Your supplier receives a purchase order with a special quantity instruction in the notes. The extraction system misses it, and an operator corrects the record. The immediate result may now be right. A different question remains: should the system change how it reads every future order from your company? A transaction exception should not quietly become a general interpretation rule.

An AWS document-processing architecture published on October 14 separates learning a new format from processing a known one. In its known-format path, a troubleshooter can fix extracted data without changing the reusable instructions. That distinction is useful for international order operations: a correction needs a defined reach, not just a successful outcome on the file currently open.

Diagnose the error before deciding what can be learned

An incorrect quantity can have several causes. A new table column may have shifted the extraction position. A note may amend only this particular order. A packaging unit may have been interpreted incorrectly. These are different problems even when the visible repair is simply replacing one value. Treating them as the same learning event makes subsequent behavior difficult to explain.

The reviewer needs the original page, the extracted field and the basis for the correction together. If the source is ambiguous, ask the sender. Keep the proposed interpretation separate from the original request while clarification is pending; do not overwrite the source merely to make validation pass. A model's most likely interpretation cannot settle a commercial disagreement about quantity, price or delivery. Extraction locates and structures the stated information; it does not give the system authority to decide a term that the parties have not made clear.

This matters on both sides of a China sourcing relationship. A buyer may send a special instruction expecting it to apply once. A supplier may save it as a permanent customer rule. The next order then receives an unintended interpretation, even though its document is normal. When sending an amendment, identify whether it changes the current order, a named template version or a continuing purchasing convention.

For the receiving team, keep an isolated exception in the current record. A repeated change in the sender's layout may justify a new format version, but the company name alone is too broad a match. Different departments can use different templates, and older documents can continue to arrive after a new one has been introduced. Similar appearance is a clue to investigate, not proof of identical meaning.

Test the files that should stay unchanged

Before promoting a correction into reusable instructions, state its scope plainly. “Read quantity from this column in this template version” can be tested. “Handle similar cases intelligently” leaves the affected population undefined. The proposed rule should identify the format it applies to and the evidence that distinguishes that format from its neighbors.

Next, select counterexamples. Include the previous template, an ordinary order without special notes and a document containing similar wording with a different purpose. Ask not only whether the new rule repairs the original failure, but whether it damages a field that was previously correct. One unusual purchase order is weak evidence for a broad permanent rule.

When the evidence is not yet sufficient, continuing to correct individual records can be the safer operating choice. There is no need to demonstrate automatic learning at every encounter. The objective is stable interpretation of orders, not the largest number of changes to the extraction instructions. More examples can clarify whether the issue is repeated structure or a genuinely one-off instruction.

Once a reusable change is accepted, give it an identifiable version and record which new documents use it. Do not silently recalculate already confirmed historical orders. If reprocessing is necessary, define the population and show the resulting differences to the responsible business team. Otherwise a technical update can look like an unexplained change in the buyer's commercial instructions.

Give the correction note a precise object as well. “Quantity corrected” is less useful than identifying the affected line and the source passage that supports the change. If the buyer later sends a revised order, retain the distinction between that document revision and the operator's earlier interpretation. The parties can then discuss the actual amendment rather than mistake a processing repair for a newly agreed commercial term.

Our discussion of retaining original values during metadata harmonization covers the related mapping problem. Here the additional question is whether a correction should affect one instance or future processing. Keeping those decisions separate lets buyers and suppliers investigate a disagreement without guessing whether the document, the extraction or the reusable rule changed.

Sources