What the documents look like
Retail checkout and loyalty operations depend on four recurring document types: promotional price tags, membership point rules, daily checkout reports, and abnormal refund slips. The price tag system supplies item prices and member prices. The membership CRM holds point rules and point transaction records. The POS system produces the daily checkout report. The payment platform provides refund records. Store checkout supervisors, CRM operations staff, and finance auditors all work with these materials during order fulfilment and after-sales question handling.
Materials update daily or per shift. A single transaction may leave different versions across the current shift's checkout report, that day's promotional price tag, and the point transaction record. Citation tracing addresses one question: which source text, version, and validity scope supports a given answer or handover request, and whether that material can support the request at hand.
The knowledge base search node can select one or more knowledge bases that share the same vector model, and outputs a citation array whose length may be zero. An empty array means the output path runs through; it does not indicate that the workflow has errored or terminated. Citation content must be passed to the answering step, and each answer must be checked against the cited source text.
Constraints at the citation tracing step
The minimum relevance setting takes effect only during semantic retrieval or result reranking. It is a filter threshold and does not perform permission or entity isolation. The same-vector-model requirement applies only to knowledge base association; it does not extend to external systems. Hybrid retrieval merges multiple result paths through RRF, which describes a process and has no corresponding switch. Thinking budget controls model reasoning intensity and is unrelated to the citation token limit.
Material timeliness requires separating the current version from historical versions. Whether a citation is adopted still depends on valid-version review; retrieval itself does not constitute authorization. Image retrieval and reranking require the corresponding input and model capability; when these are not enabled, no assertion is made about their behaviour.
Configuration and implementation checks
Hypothetical paper exercise sample: a store's daily checkout report for the shift of 12 March 2025 records "member-day points accrue at double the base points; a single refund does not retroactively claw back points already granted." The promotional price tag system for the same day records "12 March member day, category A items at 10% off member price." A draft sentence to be assessed reads "after the member's refund on 12 March, points were clawed back per the rules." The following configuration items centre on this sample.
| Configuration or check item | Hypothetical exercise starting point | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Only the base containing the above report and price tag, sharing one vector model | Compare against base identifiers; exclude those with a different vector model |
| Search mode | Hybrid retrieval | Compare against the recall scope of semantic retrieval and full-text retrieval |
| Citation limit | Hypothetical starting point 800 tokens | Compare against the combined length of the sample report and price tag |
| Minimum relevance | Hypothetical starting point 0.3 | Compare against the 0-1 range of the rerank score |
| Question optimization | Off | Compare against the extra model-call time when on |
| Reply format | Default | Compare against a json_schema fixed-structure requirement |
These numeric values are starting points for this paper exercise, not official defaults.
Three common errors
- Treating a retrieved promotional price tag as the basis for authorizing a point clawback. The input is the price tag text; the comparison object is the report's statement that clawback does not apply retroactively; the observable result is that the two give different handling directions for the same refund scenario.
- Treating an empty citation array as a workflow failure. The input is the search node output []; the comparison object is the documented statement that length may be zero; the observable result is that the downstream step should still receive the empty array and the output path continues.
- Treating minimum relevance as a member or store permission filter. The input is a low-scoring citation; the comparison object is that citation's business ownership; the observable result is that the score level does not change whether it belongs to the current member.
How to verify the setup
- Input the sample report text and the draft sentence "after the member's refund on 12 March, points were clawed back per the rules." Compare whether the draft sentence and the report's "a single refund does not retroactively claw back points already granted" share the same subject, the same period, and the same field. Record a contradiction only when the draft sentence and the report give mutually exclusive values for the same field; when the report does not state a clawback operation, record it as unsupported.
- Input an explicit paper citation array []. Compare whether the downstream answering step continues to accept the empty citation. Observe whether the downstream step produces an unsourced answer or halts, and check the design accordingly. Do not claim that a given rule's retrieval necessarily returns [].
- Input the sample price tag text. Compare whether the price tag's validity scope covers the date and item category of the refund in question. The price tag can serve as a clawback basis only when it states that it applies to that refund slip; when it does not state this, record the evidence gap.
For bases across different vector models, check identifiers first and exclude those that do not match; do not run cross-model retrieval to compare missed recalls. Image retrieval and reranking require the corresponding input and model capability; when these are not enabled, no assertion is made about their behaviour.
This page addresses only retail checkout and loyalty / order fulfilment and after-sales question handling / citation tracing. Which materials, operations, or requests sit at what authorization state during operations handover cannot be settled by treating retrieval or a form as authorization execution.