What the documents look like
In cross-border standalone-site payment risk control, copy review centres on card-network disputes and receiving accounts. The working material comes from payment platform back offices, standalone-site orders, carrier proof of delivery, and customer service email. The documents handled day to day are chargeback notices, fraudulent order lists, invoices, and logistics proof of delivery, with updates at a daily or per-order frequency. Citation tracing at the copy drafting step addresses one factual question: whether a given promotional or explanatory sentence can be directly supported by the wording in this material, and whether the effective scope of that wording covers the object and period the draft refers to.
This differs from order fulfilment work, which is organised around parcel status. Here the review object is a dispute record or a receiving account, and the evidence is a notice, an invoice, or a signed delivery document rather than a shipping event. The consequence of getting this wrong is a denied dispute, a frozen account, or interrupted advertising, so the review must stay tied to what the source documents actually state.
Constraints at the citation tracing step
The knowledge base search node outputs citations as an array, and the length may be zero. An empty array does not mean the workflow has errored or terminated; the output path still runs. Citation content must be passed to the answering step, where each item is compared against the source wording. A relevance score reflects textual similarity only. It does not perform permission or entity isolation, and retrieval itself is not authorisation.
The search scope must first be checked against the selected knowledge bases and the material version. Only knowledge bases using the same vector model may be associated together; knowledge bases with different vector models cannot be associated. Hybrid search merges multiple result paths through RRF, and this fusion is a description of the process, not a switch that can be turned on. Reranking and image retrieval are optional paths that require the corresponding model capability and must be enabled according to settings; their presence in the overall flow diagram does not mean every request executes them.
Hypothetical drill sample. Source wording one: "This batch covers order numbers A-2201 to A-2260, signed on 12 March 2025, and the carrier proof of delivery is on file." Source wording two: "The above orders correspond to invoice number INV-0312, issued on 13 March 2025." Source wording three: "Chargeback notice DP-0091 concerns order A-2233, notice date 2 April 2025, status pending evidence submission." The draft sentence to be judged: "All orders in this store are invoiced the day after signing, and all chargeback disputes have been fully resolved." The review task is to determine whether "all orders", "invoiced the day after signing", and "fully resolved" in that draft sentence can be supported by the three source sentences above.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Include only the payment risk control knowledge base using the same vector model | Check the base identifier and vector model identifier item by item; exclude bases whose identifiers do not match |
| Search mode | Hybrid search | Compare recalled item counts against semantic search and full-text search, observing differences in candidate coverage |
| Minimum relevance | Hypothetical starting point 0.35 | Compare against 0.20 and 0.50 for returned item counts and correspondence to matched source wording, recording actual observations |
| Citation limit | Hypothetical starting point 2000 tokens | Compare against 1000 and 4000 to see whether citation content is truncated and whether source wording stays complete |
| Result reranking | Enabled according to settings | Compare ranking positions against the disabled state, confirming rerank scores correspond to source wording |
| Question optimisation | Enabled according to settings | Compare query completion results against the disabled state, observing whether references are resolved |
The citation limit uses the documented token unit, not a guessed item count. The minimum relevance value applies only under semantic search or when result reranking is used, and it is a filter threshold rather than a final sorting rule.
Three common errors
- Treating the draft sentence "all orders in this store are invoiced the day after signing" as already supported. The material covers only order numbers A-2201 to A-2260, the signing date 12 March 2025, invoice number INV-0312, and the issue date 13 March 2025. It does not state a full order scope or a next-day invoicing rule, so this wording remains unsupported.
- Reading the "pending evidence submission" status of chargeback notice DP-0091 as a completed dispute. Pending evidence submission and fully resolved are different stages, and the draft sentence "all chargeback disputes have been fully resolved" lacks supporting evidence.
- Treating a similarity score or rerank score as a subject or permission filter. The score reflects textual relevance only; retrieval does not constitute authorisation and does not exclude unrelated business records.
How to verify the setup
- Input the three source sentences and the draft sentence above, then compare each clause of the draft against the source wording for subject, period, and field. When a clause has no corresponding statement with the same subject and same period in the source wording, judge it unsupported and record the gap.
- Input the original chargeback notice DP-0091, then compare the two status field values "pending evidence submission" and "fully resolved". When the status field values differ and are mutually exclusive, judge that part of the draft as conflicting with the source wording; if the source wording does not state a resolution outcome, keep it as pending verification.
- Input an explicit empty citation array
[], then compare the downstream answering step's handling behaviour. When the output path still runs and the answer does not fabricate sources, the condition is met; when the answer invents source wording or claims a reason for a retrieval failure, judge it non-conforming.
These checks describe what to compare and what to observe. They do not assert that any test, deployment, or approval has occurred, and any evidence gap remains a gap.