Your expense-report extraction tool repeatedly fails on receipts where the merchant name is genuinely cut off in the photograph — the bottom of the receipt is missing in the scan. A teammate proposes adding a retry-with-error-feedback loop to fix the failure rate.
Will retry-with-feedback help here?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Retries cannot recover information that isn't in the source. When the merchant name is genuinely missing from the photo, no amount of feedback will produce it — the model can either keep failing honestly or start fabricating. Retry-with-feedback is excellent for fixing format errors, schema violations, and arithmetic mistakes, because those are processing failures on information that's present. They're wrong for informational failures where the data simply isn't there. Knowing the difference is one of the most important judgment calls in extraction system design.
Retries can't recover information the source doesn't contain. Asking the model "you missed the merchant name, what is it?" against a receipt that doesn't show one will produce either honest failure or hallucinated answers.
Higher temperature explores more interpretations of given input — it doesn't add information that isn't there. Cranking temperature on a missing-data problem produces more creative fabrications, not recovery.
If three attempts all fail because the data isn't there, the fourth will too. Multiplying retries on structurally impossible cases is a category error.