Back to documentation

Confidence scoring and review

Karchu scores confidence at document, transaction and field level, and routes anything unclear to a review queue with the source document beside it.

Last updated 2026-09-11

Karchu does not claim to be right. It states how sure it is, at three levels, and holds back anything it is not sure about.

The three levels

Document confidence      0.98
Transaction confidence   0.96
Field confidence
  date        0.99
  merchant    0.95
  amount      0.99
  category    0.87

What sends a row to review

  • Row confidence below 0.6
  • A missing or unparseable date
  • A non finite or missing amount
  • A row that cannot be placed in the reconciled balance sequence
  • A document whose statement arithmetic does not settle
  • A document type classified below 0.7 confidence, which asks you to confirm the type

The review screen

Review shows the original document on one side and the extracted rows on the other. Selecting a row highlights the source text it came from. From there you can change the date, amount, merchant or category, delete a row that should not be there, add a row that was missed, and mark rows reviewed. Every change is recorded with the value before, the value after, who made it and when.

Limits

  • Source highlighting depends on the document carrying position information, so a low quality photograph may not highlight
  • Confidence is a measure of how clearly the value was read, not a promise the statement itself is correct

Questions

Is there one accuracy percentage for a document?
No. There are three levels: a document score, a score per transaction, and separate scores per field such as date, merchant and amount. A document can be strong overall and still hold one weak field.
What happens to a row below the threshold?
It is saved with review status requires_review rather than accepted. Unreadable values are stored empty, never filled with a guess.
Can I see where a number came from?
Yes. Every row keeps its page and the position of the source text, so selecting a transaction highlights the place on the original document it was read from, where the document allows it.