Exporters and suppliers in Ludhiana handle a steady stream of purchase orders, packing lists and invoices from buyers in many formats — PDFs, scanned pages, emails. A newer use of language models is to read these and fill the ERP automatically. When it works, it saves hours of data entry. When it fails, a wrong quantity or size breakdown goes straight into production planning.
Testing an extractor like this is closer to data testing than to chat testing:
- Field-level accuracy: for each document, compare every extracted field — buyer, style number, quantity per size, unit price, currency, delivery date — with a hand-checked answer key.
- Tax and codes: GSTIN, HSN codes and amounts must be exact; a near-miss is a failure, not a partial pass.
Read moreRead less
- Layout variety: include multi-page orders, tables that break across pages, handwritten corrections and poor scans.
- Missing data: if a field is absent, the system should leave it blank and flag it, not invent a plausible value.
- Dates and currencies: 03/04 can be March or April; USD, EUR and INR amounts must not be mixed.
For your portfolio, design twenty sample orders yourself in different layouts, build the answer key, and report accuracy per field. It is unglamorous, but it is exactly what a factory finance team would ask before trusting the tool.