queryguard-sql 0.1.2

SQL validation and policy enforcement for NL2SQL pipelines

QueryGuard SQL

• QueryGuard SQL is a small Python library for validating SQL before execution.

• It is useful for LLM-generated SQL, NL2SQL apps, analytics assistants, and any workflow where SQL should be checked before it reaches the database.

Repository: github.com/WajeehAlamoudi/QueryGuard

Install

pip install queryguard-sql

Import package:

from queryguard import QueryGuard, GuardConfig

Quick Use

from queryguard import QueryGuard, GuardConfig

config = GuardConfig(
    database_type="mysql",
    allowed_tables=["users", "orders"],
    read_only=True,
    max_rows=100,
)

guard = QueryGuard(config)

result = guard.check("SELECT * FROM users")

print(result.allowed)
print(result.final_sql)
print(result.detected_tables)
print(result.errors)

Output:

True
"SELECT * FROM users LIMIT 100"
["users"]
[]

When To Use

Use QueryGuard SQL when you need to:

• check SQL generated by an LLM
• enforce read-only SQL
• allow or block specific tables
• validate columns using a schema
• add safe row limits
• suggest similar table names
• inspect SQL before database execution

GuardConfig

GuardConfig defines the policy used by QueryGuard.

config = GuardConfig(
    database_type="postgres",
    allowed_tables=["public.users", "public.orders"],
    blocked_tables=["public.payments"],
    read_only=True,
    allow_select_star=False,
    max_rows=500,
    schema={
        "public.users": ["id", "name", "email"],
        "public.orders": ["id", "user_id", "total"],
    },
    suggest_tables=True,
    auto_fix_tables=False,
)

Parameters

Parameter	Type	Default	Description
database_type	str	required	SQL dialect, such as mysql, postgres, sqlite, bigquery, snowflake, tsql.
allowed_tables	list[str] | None	None	Tables that may be queried. None means allow all tables unless blocked.
blocked_tables	list[str]	[]	Tables that must always be blocked. This wins over allowed_tables.
read_only	bool	True	Allows only read-only query types when enabled.
allow_select_star	bool	False	Allows SELECT * when schema rules are active.
max_rows	int | None	1000	Adds or lowers row limits. None disables limit rewriting.
schema	dict[str, list[str]] | None	None	Known tables and allowed columns.
suggest_tables	bool	True	Suggests similar known table names.
auto_fix_tables	bool	False	Reserved for table auto-fix behavior. Prefer suggestions for now.

Table Rules

GuardConfig(
    database_type="mysql",
    allowed_tables=["users"],
)

Allowed:

SELECT * FROM users

Blocked:

SELECT * FROM payments

Schema-qualified table names are strict:

allowed_tables=["analytics.users"]

allows:

SELECT id FROM analytics.users

but not:

SELECT id FROM public.users

Schema Rules

If schema is provided, QueryGuard can validate selected columns.

GuardConfig(
    database_type="mysql",
    schema={
        "users": ["id", "name", "email"],
    },
)

Allowed:

SELECT id, name FROM users

Blocked:

SELECT password_hash FROM users

SELECT Star

When schema is provided and allow_select_star=False, this is blocked:

SELECT * FROM users

This remains allowed:

SELECT COUNT(*) FROM users

If no schema is provided, SELECT * is allowed because no column policy is active.

Suggestions

Suggestions use known tables from allowed_tables or schema.

guard = QueryGuard(
    GuardConfig(
        database_type="mysql",
        allowed_tables=["users", "orders"],
        suggest_tables=True,
    )
)

result = guard.check("SELECT * FROM usres")

print(result.suggestions)

Output:

{"usres": ["users"]}

QueryGuard

QueryGuard receives a GuardConfig and uses it to validate SQL.

guard = QueryGuard(config)

Methods

Method	Returns	Description
check(sql, raise_on_blocked=False)	GuardResult	Full validation result.
validate(sql)	str	Returns safe SQL or raises if blocked.
rewrite(sql)	str	Returns rewritten SQL or raises if blocked.
is_safe(sql)	bool	Returns True when SQL is allowed.
tables(sql)	list[str]	Returns detected table names.
check_many(sqls)	list[GuardResult]	Checks many SQL strings.
explain(sql)	str	Human-readable explanation.
set_policy(config)	None	Replaces the current config.

Example

result = guard.check("DELETE FROM users")

print(result.allowed)
print(result.errors)

Output:

False
["DELETE is not allowed in read_only mode"]

GuardResult

guard.check() returns a GuardResult.

result.allowed          # bool
result.original_sql     # str
result.final_sql        # str | None
result.database_type    # str | None
result.detected_tables  # list[str]
result.errors           # list[str]
result.warnings         # list[str]
result.suggestions      # dict[str, list[str]]

Example:

result = guard.check("SELECT * FROM users")

if result.allowed:
    db.execute(result.final_sql)
else:
    print(result.errors)

Supported Dialects

Common supported dialect names:

mysql
mariadb
postgres
postgresql
sqlite
bigquery
snowflake
duckdb
redshift
trino
presto
spark
tsql
sqlserver
mssql
oracle
clickhouse

QueryGuard SQL uses sqlglot for SQL parsing and generation.

Security Note

QueryGuard SQL validates SQL structure and configured policy before execution. It does not replace database permissions, parameterized queries, or least-privilege database users.

For production systems, always keep database-side protections enabled.

License

MIT License.