You submitted a batch of 10,000 invoices for extraction. 200 came back with validation failures. You want to retry the failed ones with adjusted prompts (you noticed they're mostly receipts in unusual formats). Your current plan: re-submit all 10,000 invoices to ensure consistency.
What's the right approach?
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.
custom_id exists precisely so you can correlate batch responses with their inputs and resubmit selectively. Re-submitting all 10,000 wastes 50% of the cost on documents that already extracted correctly, and creates the (small but real) risk that some now-correct extractions come back differently. Targeted resubmission with prompt adjustments is dramatically more efficient and exactly what custom_id is designed to support.
"Consistency" sounds like a virtue, but the 9,800 documents that already extracted correctly don't need to be reprocessed. The cost saving is significant; the risk of inconsistency from selective resubmission is small.
Real-time calls for 200 documents cost twice as much per token. If the prompt adjustment fixes the underlying issue, those 200 will process successfully in batch on the second attempt.
Discarding 200 documents because they're inconvenient isn't a reasoned acceptance of imperfection — it's giving up on real data. If they're recoverable with prompt adjustment, recovering them is the right move.