An invoice-extraction pipeline tested well, but in production about 15% of outputs are empty or nonsensical…

Question

An invoice-extraction pipeline tested well, but in production about 15% of outputs are empty or nonsensical. Investigation finds these invoices are scanned PDFs with no text layer, and the pipeline sends only the extracted text to Claude. What is the best fix?

Accepted Answer

Detect a missing text layer at input and send page images or OCR output instead, flagging files that stay unreadable

Answer

Add an instruction telling Claude to answer "unreadable" whenever the provided document text appears empty or garbled

Answer

Move those invoices to a more capable model, which is more robust to poor-quality or missing document content

Answer

Retry extraction up to three times whenever required fields come back empty, then send any remaining failures to a queue

An invoice-extraction pipeline tested well, but in production about 15% of outputs are empty or nonsensical. Investigation finds these invoices are scanned PDFs with no text layer, and the pipeline sends only the extracted text to Claude. What is the best fix?

Keep going

More Solution Design & Architecture questions