What the documents look like
Retail checkout and loyalty materials centre on price and membership entitlement. They do not cover short-shelf-life stock. The working set for this category includes promotional price tags, membership points rules, daily checkout reports and exception refund slips. These records come from the POS system, the membership CRM, payment platforms and the price tag system, and they are refreshed daily or once per shift, so versions turn over frequently. Store checkout supervisors, CRM operations staff and finance auditors all draw on the same records, and mistakes surface as price complaints, cash shortfalls or overages, and disputes over membership entitlement.
For retrieval in store service and internal training, the practical consequence is that one question often has to be answered from several document types that share a common wording. A training answer about points accrual has to stay consistent with the price tag in force, the checkout report for the same shift and any exception refund slip raised during it. Retrieval returns cited content; it does not decide whether a given staff member is allowed to act on that content.
Constraints at the retrieval step
The knowledge base search node can reference one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be referenced together. Search mode is chosen per test question: semantic retrieval, full-text retrieval and hybrid retrieval are compared against the same question rather than assumed to suit all material. The citation cap is expressed in tokens, so long rules and short price tags have to be tested separately to see whether the cited content is truncated. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. Question optimisation performs coreference resolution and question expansion; the original question and the optimised question are compared to check whether the subject has been altered. Retrieval output is an array and its length may be zero, which means the downstream path still runs when nothing is found. Retrieval on its own does not complete authorisation and does not separate business entities or permission scopes.
Hypothetical paper exercise. The material text for checking reads: "This store's promotional price tags run from 1 June 2024 to 30 June 2024. Membership points accrue at one point per yuan of the amount actually paid. The daily checkout report must be reconciled against exception refund slips at the end of each shift." The handover request awaiting a decision reads: "Please synchronise the June promotional points rules for use in training newly hired checkout staff." The subject in this request is the store training side, the object is checkout staff, the status is pending review, and the effective scope is the June promotion period. The authorisation status of this request is not stated.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Referenced knowledge bases | Reference only the knowledge base containing promotional price tags and points rules | Check whether the vector model identifier matches; exclude knowledge bases on a different vector model |
| Search mode | Test semantic retrieval and full-text retrieval separately | Put the same question through both and compare which recall covers the points rules |
| Citation cap | Aim at covering the complete points rule | Compare a long rule against a short price tag to see whether cited content is truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Observe the filtering effect only under semantic retrieval or reranking |
| Question optimisation | Enable it for follow-up questions containing references and check the result | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Observe after a rerank model is available and enabled per settings | Compare candidate ordering before and after reranking; RRF remains a fusion step, not a switch |
Three common errors
- Treating a retrieval result as an executed authorisation. In the hypothetical exercise the handover request is pending review, so retrieving the points rule does not mean the training side has been granted synchronisation authority, and it does not mean checkout staff have been cleared to use it. The check is to input that request, compare its review status field, and observe whether an explicit approval record exists.
- Treating minimum relevance as a permission or entity filter. That value takes effect only under semantic retrieval or result reranking, and it does not exclude unrelated businesses or unauthorised records. The check is to input the same question, compare returned content before and after filtering, and observe whether records from other stores or unauthorised records still appear.
- Referencing knowledge bases across different vector models together. Referencing requires the same vector model, and different models cannot be referenced together. The check is to verify the vector model identifier of each knowledge base, exclude those that do not match, and record what was actually observed; a cross-model retrieval is not run to compare missed recall.
How to verify the setup
- Input the material text and the handover request from the hypothetical exercise, then compare the authorisation status field of the request against the presence of an approval record. The condition is met when retrieval returns the original points rule text while the request status shows pending review and no approval record exists.
- Input the same points rule question under semantic retrieval and under full-text retrieval, then compare the returned content. The condition is met when both recall paths cover the main body of the points rule and the cited content is not truncated.
- Input a follow-up question containing a reference, compare the original question with the search terms produced by question optimisation, and check whether the subject was altered. The condition is met when the optimised question retains the points rule subject and does not substitute another store or another period. If the input is an empty array, the downstream design can be checked for how it handles empty citation output; no claim is made that a fictional rule query must return an empty array.