Accuracy benchmark · 十月 2026
We run our production extraction pipeline on real documents and compare every number, word and date with text checked by hand against the source. For numbers: 99% on the 50 pages held out from tuning.
“Correct” is the share of values in our output that appear exactly on the page. “Found” is the share of values on the page that appear in our output.
| 页 | Numbers correct | Numbers found | Words correct | Words found | Dates correct | Dates found |
|---|---|---|---|---|---|---|
| 所有页面 (82) | 98.1% | 97.3% | 96.2% | 90.2% | 99.5% | 99.5% |
| Held out from tuning (50) | 99.1% | 98.4% | 96.6% | 93.0% | 99.7% | 100% |
| 页 | Correct | Found |
|---|---|---|
| Digital PDFs (29) | 99.7% | 99.4% |
| 扫描件 (41) | 97.2% | 96.8% |
| PDFs with a scanner's text layer (9) | 98.2% | 98.6% |
| 手机照片 (3) | 96.4% | 92.3% |
| 页 | Correct | Found |
|---|---|---|
| Financial tables (statements, invoices) (32) | 98.7% | 97.5% |
| Plain text documents (11) | 92.9% | 94.3% |
Other layouts are not broken out here; every page counts in the overall results.
Of 284 transaction rows extracted from the statements in the set, the running-balance check flagged 1 for review. When a statement prints its balances, the preview also checks that opening balance + credits − debits = closing balance, and tells you when they match.
Upload a PDF, scan or photo and compare the preview with the original.
Try it on a document