Field Detection Accuracy

Explanation

Field Detection Accuracy checks whether the AI system finds and locates every field it is supposed to extract from a document, which matters in document-extraction because a field that is never detected, for example a missed PAN, date of birth, or income figure on a loan application, silently drops information and forces costly downstream errors or manual rework. It is measured by the Field Detection F1 Score, computed as 2*(Precision*Recall)/(Precision+Recall), which balances how many detected fields were genuine (precision) against how many true fields were caught (recall). To implement it, maintain a labelled ground-truth set of documents marking where each field appears, compare the model's detected fields against those labels to count true positives, false positives, and false negatives, and log precision, recall, and the resulting F1 as evidence; run this in pre-deployment validation and on sampled production documents. The threshold requires detection recall of at least 0.95 with a false-negative alert (proposed): because the control is recall-focused, the key concern is missed fields, and a recall breach raises a false-negative alert that prompts investigation of the extraction model and routing of affected documents to manual review. Its Immediate priority reflects how undetected fields propagate into downstream banking decisions.