Quick Start

Replace <app_name> with the name you chose at install.

1. Install and grant privileges

Install the app. Grant five privileges when prompted. Services start automatically.

2. Bind a training data source

Tell the app which table to learn from. The easiest way is Snowsight: Data Products > Apps > AIQu VEIL > Settings. Or bind via SQL:

CALL <app_name>.core.register_reference(
    'training_source', 'ADD',
    SYSTEM$REFERENCE('TABLE', 'my_db.my_schema.sales', 'PERSISTENT', 'SELECT')
);

For a view, use training_source_view and 'VIEW' instead of training_source and 'TABLE'.

3. Train an encoder

Pick a name, point it at your table, tell it which columns to encode. Everything else is automatic.

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

The first three arguments are required. The rest have defaults (16 latent dimensions, 100 epochs, batch size 256, no passthrough columns). Here we pass year_built as a passthrough column so it appears alongside the encoded features in the view. See Procedures for the full signature.

Poll for status:

SELECT job_id, status, progress
FROM <app_name>.core.training_jobs_v
WHERE encoder_name = 'sales';

4. Create the protected view

Generate a protected view. Masking policies are applied automatically.

CALL <app_name>.core.create_view_from_encoder('sales');

5. Query the view

Query the view. Admins see raw values. Everyone else sees encoded vectors.

As admin:

SELECT year_built, features
FROM <app_name>.app_runtime.sales_v
LIMIT 5;
-- year_built | features
-- 1985       | [350000, 1800, 3, 2]

As analyst:

SELECT year_built, features
FROM <app_name>.app_runtime.sales_v
LIMIT 5;
-- year_built | features
-- 1985       | [0.23, -0.41, 0.87, 0.12, ...]

Next steps

Once you’re running, see Configuration for scheduled retraining, GPU training, and compute scaling.