What's new#
Page classification now defaults to multi-label mode, allowing each page to be assigned multiple page classes simultaneously. Previously, pages could only receive one classification label (multi-class mode). Multi-class classification remains available as a configuration option.
Why it matters#
- Complex documents often have overlapping content types on single pages
- Bank statements - first page contains both account info AND transaction data
- Legal documents - pages mix contract terms, signatures, and exhibits
- Better extraction targeting - extract account_info AND transactions from the same page
Highlights#
- Single page can receive multiple relevant classifications
- More accurate representation of complex document structure
- Improved downstream extraction accuracy
- Backward compatible - multi-class mode still available
Technical details#
Multi-class (old default): Each page gets exactly one label - mutually exclusive
Multi-label (new default): Each page can get multiple labels - non-exclusive
Example: A bank statement's first page that contains account details AND the start of transaction history now gets both account_info and transactions labels instead of forcing a choice.
How to use#
Multi-label classification works automatically with existing page classification prompts:
doc_ai = DocumentAI()
result = doc_ai.parse_and_wait(
file="bank_statement.pdf",
page_classification=PageClassificationConfig(
page_classes=["account_info", "transactions", "summary"]
)
# Multi-label is now default - no config change needed
)
# Pages can now have multiple classifications
for page in result.pages:
classifications = page.classifications # Can contain multiple labels
To revert to multi-class behavior:
page_classification=PageClassificationConfig(
page_classes=["account_info", "transactions", "summary"],
classification_mode="multiclass" # Explicitly set multi-class
)
Status#
✅ Multi-label by default is live now. Existing prompts will automatically benefit from multi-label classification. 🚧 Choosing between mutli-label and multi-class is coming to the API and SDK soon.