Accuracy methodology
What 98% accuracy means.
ExtractInvoice reports headline accuracy on core invoice fields: vendor, dates, totals, currency, document type, invoice number, terms, and PO number when those fields are present in the labeled ground truth. Line items are measured separately because receipt modifiers and zero-amount options can otherwise distort the headline number.
Measured fields
- Vendor name
- Invoice number
- Invoice date and due date
- Subtotal, tax, total, and currency
- Document type
- Payment terms when present on the invoice
- PO number when present on the invoice
How to read the claim
The headline number is a core field-level accuracy claim, not a promise that every document will be perfect. The benchmark runs below include totals, dates, vendor names, invoice numbers, currency, document type, terms, and PO numbers when present.
Line items are still extracted and tested, but priced sale rows and strict receipt-modifier reconstruction are reported as their own quality metrics. The product is designed around review: approve clean fields quickly, correct low-confidence fields before exporting, and keep the final accounting data under your control.
Current benchmark runs
These are reproducible internal spot checks from the labeled test corpus. The stress benchmark intentionally overrepresents files that are harder than clean digital PDFs, so it is useful for finding review cases rather than measuring only ideal uploads.
100-item coverage benchmark
98.5%
100 labeled invoices and receipts with stratified edge-case coverage
Core field accuracy; strict line-item accuracy was tracked separately at 89.6%
Admin Testing Center: seed=1293106513, count=100, stratified=trueRepeat coverage benchmark
99.2%
99 completed labeled invoices and receipts from the same seed
Core field accuracy; strict line-item accuracy was tracked separately at 90.5%
Admin Testing Center: seed=1293106513, count=100, stratified=trueWhat usually works best
Typed invoices, native PDFs, clear scans, and well-lit photos typically produce the strongest results. Standard vendor layouts also tend to produce higher confidence.
What might need review
These cases often still parse correctly. They are the cases where confidence scores and a quick human check matter most.
- Handwritten invoices
- Low-resolution scans or blurry photos
- Unusual table layouts
- Missing labels, cropped pages, or faded text
- Invoices with multiple currencies or manual corrections
Try it against your own invoices.
The best test is your actual workflow, with the vendors and layouts you see every day.