What the documents look like
Cross-border material for supermarket checkout and member promotions centres on promotional price tags, member points rules, daily checkout reports and exception refund slips. Price and member benefit are the core subjects; short-shelf-life stock does not belong to this page. Source systems named in the brief are the POS system, member CRM, payment platform and price tag system. Material updates on a daily or per-shift rhythm, and in a cross-border setting one promotion may exist in several language versions, so the tag wording, the points rule clause and the refund slip explanation each need their source version checked line by line. Checkout supervisors, CRM operations staff and finance auditors hand work over to one another, and at that point each item needs a clear authorization state. A retrieval result is not an approval and not an execution instruction.
Constraints at the citation tracing step
Citation tracing requires that every citation in an answer can be traced back to source text held in the knowledge base. A knowledge base search node can be linked to one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. An empty array means the output path still runs; it does not mean the workflow reported an error and stopped. The minimum relevance value is a filter threshold and takes effect only under semantic retrieval or result reranking. It does not act as a permission filter or an entity filter. Whether a retrieved item is adopted still depends on a manual review against the valid version.
Hypothetical paper exercise: a store price tag reads "Store member day, all drinks 10% off, event dates 1 March 2025 to 3 March 2025, points calculated on the amount actually paid." The member points rule reads "For discounted items on member day, points are calculated on the discounted amount; on refund, the corresponding points are deducted." A draft sentence reads "During the member day drink discount, points are calculated on the original price, and on refund no points are deducted." The question is whether this draft sentence agrees with the two source texts.
Configuration and implementation checks
The following items relate to this page's exercise. They list documented settings or implementation checks only. The numbers are starting values for this hypothetical exercise, with a reason and a comparison object.
| Configuration item | Hypothetical starting value | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Two bases sharing one vector model: price tag base and points rule base | Only bases with the same vector model can be linked together; compare against identifiers of bases using a different vector model and exclude those that do not match |
| Search mode | Hybrid retrieval | The tag contains dates and discount wording, the rule contains clause sentences; compare recall coverage against semantic retrieval alone |
| Minimum relevance | Hypothetical start 0.5 | Used to filter low-relevance fragments; compare against 0.3 and 0.7 and observe citation count and source coverage |
| Citation limit | Hypothetical start 2000 tokens | Controls total cited length; compare against 1000 and 4000 and observe whether rule clauses are truncated |
| Result reranking | ReRank enabled | Secondary ordering when candidates are numerous; compare against the disabled state and observe whether citation order sits closer to the source text |
| Question optimisation | Enabled | Multilingual follow-up questions often lack a subject; compare against the disabled state and observe whether the retrieval wording is completed |
Three common errors
- Treating minimum relevance as a permission or entity filter. Input is the retrieval request from the exercise above; compare whether returned citations are ordered only by similarity; observe whether records from another store or records without authorization appear. A relevance score does not perform isolation, and retrieval is not authorization.
- Treating an empty citation array as a workflow failure. Input is a paper-constructed empty array []; compare whether the downstream node still executes; observe whether it is misjudged as an error and stopped. The node output length may be zero, and an empty array is not an error report.
- Treating "calculated on the original price" in the draft sentence as equivalent to "calculated on the discounted amount" in the source. Input is the draft sentence and the two source texts; compare the value of the points calculation field; observe whether one field carries mutually exclusive values. Only a clear mutual exclusion in the same field counts as a contradiction; a field that is not stated remains to be verified.
How to verify the setup
- Citation source check. Input is the price tag text, the points rule text and the draft sentence given above. Compare the points calculation method in the draft sentence with the corresponding field in the source texts. Observe whether the draft sentence agrees with the sources. The expectation is that the draft sentence and the sources differ on the points calculation field, so the draft sentence is not supported by the source text.
- Empty result continuation check. Input is a paper-constructed citation array []. Compare whether the downstream answer step still executes. Observe whether the path runs through. The expectation is that the downstream step continues and the empty array is not judged as an error.
- Cross-model linking check. Input is two knowledge base identifiers that use different vector models. Compare whether the identifiers match. Observe whether the joint linking condition is met. The expectation is that a mismatch excludes the link, and no cross-model retrieval is run to compare missed recall.
All of the above is a paper check. No deployment or test has taken place. Missing evidence stays missing, and a pending review does not show that an operation lacked authorization.