Test a representative document, compare material fields with their source positions, and keep extraction defects separate from missing commercial terms. Fix or flag those defects before building the buying comparison.
In this guide
How should you record a shifted table cell?
This deliberately constructed defect is not an observed failure of Docling. It illustrates how to compare a source field, working copy, and corrected value. Keep the freight question open even after repairing its location. A text extractor cannot decide whether an unexplained shipping charge covers the whole order or a particular destination.
| Field / location | Authored source values | Faulty extraction | Review result |
|---|---|---|---|
| Item B unit price / p1 row2 | USD 12 per unit | USD 40 | Restore USD 12 from source row |
| Freight / p1 footer | USD 40; scope not stated | Attached to item B | Keep as quote-level amount; ask shipping scope |
| Item B quantity / p1 row2 | 10 | 10 | Unit subtotal 12 × 10 = USD 120; freight separate |
Check fields from the source outward
- Open the original document beside the extracted result and identify supplier, quote ID, revision, and page boundaries.
- Trace each product row's description, quantity, unit, currency, and price to its correct source location.
- Inspect decimal points, minus signs, currency symbols, and characters that OCR may confuse.
- Check shipping, tax, minimum order, payment, validity, and lead-time conditions, including footnotes and continuation pages.
- Classify discrepancies as extraction defect, source ambiguity, or source not stating a term.
- Correct a working field with its source reference or leave it unresolved; retain the raw extracted output for diagnosis.
Establish a review sample before batching
Record the layouts and defects found in the trial. Include examples with merged cells, continued tables, scanned text, and different currencies if those occur in your real source set. Use the observations to decide which fields require review on every document.
A count of successfully converted files tells you whether the software produced outputs. It does not establish that the commercial fields are correct. Keep a separate list of missing outputs and a separate list of records whose contents need review.
Choose a document that reveals the difficult layout
A short clean PDF may be a poor test for a folder containing scans, multi-page tables, and footnotes. Pick a permitted example containing the structures you actually receive. Keep its file version and page references with the trial output.
Use Docling or your chosen document application in your own environment. Selectable text, OCR, and table reconstruction can need different handling. An extraction that produces readable sentences can still detach a price from its row or a condition from its heading.
Ask for a discrepancy report with source positions
Review this extraction against supplied SOURCE PASSAGES/PAGES: [material]. EXTRACTED FIELDS: [rows with IDs] MATERIAL FIELDS: [currency, quantity, unit price, freight, tax, payment, validity] Return field-level discrepancies with source file/page, extracted value, source-supported value, and classification: extraction defect, source ambiguity, or term not stated. Keep product-row and quote-level conditions separate. Do not fill missing charges with zero or infer an exchange rate. Identify which fields need human source inspection.
Move checked terms into the buying decision
After the field review, compare like-for-like requirements and prepare supplier questions for unresolved terms. If the extraction created duplicate or shifted rows, resolve those before importing into the working table.