What the documents look like
Retail checkout and loyalty copy drafting centres on price and member benefit wording. The working materials are promotional shelf tags, member point rules, end-of-day checkout reports and exception refund slips. These come from the POS system, the member CRM, the payment platform and the shelf-tag system. Update cadence is daily or per shift, so any price, point multiplier or refund condition written into copy must correspond to the material in force for that period. Mistakes surface as price complaints, cash shortfalls or overages, and member benefit disputes, which makes traceability from each drafted sentence back to a specific source passage the central requirement.
This page covers one task only: drafting product and marketing copy in this category, and tracing each drafted sentence to its cited source. It does not cover identity verification, report analysis or any other task.
Constraints at the citation tracing step
The knowledge base search node returns cited content as an array, and the array length may be zero. An empty array does not mean the workflow has errored or stopped. The citation limit is controlled in tokens, not by a top-k item count. The minimum relevance value applies only under semantic retrieval or when result reranking is used. Hybrid retrieval merges multiple result paths through RRF, and reranking produces a 0-1 score that can be used for filtering. Retrieval is limited to the selected knowledge bases, and only knowledge bases sharing the same vector model can be associated together. Retrieval does not perform authorization. A similarity score is a relevance threshold and does not exclude unrelated businesses or unauthorized records.
Hypothetical paper exercise sample. Source material wording: "From 1 March 2025 to 31 March 2025, members who purchase Brand A cooking oil reaching 100 yuan accumulate points at double rate. The point redemption ratio remains 100 points to 1 yuan. On refund, points already granted are deducted back." Drafted sentence to be assessed: "In March, members who buy Brand A cooking oil reaching 100 yuan enjoy double points, and points can be spent directly as cash."
Differences to check: the drafted sentence omits the campaign end date and the refund deduction condition. The source material does not contain the phrase "spent directly as cash"; it states a redemption ratio of 100 points to 1 yuan. The drafted sentence changes how the redemption condition is expressed.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting value | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the promotion rules base and the member benefits base | The two bases must share the same vector model; check the identifier first and exclude any base that does not match |
| Search mode | Hybrid retrieval | Promotion codes and member tier names need exact hits, and semantically close content also needs recall |
| Citation limit | Set to 2000 tokens | Hypothetical trial starting value; compare citation completeness against 1000 and 4000 |
| Minimum relevance | Set to 0.5 | Hypothetical trial starting value; compare the number of remaining results after filtering against 0.3 and 0.7 |
| Result reranking | Enabled | When candidates are numerous, apply second-pass ordering and filter using the rerank score |
| Question optimization | Enabled | For follow-up turns, complete references; for standalone questions, compare with the setting disabled and observe elapsed time |
The citation limit is a token budget for cited content. It is not a thinking budget and the two must not be interchanged. RRF is a merge process for combining multiple retrieval paths; it is not a switch that is turned on. Image retrieval applies only when an image input is present and the model supports it, and reranking applies only when a rerank model is available and the setting is enabled. The optional paths shown in the overall retrieval flow are not all executed on every request.
Three common errors
- Judging the drafted phrase "points can be spent directly as cash" as supported by the material. The source states a redemption ratio of 100 points to 1 yuan, and the drafted sentence alters how the redemption condition is expressed. This is an unsupported rewrite.
- Treating the retrieval score as an authorization filter. The minimum relevance value is a 0-1 filtering threshold and does not exclude unauthorized records. Authorization decisions require a downstream implementation.
- Treating an empty citation array as a workflow error. The cited content output length may be zero, and downstream steps can still continue. The check is whether the downstream design handles an empty result.
How to verify the setup
- Input the three source sentences and the drafted sentence given above. Compare whether the drafted sentence contains "spent directly as cash" and whether it omits the campaign end date and the refund deduction condition. If the drafted sentence contains wording the material does not state, or lacks conditions the material does state, record it as unsupported. Do not treat an unstated condition as a contradiction; it remains unsupported or pending verification.
- Input the vector model identifiers of the promotion rules base and the member benefits base. Compare whether the two identifiers match. If they differ, exclude that base, record the observed result, and do not run a cross-model retrieval. Cross-model bases cannot be associated together; do not infer any retrieval outcome from an association that is not permitted.
- Input a clearly labelled paper empty array [] and check whether the downstream node continues execution without error. If the downstream design does not handle an empty result, record the gap. Do not claim that a fictional rules query necessarily returns [].
Each check is stated as an intended comparison, not a completed test. Where the material, the model identifier or the downstream design is not available, the evidence remains missing and the item stays pending.