What the documents look like
Payment risk control for cross-border standalone stores centres on card network disputes and the receiving account. Order fulfilment status is a separate matter. The working materials are dispute notifications, fraud order lists, invoices, and proof of carrier delivery. These arrive from payment backends, store order records, carrier acceptance records, and customer service email. Updates occur daily or per order. In store service and internal training, citation tracing means every outward statement can be traced back to a specific material, and each trace records the subject, object, status, and validity scope of that material.
Two features of this material shape the tracing step. Dispute notifications carry a status such as pending review, and that status is not an outcome. A draft sentence that asserts a completed action can be checked against the status field in the notification. The second feature is that retrieval is information access, not authorization. A search result, a form value, or an offered option does not establish that business verification, approval, a database write, or external synchronisation has occurred.
Constraints at the citation tracing step
The knowledge base search node outputs citations as an array, and the length can be zero. A zero-length array means the search path completes without a content hit. The retrieval output is handed to the answering stage, where each answer sentence is checked against the cited source text. Citation display is checked against a traceable sample, and the model's thinking process or the fluency of the answer does not replace source evidence.
Document timeliness is checked with a current version and a historical version of the same material, and whether a citation is adopted still depends on review of the valid version. Retrieval is not authorization. Form collection of values is not business verification, approval, a database write, or external synchronisation. The search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together.
Hypothetical paper exercise sample: the source text available for checking reads "Dispute notification number DB-2024-0117, object order SO-8823, status pending review, validity scope limited to 17 January 2024 through 31 January 2024." A draft sentence awaiting judgement reads "The dispute has been fully processed and the funds have been returned." The judgement point: the source states the status is pending review, the draft sentence claims full processing, and the two take mutually exclusive values on the same field, so the draft sentence is unsupported. The source does not state any fund return, so that part remains to be verified.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Comparison method |
|---|---|---|
| Associated knowledge bases | Knowledge bases sharing one vector model | Check the vector model identifier of each selected base and exclude those that differ |
| Citation limit | 3000 tokens as a hypothetical trial starting point | Compare against the length of the sample citation content and observe whether truncation appears |
| Minimum relevance | 0.5 as a hypothetical trial starting point | Compare against retrieval result scores and observe how many results remain after filtering |
| Search mode | Semantic search | Compare against full-text search results and observe differences on identifier-type queries |
| Question optimisation | Enabled | Compare against the disabled state and observe changes in retrieval results for follow-up turns |
| Result reranking | ReRank enabled | Compare against unranked results and observe changes in result positions |
The numeric values above are hypothetical trial starting points for this exercise. The reason for choosing them is that the sample citation content is short, and they are compared against truncation behaviour before any adjustment. The citation limit is measured in tokens, not a guessed item count. A relevance score is a filter threshold, not a category, entity, or permission filter, and it does not exclude unrelated businesses or unauthorized records.
Three common errors
- Treating retrieval output as authorization to execute. The input is the citation array returned by the search node. Compare the citation content against the business approval status. The observable result is that a citation exists while the approval record is absent. Retrieval is information access; approval is a separate business action.
- Treating a pending review status as completed processing. The input is the sample source text stating "status pending review". Compare the draft sentence "has been fully processed". The observable result is that the two take mutually exclusive values on the same field. A pending review does not demonstrate a lack of operating authority, and it does not equal completed processing.
- Associating knowledge bases that use different vector models. The input is the vector model identifier of two knowledge bases. Compare whether the identifiers match. The observable result is that the identifiers differ. Cross-model association is not permitted; check the identifiers and exclude those that do not conform, without running a cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the sample source text "Dispute notification number DB-2024-0117, object order SO-8823, status pending review". Compare the draft sentence "The dispute has been fully processed". The observable result is that the draft sentence and the source status field take mutually exclusive values, and the judgement is unsupported. The expected phenomenon is that the draft sentence is marked as pending verification. No claim is made that this was actually tested.
- Input the citation array output by the search node. Compare the array length against the downstream handling design. The observable result is whether the downstream path still completes when the array is empty. The expected phenomenon is that the downstream can handle a citation length of zero. A paper input of an empty array can be used to check the downstream design. No claim is made that a fabricated-rule retrieval necessarily returns an empty array, and no guarantee is made that the first result is always a hit, that all relevant content is recalled, or that an unrelated query returns nothing.
- Input the vector model identifiers of two knowledge bases. Compare whether the identifiers are the same. The observable result is that when the identifiers differ, joint association is excluded. The expected phenomenon is that knowledge bases with different models are not selected together. No claim is made that joint association would produce missed recalls or redundant results.
For the search mode and question optimisation items, the check is observational. Full-text search suits identifiers, model numbers, proper nouns, and error codes, while semantic search suits natural language questions. The overall retrieval flow may combine problem optimisation, semantic recall, full-text recall, image handling, RRF fusion, and ReRank, but the optional paths listed in the documentation flow diagram do not all execute on every request. Image retrieval requires a corresponding image input and model capability, and reranking requires the model and an enabled setting. RRF is a fusion process for mixed results, not a switch. Thinking budget is a reasoning budget and is not used as a citation token limit. Where the source material does not state a system error or an automatic business action, that remains unknown.