AI_EXTRACT (Document AI legacy models)¶

Syntax¶

AI_EXTRACT ( model => <model> ,
            file => <file> ,
            responseFormat => [] )

Arguments¶

model => model

Specifies the Document AI Arctic-TILT model for extraction stored in the Snowflake Model Registry; for example, my_db.my_schema.my_model.

file => file

A FILE for extraction.

Returns¶

{
  "error": null,
  "response": {
    "invoice_items": [
      "NEW CRUSHED VELVET DIVAN BED",
      "Vintage Radiator",
      "Solid Wooden Worktop",
      "Sienna Crushed Velvet Curtains"
    ],
    "invoice_number": "123/20",
    "tax_amount": "77.57",
    "total_amount": "465.43 GBP",
    "vendor_name": "UK Exports & Imports Ltd"
  }
}

{
  "error": null,
  "response": {
    "table1": {
      "gross": ["10", "31", "10"],
      "item": ["apples", "banana", "pear"],
      "net": ["9", "30", "10"],
      "tax": ["1", "1", ""]
    },
    "table2": {
      "name": ["John", "Ana", "Lisa"],
      "surname": ["Smith", "Nixon", "Gonzales"]
    }
  }
}

Access control requirements¶

Users must use a role that has been granted the SNOWFLAKE.CORTEX_USER database role. For information about granting this privilege, see Cortex LLM privileges.

Additionally, you must have the OWNERSHIP privilege on the model.

Usage notes¶

• The model must be in the Snowflake Model Registry. To migrate models to the Snowflake Model Registry, go to Document AI UI, and when prompted, follow the instructions on the integration banner.

• The Document AI model should not have more than 100 entities.

• If not set explicitly, the latest available model version is used by default (the version set when the model was published or trained in the Document AI UI). To set the default version of a model, use the ALTER MODEL command as shown in the following example:

  ALTER MODEL my_model SET DEFAULT_VERSION = new_version;
  

  Copy

• Confidence scores are not supported.

• AI_EXTRACT uses token-based billing, which is different from compute time-based billing used in the <model_build_name>!PREDICT method. For more information on the AI_EXTRACT cost for Document AI legacy models, see the Snowflake Service Consumption Table.

  • Entity extraction cost is labeled as arctic-tilt-entity.

  • Table extraction cost is labeled as arctic-tilt-table.

Regional availability¶

The following regions are available:

• AWS Canada (Central)

• AWS EU (Frankfurt)

• AWS EU (Ireland)

• AWS US East (N. Virginia)

• AWS US East (Ohio)

• AWS US West (Oregon)

• Azure Australia East (New South Wales)

• Azure East US 2 (Virginia)

• Azure Southeast Asia (Singapore)

• Azure West Europe (Netherlands)

• Azure West US 2 (Washington)

If your region is not listed, use cross-region inference.

Examples¶

The following example extracts the features defined in the Document AI model:

SELECT AI_EXTRACT(
  model => 'my_db.my_schema.my_model',
  file => TO_FILE('@files_db.files_schema.files', 'agreement.pdf'),
  responseFormat => []
);

The following example extracts information from all files in a directory on a stage:

SELECT AI_EXTRACT(
  model => 'my_db.my_schema.my_model',
  file => TO_FILE('@db.schema.files', relative_path),
  responseFormat => []
) FROM DIRECTORY (@db.schema.files);

Legal notices¶

Refer to Snowflake AI and ML.