What the documents look like
Payment risk control for cross-border standalone stores centers on card network disputes and receiving accounts. The working materials come from the payment backend, the order system, and the carrier. A single order can carry a chargeback notice, an order record, an invoice, and a signed delivery receipt at the same time. These materials update daily and per order. The same order may show a dispute date in one document, an invoice date in another, and a delivery date in a third. The parties, amounts, dates, and status fields across these documents need to line up item by item.
The materials sit in different systems, so citation tracing has a specific job: for any answer or handover request, identify which source document and which version of that document the statement rests on, and whether that source falls inside a valid scope. A chargeback notice records a date and a disputed amount. It does not record whether a contract took effect or whether an authorization was granted. A delivery receipt with an empty signatory field does not show that delivery was confirmed by a named person. These distinctions matter because the consequence of a wrong reading is a lost dispute, a frozen account, or interrupted ad delivery.
Constraints at the citation tracing step
The knowledge base search node outputs an array of citations, and the length can be zero. A zero-length array does not mean the workflow has thrown an error or stopped; the output path still runs. The citation content then passes to the answering stage, where each item is checked against the source text one by one.
The retrieval scope is limited to the selected knowledge bases, and only knowledge bases that share the same vector model can be associated together. Identifiers must be checked first, and any knowledge base on a different vector model is excluded. Semantic retrieval and the minimum relevance score affect recall and filtering only. They do not perform permission isolation or entity isolation. Retrieval itself does not constitute authorization. A form collects values; it does not mean business verification, approval, a database write, or external synchronization has occurred.
Hypothetical paper exercise. The source materials read as follows. First: "Order A-4471 was signed for on 12 March 2026; the signatory field is blank." Second: "The payment backend shows a chargeback notice for this order dated 20 March 2026, disputed amount USD 480." Third: "Invoice INV-2098 states the buyer is Company B, amount USD 480, invoice date 5 March 2026." The draft sentence under review reads: "Order A-4471 has completed delivery with a clear signatory, so a chargeback rebuttal can be submitted on this basis." This sentence turns a blank signatory field into a clear one, and it is not supported by the source.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Include only the same-vector-model base holding the three materials above | Compare against the identifier of a different-vector-model base; exclude the mismatch |
| Search mode | Hybrid retrieval | Compare against semantic retrieval alone for hits on identifier-type fields |
| Minimum relevance | Hypothetical starting value 0.3 | Compare against 0.2 and 0.4 for recalled item count and irrelevant fragments |
| Citation limit | Hypothetical starting value 2000 tokens | Compare against 1000 and 3000 for citation completeness |
| Citation handoff | Pass the citation array to the answering stage for item-by-item checking | Compare against adopting the answer text directly for traceability |
| Material version check | Retrieve with the current version and a historical version separately | Compare whether returned sources point to a valid version |
The starting values above are exercise choices, not official defaults. The relevance score filters low-relevance results and applies only in semantic retrieval or when result reranking is used. It is not a category, entity, or permission filter, and it does not exclude unrelated businesses or unauthorized records. RRF is a fusion step that combines multiple recall paths; it is not a switch that is turned on. The citation limit is measured in tokens, not in a guessed item count.
Three common errors
- Reading a blank signatory field as completed delivery. Input: the delivery receipt text stating the order was signed for on 12 March 2026 with a blank signatory field. Comparison: the value of the signatory field against the draft sentence's claim of a clear signatory. Observable result: the two do not match, and the sentence is not supported. The evidence gap is recorded.
- Treating a chargeback notice date as contract effectiveness or granted authorization. Input: the chargeback notice text dated 20 March 2026 for USD 480. Comparison: the fields the notice actually carries against the authorization status field. Observable result: the notice states a date and an amount only; the authorization status remains unknown.
- Treating a retrieval hit as a verified source. Input: the citation array returned by the search node. Comparison: each array entry against the source material text. Observable result: the citation can be traced back, but business verification and rebuttal submission still require downstream implementation.
How to verify the setup
- Input the three source texts above together with the draft sentence. Compare the phrase "clear signatory" against the blank signatory field. The condition is met when the phrase is observed to be unsupported and the evidence gap is recorded.
- Input a query containing only the order number. Compare the returned citations against the source texts. The condition is met when each citation item can be traced back individually. If the citation array is empty, check the downstream design against a paper input of an empty array, without asserting that a rule will return an empty array.
- Input the identifiers of two knowledge bases on different vector models. Compare whether the identifiers match. The condition is met when the mismatched base is excluded, no cross-model retrieval is executed, and only the actual observation is recorded. Do not predict missed recall or redundant returns as a result of associating them.
For the material version check, retrieve with the current version and a historical version and compare whether the returned sources point to a valid version. Whether a citation is adopted still depends on review against the valid version. Image retrieval requires image input and a model with the corresponding capability; reranking requires a rerank model and must be enabled by configuration. The optional paths in the overall retrieval flow do not all execute on every request.