A document extraction prompt should return a field only with the page and region that support it, along with an explicit unknown or conflict state. OCR text may separate a value from its label or move a table cell into the wrong row. A prompt can help interpret the layout, but it cannot make a low-quality scan legible or settle a disagreement between signed pages by assertion. Preserve the raw document ID, page numbering convention, extraction version, and field-level evidence. Validate types and reconcile business rules outside the model before writing a system of record.
Make this comfortable
Document prompts: anchor each field to a page and resolve conflicts
Decision in practice
A warehouse packet contains delivery note DN-684. Page 2 lists 47 crates in the dispatch table; page 5 has a handwritten correction to 44, signed by a receiver but not by the driver. A text-only extraction joins the 44 with the wrong row. The assistant must return both observed quantities, their page anchors, and a conflict state. The receiving system blocks automatic inventory adjustment until an operator verifies whether the correction applies to this shipment. A clearer prompt may improve field placement, but it cannot manufacture the missing driver acknowledgement.
Document: DN-684; pages use visible printed numbers.
quantity_observations: 47 at page 2/table row 6; 44 at page 5/note box.
conflict: true; accepted_quantity: unknown.
Review reason: handwritten correction lacks driver acknowledgement.
Write gate: no inventory update until authorized reconciliation.Performance and operating cost
Reading P pages with C candidate fields produces at least O(P + C) extraction and validation work, while image processing and model calls dominate wall time. Keeping page crops for disputed fields increases storage and review time. Ask only for the fields needed by the workflow, and route ambiguous handwriting to a reviewer. A high field confidence score does not resolve a disagreement between two visible values. The expensive error is a silent wrong-row match that propagates into inventory, so measure field-to-region accuracy as well as text accuracy.
Common Mistakes
- Do not return a value without its page or region evidence.
- Do not treat OCR confidence as proof that conflicting pages agree.
- Do not update inventory from an unresolved extracted quantity.
Connected lessons
- Prompt engineering applications
- Prompt Engineering
- Document extraction: separate observed fields from inferred values
- Evidence IDs: make generated claims auditable against supplied records
- Human handoff: preserve evidence and the reason for uncertainty
- Image prompts: verify claims against named regions and crops
- Audio prompts: mark overlapping speech before assigning speakers
- Video prompts: disclose sampling gaps around short events
- Cross-modal evidence: keep conflicting observations separate
- Project: reconcile a warehouse evidence packet
- Multimodal prompt evidence decisions
Continue with: Study extraction prompts: preserve the result behind each claim.
Continue with: Email prompts: verify attachment version and coverage.
Continue with: Invoice prompts: anchor each extracted field.
prompt engineering
multimodal evidence