Documentation
This documentation describes the demo parser that actually runs on this site — not a future product. If the docs and the demo ever disagree, the demo is the truth; tell us at [email protected].
What the parser extracts
| Field | How it is found | Confidence when found |
|---|---|---|
| Vendor | First non-empty line of the document | 99% |
| Date | ISO date (2026-10-24), “24 Oct 2026”, or day-first numeric date (24/10/2026) | 99% / 99% / 90% |
| Currency + Total | A line starting “Total” (or “Amount due”) followed by a three-letter code and an amount | 99% |
| Tax ID | VAT-style patterns, e.g. NL003456789B01, DE123456789, GB123456789 | 99% |
| Category | Vendor keywords: hotel, taxi/cab, café/restaurant, software/labs; otherwise “Unclassified — review” | 92% / 60% |
| Line items | Lines ending in an amount, excluding date, tax-ID, invoice and total lines | 95% |
Checks that decide the verdict
1) Line-item sum vs stated total (must match to the cent). 2) A tax ID was recognised. 3) A date was parsed. If every check passes, the verdict is “Ready to post”. If any check fails, the verdict is “Routed to review queue”, and the missing fields are listed. Nothing is ever invented to make a check pass.
Limits
Plain text only, max 20,000 characters per extraction. No image OCR. No duplicate detection across documents in the demo (each extraction is independent). No live ledger posting — the posting button is a labelled preview. Numeric dates are read day-first; a US-style 10/24/2026 date would be misread, and would surface in review because the line-item and date checks are shown, not hidden.
Next: Getting started · Demo guide · Live demo.