1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
51 small wins to finish your pathNext question →
How do you evaluate structured data extraction (e.g. fields from invoices or contracts)?
30-second answerSay your answer out loud first, then reveal.
Field-level outcomes
| Outcome | Meaning |
|---|---|
| Correct | Extracted value matches ground truth (after normalisation) |
| Wrong | Value present but incorrect |
| Missed | Ground truth has value; extraction empty |
| Hallucinated | Extraction has value; ground truth empty |
Matching rules by field type
- Invoice number, GSTIN, amounts → exact after normalisation (₹1,20,000.00 = 120000).
- Dates → parse to ISO; handle DD/MM vs MM/DD.
- Vendor name → fuzzy match (case, punctuation, "Pvt Ltd" vs "Private Limited").
- Line items → align rows (e.g. Hungarian matching on description + amount), then compare fields.
Reporting
| Field | Accuracy | Missed | Hallucinated | Criticality |
|---|---|---|---|---|
| Total amount | 98.5% | 0.5% | 0% | High |
| Due date | 94% | 4% | 1% | Medium |
| Line items (row F1) | 89% | — | — | High |
Also measure: straight-through rate (documents with all critical fields correct), and confidence calibration (are low-confidence fields the ones that are wrong?) to set review thresholds.
Related
This is what real progress feels like.