I’m putting together a workflow to extract line items from scanned invoices: product descriptions, quantities, and totals. Some scans are clean, while others have skewed pages or faint text. I’m unsure whether to send the document straight to a DeepSeek model or run OCR first and pass along the recognized text.

For anyone who has tested invoice extraction, did separating OCR from the language model make the results easier to check? I’d also be interested in how you preserve row boundaries so the model doesn’t accidentally combine neighboring items.