What the documents look like
Copy drafting for supermarket checkout and membership promotions draws on four document families: promotional price tags, membership point rules, checkout daily reports, and exception refund slips. Price and membership benefit are the central objects. Short-shelf-life inventory does not enter this task. Store checkout supervisors, CRM operations staff, and finance auditors each supply a different side of the material, and the same product's price description may appear simultaneously in a price tag, a daily report, and a refund slip. Whether field definitions agree across those sources determines whether a draft sentence can be cited at all. Materials update daily and per shift, so retrieval must account for wording differences for the same product across shifts. The search scope is the whole knowledge base; a collection is only a classification device and does not affect search results.
This category differs from convenience-store fresh-food ordering, which handles short-shelf-life stock. Checkout promotion work centers on price and membership entitlement, and the consequence of error is a price complaint, a cash shortfall or overage, or a membership benefit dispute. The retrieval step serves copy drafting for goods and marketing text; it does not decide whether a draft sentence is legally compliant or whether a promotion is approved.
Constraints at the retrieval step
The knowledge base search node can be associated with one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Retrieval returns cited content only. It does not perform permission isolation or entity filtering, and a similarity score cannot be used to exclude records the requester is not authorized to see.
Semantic retrieval, full-text retrieval, and hybrid retrieval have different strengths. Queries for numbers, model codes, and proper nouns rely more on full-text hits. Natural-language phrasing relies more on semantic recall. Hybrid retrieval merges multiple result paths through RRF, and reranking requires an available model and must be enabled in settings. The citation cap is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval.
A hypothetical paper exercise sample: Material text one reads "From 1 June 2024 to 30 June 2024, a store's A-class laundry detergent promotional price tag states: member price 19.9 yuan, non-member price 25.9 yuan." Material text two reads "Membership point rule article 3: member-price goods accrue points on the amount actually paid, and corresponding points are deducted back on refund." The draft sentence to be judged reads "This laundry detergent's June member price is 19.9 yuan, and points are not deducted back after a refund." The comparison objects are the promotional price tag and the point rule, the period is June, and the fields are member price and point deduction on refund. The draft sentence and material text two take opposite values on the point-deduction field, which is a mutually exclusive value on the same field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Promotional price tag base, membership point rule base | Both bases must use the same vector model; check the identifier first, then associate |
| Search mode | Hybrid retrieval | The draft sentence contains the proper expression "member price" plus natural-language description; compare the difference between semantic and full-text hits |
| Citation cap | Hypothetical starting point 2000 tokens for this exercise | Price tag and point rule entries vary in length; compare whether cited content is complete at different caps |
| Minimum relevance | Hypothetical starting point 0.5 for this exercise | Label relevant and irrelevant samples first, then compare whether irrelevant fragments decrease after threshold adjustment |
| Question optimization | Enabled | The draft sentence contains a reference; compare whether the subject is altered incorrectly between the original and optimized question |
| Result reranking | Enabled | When candidates are numerous, compare ranking changes before and after reranking |
The citation cap uses the documented unit of tokens. The minimum relevance value is a filter threshold and not a final ranking rule. When question optimization, hybrid retrieval, image retrieval, or result reranking is enabled, the final result may pass through multi-path fusion and does not necessarily follow a single vector similarity order.
Three common errors
- Treating the minimum relevance value as a permission or entity filter. Input is the draft sentence and the price tag base; the comparison object is the similarity score against the record ownership field; the observable result is that low-scoring records are filtered while unauthorized records may still be recalled. The relevance setting does not perform the function of excluding unauthorized business records.
- Forcing association of knowledge bases built on different vector models. Input is two knowledge base identifiers with different vector models; the comparison object is whether the vector model identifiers match; the observable result is that association should not proceed when the identifiers differ. The documentation does not describe the runtime result of such an association; only the identifier check and the exclusion action are recorded.
- Treating a draft sentence awaiting review as an approved copy. Input is the draft sentence and the review record; the comparison object is the review status field; the observable result is that a pending status cannot be taken as the copy having taken effect. Review approval, copy publication, and external display are three distinct facts.
How to verify the setup
- Input material text one, material text two, and the draft sentence from the hypothetical sample. Compare the draft sentence's value on the point-deduction field against material text two. The observable result is that when the two take mutually exclusive values, the draft sentence is judged unsupported and must return to the material for verification. The expectation is to identify this field contradiction; no actual test is claimed.
- Input the draft sentence from the hypothetical sample. Compare the vector model identifiers of the associated knowledge bases. The observable result is that joint association is permitted only when the identifiers match. The expectation is to exclude bases whose identifiers do not match, without running a cross-model retrieval to compare missed recalls.
- Input an explicit paper empty result
[]. Compare how the downstream node handles a cited-content length of zero. The observable result is that the output path still runs through. The expectation is to check the downstream design's handling of an empty citation, without claiming that a fictional rule retrieval necessarily returns[].
Missing evidence remains missing. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where a document field is not provided, the value is unknown.