What you can do today
- Extract invoice supplier details, invoice numbers, dates, totals and line items using deterministic rules.
- Match named fields to literal labels and aliases, with text, decimal, whole-number, date and yes/no types.
- Review named-field previews beside the PDF and download JSON with source references; missing or ambiguous values remain explicit.
- Save reusable named-field schema versions in Organization knowledge.
- Choose exact column names for suitable aligned tables.
- Keep original strings, missing values, exceptions and evidence references with the result.
Choose the values you need from a purchase order
Define the order number, order date and delivery date by their printed labels. Review each candidate beside the PDF, then download a JSON preview that retains the original text and source references.
Supported scope and limits
- Named fields require literal labels in supported layouts. Their JSON preview is unapproved and is not saved to the invoice register.
- General table extraction requires aligned rows and exact single-word headers.
- Written instructions are retained as context; they do not currently drive arbitrary AI extraction.
- General nested JSON schemas, contracts and arbitrary document layouts are not supported.
Planned extensions
- Nested objects and repeating groups.
- Model-based extraction with evidence and explicit missing or conflicting states.
- Nested schema versions and broader document-specific evaluation sets.
How to check the result
- Compare field values with the source, including leading zeros.
- Check that line-item counts and page coverage are complete.
- Resolve missing or conflicting fields before approval.