

Trailing-underscore convention — config kwargs end with _ (database_, routing_, auth_, result_transformer_, bookmark_manager_). No query parameter name may end with _; pass those via parameters_={"key_": val}.

Never f-string or format Cypher. Always $param — prevents injection and enables plan caching.

result_transformer_ — reshape before return:

pythonCopy code
import neo4j
df      = driver.execute_query("MATCH (p:Person) RETURN p.name, p.age", database_="neo4j",
                                result_transformer_=neo4j.Result.to_df)
record  = driver.execute_query("MATCH (p:Person {name:$n}) RETURN p", n="Alice", database_="neo4j",
                                result_transformer_=neo4j.Result.single)   # raises if 0 or 2+ results

Result.single() raises ResultNotSingleError on zero results (not just 2+). Use single(strict=False) for None-on-empty.

Managed Transactions (execute_read / execute_write)

Use for large results or multiple queries in one transaction.

pythonCopy code
with driver.session(database="neo4j") as session:

    def get_people(tx):
        result = tx.run("MATCH (p:Person) WHERE p.name STARTS WITH $pfx RETURN p.name AS name",
                        pfx="Al")
        return [r["name"] for r in result]   # consume INSIDE callback — Result invalid after tx closes

    names = session.execute_read(get_people)

    def create_person(tx):
        tx.run("CREATE (p:Person {name: $name})", name="Carol")

    session.execute_write(create_person)

Result lifetime — Result is a lazy cursor backed by the open transaction. Returning it unconsumed raises ResultConsumedError. Always collect to list inside the callback.

Callback may retry on transient failures — keep callbacks idempotent; move side effects (HTTP calls, emails) outside the callback.

Timeout/metadata via @unit_of_work (named functions only — cannot decorate lambdas):

pythonCopy code
from neo4j import unit_of_work

@unit_of_work(timeout=5.0, metadata={"app": "svc", "user": user_id})
def get_people(tx):
    return [r["name"] for r in tx.run("MATCH (p:Person) RETURN p.name AS name")]

session.execute_read(get_people)

Implicit Transactions (session.run)

Use only for LOAD CSV, CALL {} IN TRANSACTIONS, or quick scripts. session.run() does a single immediate retry on idempotent (DBMS-marked) errors only [6.2+]; other errors do not retry.

pythonCopy code
with driver.session(database="neo4j") as session:
    result = session.run("CREATE (p:Person {name: $name})", name="Alice")
    summary = result.consume()   # call consume() to guarantee commit before proceeding
    print(summary.counters.nodes_created)

# Opt out of one-shot retry [6.2+] — driver- or session-level
driver = GraphDatabase.driver(URI, auth=AUTH, disable_auto_commit_retries=True)
with driver.session(database="neo4j", disable_auto_commit_retries=True) as session:
    session.run("...")

Async API

Mirror of sync API — replace GraphDatabase with AsyncGraphDatabase, await every call.

pythonCopy code
from neo4j import AsyncGraphDatabase
import asyncio

# Singleton — same rule as sync: never create per-request
driver = AsyncGraphDatabase.driver(URI, auth=AUTH)

async def main():
    records, _, _ = await driver.execute_query(
        "MATCH (p:Person) RETURN p.name AS name",
        database_="neo4j", routing_=RoutingControl.READ,
    )
    print([r["name"] for r in records])
    await driver.close()

asyncio.run(main())

FastAPI lifespan pattern:

pythonCopy code
from contextlib import asynccontextmanager
from fastapi import FastAPI

_driver = None

@asynccontextmanager
async def lifespan(app: FastAPI):
    global _driver
    _driver = AsyncGraphDatabase.driver(URI, auth=AUTH)
    await _driver.verify_connectivity()
    yield
    await _driver.close()

app = FastAPI(lifespan=lifespan)

Parallel queries with asyncio.gather:

pythonCopy code
results = await asyncio.gather(
    driver.execute_query("MATCH (a:Artist) RETURN a.name AS name", database_="neo4j"),
    driver.execute_query("MATCH (v:Venue)  RETURN v.name AS name",  database_="neo4j"),
)

Never use sync GraphDatabase in asyncio — blocks the event loop.

Full async patterns → references/async.md

Error Handling

pythonCopy code
from neo4j.exceptions import (
    Neo4jError, ServiceUnavailable, TransientError,
    AuthError, ConstraintError,
)

try:
    driver.execute_query("...", database_="neo4j")
except AuthError:
    ...  # bad credentials
except ServiceUnavailable:
    ...  # no servers reachable
except ConstraintError as e:
    # unique/existence constraint violation — catch BEFORE Neo4jError (it's a subclass)
    print(e.code, e.message)
except TransientError as e:
    # raised only after retries exhausted (execute_query retries automatically)
    print(e.code)
except Neo4jError as e:
    print(e.code, e.message, e.gql_status)

Catch ConstraintError before Neo4jError — it is a subclass and will be swallowed otherwise.

Result Access & Null Safety

pythonCopy code
record = records[0]
record["name"]               # by key — KeyError if absent
record[0]                    # by index
record.get("name")           # None for absent key OR graph null
record.get("name", "Unknown")
d = record.data()            # dict — values still driver objects for Node/Rel/temporal types

record.data() is not JSON-safe if result contains Node, Relationship, Path, or neo4j.time.* values. Project scalar fields in Cypher instead of returning whole nodes.

pythonCopy code
# ❌ raises TypeError on json.dumps
records, _, _ = driver.execute_query("MATCH (p:Person) RETURN p", database_="neo4j")
json.dumps(records[0].data())

# ✅ project scalars
records, _, _ = driver.execute_query(
    "MATCH (p:Person) RETURN p.name AS name, p.age AS age", database_="neo4j")
json.dumps(records[0].data())   # safe

Node/Relationship/temporal access:

pythonCopy code
node = record["p"]           # neo4j.graph.Node
node.element_id              # stable within this transaction only
node.labels                  # frozenset({'Person'})
dict(node)                   # all properties as plain dict

rel  = record["r"]           # neo4j.graph.Relationship
rel.type                     # 'KNOWS'

dt = record["created_at"]    # neo4j.time.DateTime
dt.to_native()               # datetime.datetime (loses sub-µs precision)

Full type mapping table → references/data-types.md

Batch Writes with UNWIND

Pass list[dict] — only shape the driver serializes correctly for UNWIND.

pythonCopy code
people = [{"name": "Alice", "age": 30}, {"name": "Bob", "age": 25}]
driver.execute_query(
    "UNWIND $rows AS row MERGE (p:Person {name: row.name}) SET p.age = row.age",
    rows=people,
    database_="neo4j",
)

Custom objects and dataclasses must be converted to dict before passing as parameters.

Performance

• Always set database_ / database= — omitting triggers a home-database round-trip per call.
• execute_read routes to replicas automatically; use routing_=RoutingControl.READ with execute_query.
• Batch writes: one execute_write callback for the whole list > one tx per item.
• Large results: stream lazily inside execute_read callback; execute_query is always eager.

Connection pool tuning:

pythonCopy code
driver = GraphDatabase.driver(URI, auth=AUTH,
    max_connection_pool_size=50,        # default 100
    connection_acquisition_timeout=30,  # seconds to wait for free connection
    max_connection_lifetime=3600,       # seconds; recycles stale connections
    connection_timeout=15,
    keep_alive=True,
)

Session exhaustion: each open session holds a connection. Always use with driver.session(...) as session.

Full performance patterns → references/performance.md

Common Errors

References

Load on demand:

• references/async.md — full async patterns: managed transactions, result methods, concurrency
• references/data-types.md — complete Python↔Cypher type mapping, temporal conversion, graph object API, spatial types (CartesianPoint/WGS84Point)
• references/performance.md — connection pool, lazy streaming, threading vs asyncio, bookmarks/causal consistency
• references/transactions.md — explicit transactions, rollback, commit uncertainty, unit_of_work details

Docs:

• https://neo4j.com/docs/python-manual/current/
• https://neo4j.com/docs/api/python-driver/current/

Checklist

• Package installed as neo4j (not neo4j-driver)
• One Driver instance created at startup; shared everywhere
• verify_connectivity() called at startup
• database_ / database= set on every call
• $param placeholders used — no f-strings or .format()
• Result consumed inside tx callback (not returned raw)
• Sessions used as context managers (with driver.session(...) as session)
• ConstraintError caught before Neo4jError
• AsyncGraphDatabase used in asyncio code (not sync driver)
• Async driver created once at app startup (not per request)
• Side effects outside execute_read/write callbacks
• UNWIND batches use list[dict]