Skip to content
Integrated Quantum

Getting Started

This guide uses Snowflake as the example platform. The workflow is similar across platforms.

Prerequisites

Section titled “Prerequisites”
  • Snowflake Enterprise Edition (required for masking policies)
  • VEIL installed and services running
  • A table with numeric feature columns

1. Verify installation

Section titled “1. Verify installation”

These procedures require the app_admin role.

Check that services are running:

CALL veil.core.service_status();

Wait for both services to reach ready (blocks up to 10 minutes):

CALL veil.core.wait_for_services();

Test that encoding works (app_user or app_admin):

SELECT veil.core.encode('sample', ARRAY_CONSTRUCT(1.0, 2.0, 3.0, 4.0));

Returns an encoded vector like [0.23, -0.41, 0.87, 0.12].

2. Protect your data

Section titled “2. Protect your data”

Create a feature view that bundles columns into an array, then protect it.

Create a feature view (hides column names from queries):

CALL veil.create_feature_view(
    'my_db.my_schema.sales',
    'sales',
    ARRAY_CONSTRUCT('price', 'sqft', 'bedrooms', 'bathrooms'),
    ARRAY_CONSTRUCT('year_built')
);

Protect the features column:

CALL veil.protect_column(
    'my_db.my_schema.sales_v',
    'features'
);

Now query the view:

SELECT year_built, features FROM my_db.my_schema.sales_v LIMIT 3;

As an admin, you see raw feature arrays. As an analyst, you see encoded vectors. The analyst can group by year_built and use features for ML, but cannot see raw values or join to external records.

Passthrough column validation

Section titled “Passthrough column validation”

Passthrough columns are validated against an identifier denylist. Common identifier names (id, *_id, email, ssn, phone, name, address, account, customer, user) are rejected by default.

To allow identifier columns explicitly:

CALL veil.create_feature_view(
    'my_db.my_schema.sales', 'sales',
    ARRAY_CONSTRUCT('price', 'sqft', 'bedrooms', 'bathrooms'),
    ARRAY_CONSTRUCT('parcel'),
    TRUE
);

3. Train a custom encoder

Section titled “3. Train a custom encoder”

The bundled sample encoder is generic. For production, train on your data.

Start training (runs on CPU by default, GPU optional):

CALL veil.train_encoder(
    'sales',
    'my_db.my_schema.sales',
    ARRAY_CONSTRUCT('price', 'sqft', 'bedrooms', 'bathrooms'),
    16,
    100,
    256,
    ARRAY_CONSTRUCT('year_built')
);

Monitor progress:

SELECT encoder_name, status, progress
FROM veil.training_jobs_v
WHERE encoder_name = 'sales';

Once complete, generate the view from encoder metadata:

CALL veil.create_view_from_encoder('sales');

Protect the features column:

CALL veil.protect_column('my_db.my_schema.sales_v', 'features');

Retraining an existing encoder is allowed. The previous version is archived automatically.

4. Remove protection

Section titled “4. Remove protection”
CALL veil.unprotect_column('my_db.my_schema.sales_v', 'features');

Single column protection

Section titled “Single column protection”

For simple cases (one sensitive value like salary), you can protect a column directly without a feature view:

CALL veil.protect_column('my_db.my_schema.employees', 'salary', 'sample');

Encoder version history

Section titled “Encoder version history”

View archived encoder versions:

SELECT * FROM veil.encoder_versions_v WHERE encoder_name = 'sales';

GPU training

Section titled “GPU training”

Training runs on CPU by default. To switch to GPU (requires app_admin):

CALL veil.core.set_training_compute('gpu');
CALL veil.core.start_services();

To switch back to CPU:

CALL veil.core.set_training_compute('cpu');
CALL veil.core.start_services();

Next steps

Section titled “Next steps”