What the documents look like
In retail checkout and loyalty promotion work, comparison materials come from the POS system, the membership CRM, payment platforms, and the price tag system. The documents handled day to day are promotional price tags, membership point rules, checkout daily reports, and abnormal refund slips. These materials are produced daily or per shift, and price and membership benefits are the focus of reconciliation. Short-shelf-life inventory is outside this task.
A promotional price tag states the item, the promotional price, and the effective interval. A membership point rule states the point target, the multiplier, and the validity period. A checkout daily report records the shift, the cashier, and the amount. An abnormal refund slip records the refund target, the amount, and the approval status. The four material types update on different rhythms, and the same transaction may sit at different processing stages in the daily report and the refund slip.
Constraints at the retrieval step
Retrieval only locates content relevant to a question. It does not decide whether material has been authorized, whether it has taken effect, or whether it has been approved. The promotional price on a tag and the transaction price in a daily report belong to different fields, so a numeric difference alone does not establish a violation. The multiplier in a membership point rule and a point reversal in a refund slip belong to different objects and cannot substitute for each other.
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases on different vector models cannot be associated together; the vector model identifier must be checked first, non-matching bases excluded, and the actual observation recorded without inferring runtime results from it. Referenced content is output as an array, and its length can be 0, meaning the downstream chain still proceeds when nothing is found. The reference cap uses tokens as its unit. 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 an extra model call.
Hypothetical paper exercise sample: the source material for comparison is three sentences. Sentence one: "This store, from March 1, 2025 to March 31, 2025, sets the promotional price of item A at 9.9 yuan, tag number QJ-031." Sentence two: "The membership point rule states that item A accrues points at a multiplier of 1 during the promotional period, and points are credited on the day after the transaction." Sentence three: "The March 12 checkout daily report records item A's transaction price at 9.9 yuan, cashier Li, morning shift." The draft sentence to be judged is: "Item A's March 12 transaction price matches the promotional price, membership points have been credited according to the rule, and this transaction shows no discrepancy." This draft sentence covers two facts, price and points. The point crediting result is not stated in the material, so that part is unsupported.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the single base holding promotional price tags and membership point rules | That base shares the vector model of the question under judgment; a base on a different vector model is excluded after comparing identifiers |
| Search mode | Hybrid retrieval | Tag numbers contain letters and digits, so semantic and full-text retrieval are used together; compare tag-number hits against semantic retrieval alone |
| Reference cap | Set to 800 tokens as the assumed starting point for this exercise | The tag and point rule together are about 600 tokens; compare against 400 tokens to see whether content is truncated |
| Minimum relevance | Set to 0.3 as the assumed starting point for this exercise | Label relevant and irrelevant samples first; compare against 0.5 to see whether irrelevant fragments are filtered out |
| Question optimization | Enabled for this assumed exercise | Follow-up questions involve coreference; compare against the disabled state to check whether the subject is altered incorrectly |
| Result reranking | Enabled for this assumed exercise | Reranking applies when candidates are numerous; compare against the disabled state for ordering on number-type questions |
Three common errors
- Treating a retrieval hit as authorization confirmation. When retrieval returns the text of a promotional price tag, that only shows the text was recalled. It does not show the promotion was approved or has taken effect. The effective interval and approval field on the tag should be checked, and a missing field recorded as pending verification.
- Treating a numeric difference between different fields as a direct contradiction. The transaction price in the daily report and the promotional price on the tag are price fields; point crediting is another field. Comparison must use the same subject, the same period, and the same field. Only mutually exclusive values in the same field constitute a contradiction.
- Treating a pending-review status as a lack of operating authority. A refund slip marked pending review shows the approval process is incomplete. It does not show the checkout supervisor lacks operating authority. Signing, reaching cooperation, and contract effectiveness are three distinct facts and should be checked separately.
How to verify the setup
- Input the three hypothetical source sentences above and the draft sentence, then compare "membership points have been credited according to the rule" in the draft against sentence two of the source. The source states only the rule and does not state the crediting result. The observed outcome should be that this clause is unsupported and requires additional material or a pending-verification mark.
- Input the price portion of the draft sentence and compare the daily report transaction price of 9.9 yuan against the tag promotional price of 9.9 yuan. The two share the same subject, the same period, and the same field, and the values match. The observed outcome should be that the price portion is supported while the point portion remains pending verification.
- Input a query containing only tag number QJ-031 and compare what hybrid retrieval returns against semantic retrieval alone. The observed outcome should be that hybrid retrieval more readily matches the number. If neither mode returns the tag number, record it as a miss and do not infer a cause.
The knowledge base search node may associate one or more knowledge bases sharing the same vector model. Search mode should be compared across semantic, full-text, and hybrid retrieval for the test question, without presuming one mode suits all material. The reference cap is measured in tokens; test with material of different lengths to see whether content stays complete, then adjust the cap. Minimum relevance takes effect only under semantic retrieval or when result reranking is used; label relevant and irrelevant samples before tuning the threshold. Question optimization performs coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject has been altered incorrectly.