What the documents look like
Retail checkout and loyalty material centres on price and member benefit. It does not cover short-shelf-life stock. The source systems named in the brief are the POS system, the member CRM, the payment platform and the price tag system. The daily documents are promotional price tags, member points rules, checkout daily reports and exception refund slips. Updates arrive daily or per shift.
For a cross-border, multilingual retrieval task, the practical consequence is that the same promotion can exist as a Chinese price tag and an English points rule, and the two texts need to be compared on the same subject, the same period and the same field. A retrieval result should keep the original wording so staff can check it by hand. Retrieval returns candidate text; it does not decide whether a price is effective or whether a member is entitled to a benefit.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. The check is to read the vector model identifier for each library, exclude those that do not match, and record what was observed. A cross-model search is not run in order to compare missed recalls.
Search mode can be semantic, full-text or hybrid. The mode is compared against a test question rather than assumed to suit all material. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Question optimisation performs coreference removal and question expansion; the original question and the optimised question are compared to check whether the subject was altered. Reranking requires a rerank model and must be enabled by configuration. Image retrieval requires an image input and the relevant model capability. Retrieval and recall do not perform permission or entity isolation, and a similarity score is not an authorisation decision.
Hypothetical paper test case. A store checkout supervisor submits a promotional price tag whose source text reads: "1 March 2025 to 31 March 2025, members buying Brand A laundry detergent, original price 39.9 yuan, member price 29.9 yuan, limit 2 items per person." A separate English member points rule reads: "Members earn 1 point per 1 yuan spent on A brand products during March 2025." The draft sentence to be assessed is: "Members buying Brand A laundry detergent in March 2025 receive double points." Neither source states double points. The draft sentence is unsupported by the supplied text.
Configuration and implementation checks
The table below concerns this page's sample. Values are hypothetical starting points for this paper exercise, not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the Chinese price tag library and the English points rule library that use the same vector model | Read the vector model identifier for both libraries, exclude any that differ |
| Search mode | Test semantic search and full-text search separately first | Use the sample question to compare whether returned citations contain the price and points source text |
| Citation limit | Set to 800 tokens for this exercise | Test with price tag and points rule material of different lengths, compare against a shorter limit for completeness |
| Minimum relevance | Set to 0.3 for this exercise | Mark relevant and irrelevant samples first, then adjust the threshold and observe whether unrelated fragments are filtered |
| Question optimisation | Off first, retrieve with the original sample question | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Not enabled first | Confirm whether a rerank model is available and enabled by configuration; observe the original ordering while it is off |
Three common errors
- Linking knowledge bases that use different vector models. Input: two knowledge bases whose vector model identifiers differ. Compare whether the identifiers match. Observable result: a mismatched library is excluded and the observation is recorded. Linking them anyway is a configuration error; no claim is made about what a cross-model search would return.
- Treating minimum relevance as a permission or entity filter. Input: the sample question. Compare whether returned content includes promotional material from an unauthorised store. Observable result: the similarity score is not used as an authorisation decision. Minimum relevance filters low-relevance results only; it does not exclude unrelated businesses or unauthorised records.
- Treating a draft promotional price tag sentence as an effective price. Input: the draft sentence and the two source texts. Compare whether double points appear in either source. Observable result: the draft sentence is unsupported. The source text does not state double points, so the draft cannot be treated as effective.
How to verify the setup
- Retrieve against the Chinese price tag library and the English points rule library with the sample question "March 2025 Brand A laundry detergent member price". Input: the sample question and the two libraries. Compare whether the returned citations contain "29.9 yuan" and "limit 2 items per person". Observable result: the citations contain that source text. If the output is an empty array, record the gap; do not claim that a fictional rule would necessarily return an empty array.
- Compare the draft sentence "Members buying Brand A laundry detergent in March 2025 receive double points" with the two source texts. Input: the draft sentence and the two source texts. Compare whether double points appear in either source. Observable result: the draft is judged unsupported and the evidence gap is retained.
- Test the same question under semantic search and full-text search. Input: the sample question. Compare the citations returned by the two modes. Observable result: semantic search may match "March 2025" and "1 point per 1 yuan" in the English points rule, while full-text search may match "29.9 yuan" in the Chinese price tag. The two modes are expected to differ in emphasis; a first-position hit or full recall is not guaranteed.