Replace plain-text extraction with a layout-aware pipeline capable of handling text, formulas, tables, and figures as distinct region types.
Scope:
Build a gold set of 150 hand-verified questions across all 12 subjects: text-only, figure-bearing, table-bearing, and NAT
Render source pages at 300 dpi; run layout detection to classify regions
Route each region type to the appropriate handler (text, formula, table, figure)
Escalate low-confidence regions to a vision-capable model with a strict output schema that includes an explicit refusal option
Refusals must route to a human review queue, never silently auto-fill
Do not scale until field-level accuracy exceeds 95% on the gold set
Exit criterion: 95%+ accuracy on the gold set, measured before any bulk run.
Replace plain-text extraction with a layout-aware pipeline capable of handling text, formulas, tables, and figures as distinct region types.
Scope:
Build a gold set of 150 hand-verified questions across all 12 subjects: text-only, figure-bearing, table-bearing, and NAT
Render source pages at 300 dpi; run layout detection to classify regions
Route each region type to the appropriate handler (text, formula, table, figure)
Escalate low-confidence regions to a vision-capable model with a strict output schema that includes an explicit refusal option
Refusals must route to a human review queue, never silently auto-fill
Do not scale until field-level accuracy exceeds 95% on the gold set
Exit criterion: 95%+ accuracy on the gold set, measured before any bulk run.