​

Extraction Citation

Today, we’re launching Extraction Citation, which allows you to trace any extracted field back to its location in the original document.

• Improves confidence in extraction accuracy
• Enables better debugging and fine-tuning of your extraction schemas
• Citation coordinates are available in API response as well

​

Long Document Processing

We’ve added support for handling documents of any size or orientation, particularly ultra-long or wide formats.

• Enhanced processing for extra-long receipts and wide documents
• Maintained high accuracy across unusual aspect ratios and complex layouts
• Smart handling of multi-page documents with continuous content flowing across pages

​

Confidence intervals

Confidence intervals have been a major work in progress for us over the last couple months and I’m excited to announce it’s finally live!

• We return confidence scores for all extracted values and each page of the OCR result.
• From the API, scores are returned in a separate confidence object. You can read more on the format in the extract documentation.

​

Improved schema builder

You can now upload multiple files to build and evaluate each template.

• Running Suggest schema will analyze all of the example files and attempt to build a generalized schema that covers the most common use cases.
• Between this and confidence intervals, it’s now way easier to test a variety of different document formats to maximize accuracy.

​

Better observability

We made some big updates to our Usage view! You can now click into each API request to view and download:

• A document preview
• OCR + Extracted values + Confidence scores
• The full JSON result

Plus a few more improvements:

• We did a revamp on our API documentation. Much cleaner now!
• Added the Mistral model to our OCR Benchmark.
• Improved observability and retry for web-hooks.