Digitization
Confidence
Confidence scores tell you how sure a model is about each extracted field — the signal that drives what gets flagged for review and how much you can trust a given model over time.
Quick Start
- 1
Every extracted field gets a score
As a model extracts each field from a document, it attaches a confidence score reflecting how certain it is about that value.
- 2
Low-confidence fields are flagged
Fields below the confidence threshold are highlighted in Review so reviewers know exactly where to look first.
- 3
Corrections feed back into the count
Each field a reviewer corrects is logged — that's what shows up as Corrections on the Jobs and Review pages.
- 4
Use it to judge a model
A model with consistently high confidence and few corrections is a good candidate to trust with less review over time.
Reading confidence scores
Each field carries its own score rather than the document as a whole — a document can have several high-confidence fields and one low-confidence one, and only that field gets flagged.
- Confidence comes from the underlying Azure Form Recognizer model — it reflects that model's certainty, not NNIPA's own judgment of correctness.
- A field with high confidence can still be wrong if the model was trained on documents that look different from what you're feeding it.
- Watch the Corrections trend for a given model over time — rising corrections despite stable confidence can mean the source documents have drifted from what the model expects.