Enhancement: Extract and evaluate sensitivity metadata via stamp detection

[flefevre]

Description:

It would be useful to enhance the metadata extraction capabilities of Docling by integrating detection of document stamps (e.g., printed or scanned labels like "Confidentiel Défense", "Medical Data", etc.) to infer the sensitivity level of a document.

Motivation:
Some documents include visible stamps or markings that indicate their confidentiality or sensitivity level. Automatically detecting such stamps during metadata extraction would allow for better compliance with data handling policies, such as:

• Preventing further data extraction if a document is marked with high-sensitivity labels (e.g., "Confidentiel Défense", "Medical Data", "Classified", etc.)
• Tagging documents appropriately in downstream processes
• Ensuring ethical or legal handling of documents

Proposed enhancement:

• Implement a module that scans for common sensitivity-related stamps in documents (PDFs, images, OCR output, etc.)
• Maintain a configurable list of keywords/phrases indicating restricted content (e.g., "confidential", "secret", "medical", etc.)
• Flag or halt extraction processes when such terms are detected
• Possibly integrate with existing metadata fields or create a new sensitivity_level metadata field

Benefits:

• Better compliance with data protection regulations (e.g., GDPR, HIPAA, national security laws)
• Prevent unintentional leaks or processing of restricted content
• Add value for users working in government, healthcare, or legal domains

Happy to discuss implementation ideas