Receipt OCR missing fields: review and recovery checklist

Distinguish a missing value from a failed run

A blank result can mean the model found no reliable value, the selected cell was never run, the provider failed, or the prompt did not request the field clearly. These cases need different fixes.

• If there is no run activity, run the selected field manually.
• If the provider timed out or was cancelled, retry after checking the connection; the previous value should remain intact.
• If OCR text exists but a typed field is blank, refine only that field’s extraction instruction.
• If the source image is unreadable, retain an incomplete status rather than inventing a value.

Check image quality before changing the schema

Many extraction problems originate in the attachment rather than the table. Cropped totals, glare, extreme perspective, multiple receipts in one image, and low resolution can all produce incomplete results.

• Rotate the image so the text is upright.
• Use one receipt per attachment where possible.
• Include the merchant header and the section containing totals and taxes.
• Prefer the original image or PDF page over a compressed chat screenshot.

Rerun the smallest possible unit

Do not rerun every AI field when only one value is wrong. Select the affected row and field, preserve the verified cells, and rerun the narrow extraction step. This keeps the review understandable and reduces model cost.

For example, if Vendor Name and Plain Text are correct but Structured Data lacks currency, rerun the structured field with an explicit currency requirement instead of regenerating the vendor and OCR text.

Validate structured and financial values

• Compare merchant, address, date, currency, subtotal, tax, and total with the source image.
• Check that JSON is syntactically valid before another workflow consumes it.
• Do not infer a currency only from the reviewer’s location.
• Keep ambiguous categories editable and subject to human approval.
• Confirm there is exactly one final record for each source receipt.

Keep a short review record

For a repeated receipt process, record the source file, affected fields, visible failure, action taken, reviewer, and final decision. This does not need to become a heavy audit system. A compact note makes it possible to distinguish an unreadable source from a provider failure or a prompt that consistently misses the same value.

Review patterns across several receipts before changing the template for everyone. If one unusual layout fails, handle that row manually. If the same field is missing across clear receipts from several merchants, update and retest the field instruction with a small controlled set. Preserve both successful and incomplete cases so the new instruction is not optimized only for one convenient example.

• Record the original attachment name and the field that was rerun.
• Note whether the prior cell value was preserved, replaced, or corrected manually.
• Check the adapted prompt against more than one receipt layout.

When to leave the row incomplete

Leave the row visibly incomplete when the source does not contain the requested value, the image cannot support a reliable answer, or the value would affect accounting and cannot be independently checked. A reviewable incomplete record is safer than a plausible fabricated number.

See the completed ten-receipt example for the expected relationship between attachments, searchable fields, structured output, and the reviewer’s acceptance checks.