ai_parse_document improvements for signatures and layouts

The ai_parse_document function now includes enhancements to isolate signature regions and recognize unstructured document layouts with higher accuracy.

Use cases

Imagine you have scanned contracts or invoices and need to extract specific content blocks, including validation of signed fields.

• Detect signature sections in PDFs or images.
• Extract structured elements from complex layouts.
• Save rendered page images to Unity Catalog volumes.

How to use

You can call the function directly by passing the binary document content in a SQL query.

SELECT
  path,
  ai_parse_document(
    content,
    map(
      'version', '2.0',
      'imageOutputPath', '/Volumes/catalog/schema/volume/directory/'
    )
  ) AS parsed
FROM my_catalog.my_schema.documents;

What happens in this code:

1. The source table provides the binary file content.
2. The ai_parse_document function processes the file using schema version 2.0.
3. The imageOutputPath parameter saves rendered pages to the specified Unity Catalog volume.
4. The structured result is returned as a VARIANT type containing detected elements such as text, tables, and signatures.

Limitations and requirements

• Requires Databricks Runtime 17.3 or higher.
• Supports PDF, JPG/JPEG, PNG, TIFF/TIF, DOC/DOCX, and PPT/PPTX file formats.
• 500-page limit per document when the pageRange parameter is omitted.