What the documents look like
Cross-border standalone-site payment risk work centres on card-network disputes and receiving accounts, so the working materials differ from ordinary parcel-status fulfilment questions. The documents handled day to day include payment chargeback notices, fraudulent-order lists, invoices, and proof of delivery signatures. These materials sit across Stripe or PayPal backends, Shopify orders, carrier signature records, and customer-service email, and they update daily or per order. During an operations handover, both sides need to check the source, version, and current authorisation state of each item. Retrieval output and form entries supply candidate content only; they do not constitute operational authorisation.
Constraints at the citation tracing step
The knowledge-base search node outputs a citation-content array whose length may be zero. An empty array does not mean the workflow has errored out or terminated. Linked knowledge bases must share the same vector model before they can be linked together; bases with different vector models must not be linked jointly. The minimum relevance value takes effect only under semantic retrieval or result reranking, and it acts as a filtering threshold. It does not perform permission or entity isolation. Retrieval is the process of obtaining candidate content, and business verification, approval, database writes, and external synchronisation each require their own downstream implementation.
Hypothetical drill sample: a standalone-site risk specialist hands over a batch of chargeback materials on 10 March 2025. The retrieval scope is the already-selected knowledge base A and knowledge base B, and both bases carry the same vector-model identifier. The source text supplied for checking reads: "Chargeback notice number DR-0310-07, target order SO-8841, status awaiting supplementary proof of delivery, valid until 20 March 2025." A second source text reads: "Fraudulent-order list, 10 March 2025 version, contains order SO-8841, marked pending manual review." The draft sentence to be judged reads: "This order has completed chargeback evidence submission, proof of delivery has been submitted, and no follow-up is needed after handover." The judgement must check whether the materials state that proof of delivery was submitted, whether the status has moved from awaiting supplement to completed, and whether the subject and period of the draft sentence match the materials.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Knowledge base A and knowledge base B, both with the same vector-model identifier | Check whether the two vector-model identifiers match; exclude any base whose identifier differs |
| Search mode | Hybrid retrieval | Compare recall of chargeback notice numbers and natural-language questions against semantic retrieval and full-text retrieval separately |
| Citation limit | Hypothetical trial starting point of 2000 tokens for this page | Compare against 1000 tokens and observe the number of cited passages and truncation |
| Minimum relevance | Hypothetical trial starting point of 0.3 for this page | Compare against 0.5 and observe changes in candidate count; this value is not used to exclude unauthorised records |
| Question optimisation | Enabled | Compare retrieval results in multi-turn follow-up scenarios against disabled; enabling adds one model call |
| Result reranking | ReRank enabled | Compare ordering against disabled; a rerank model must be available and enabled in settings |
Three common errors
- Treating retrieval output as an authorisation credential. The input is the citation array emitted by the node; the comparison object is the operational authorisation state in the handover request; the observable result is that cited content exists while the materials state no approval or effective date, in which case authorisation remains unsupported.
- Writing a pending-review status as completed. The input is the "pending manual review" marker in the fraudulent-order list; the comparison object is the draft sentence "has completed chargeback evidence submission"; the observable result is that the same subject, same period, and same field hold mutually exclusive values, which is a contradiction.
- Forcing knowledge bases with different vector models into a joint link. The input is the vector-model identifiers of knowledge base A and knowledge base B; the comparison object is the same-vector-model requirement; the observable result is that when identifiers differ, joint linking must not proceed, and only the identifiers are recorded and the mismatched base excluded.
How to verify the setup
- Input the chargeback notice source text and the draft sentence from the hypothetical sample; compare "awaiting supplementary proof of delivery" against "proof of delivery has been submitted"; the observable result is that the status field holds mutually exclusive values, so the draft sentence remains unsupported.
- Input the fraudulent-order list source text and the handover request from the hypothetical sample; compare "pending manual review" against "no follow-up needed"; the observable result is that the request lacks evidence of completed review, so the item stays pending verification.
- Input an explicit paper value of an empty array
[]; compare it against the design of the downstream answering step; the observable result is that the downstream step accepts the empty citation, and no claim is made that retrieval will necessarily return an empty array. All three verification items use the hypothetical source text as input, and no claim is made that they have been tested.