What the documents look like
Price and display checks happen during day-to-day trading, and the recurring question is whether standards keep being met after opening support has ended. The material comes from supervisor systems, POS price lists, store cameras and head-office promotion calendars. The working documents are store-visit photographs, price tag lists, display standards and rectification notices. Operating reports and exception reviews organise this material weekly or monthly into a retrievable chain: price tag lists for comparing the price shown in a store against the promotion calendar, store-visit photographs and display standards for judging display execution, and rectification notices for tracking closure. The retrieval step recalls these documents by subject, store, period and field so that a review can compare them. Retrieval itself performs no authorisation decision.
Constraints at the retrieval step
The knowledge base search node can associate one or more libraries that use the same vector model. Libraries using different vector models cannot be associated together. Search modes include semantic retrieval, full-text retrieval and hybrid retrieval; hybrid retrieval merges the semantic and full-text result sets through RRF, which is a fusion process and not a switch. The minimum relevance value is a filter threshold and applies only in semantic retrieval or when result re-ranking is used; it does not perform permission, entity or store isolation. The citation limit is measured in tokens. The citation output is an array and its length may be 0, so the downstream chain still runs with an empty result. Question optimisation performs reference resolution and question expansion and makes an additional model call. Image retrieval requires image input and the corresponding model capability, and re-ranking requires a model and is enabled according to settings; both are optional paths and do not run on every request.
Hypothetical paper exercise: a chain operator intends to review promotion-calendar execution for one store in the East China region for March 2025. The source material for comparison reads: "The head-office promotion calendar states that from 1 March 2025 to 31 March 2025 the recommended retail price of Product A is 9.9 yuan. The store price tag list states that on 15 March 2025 the store marked Product A at 12.9 yuan. The rectification notice states that the store received the notice on 20 March 2025 and the status is awaiting store confirmation." The draft sentence to be judged is: "In March 2025 this store's price execution for Product A did not conform to the head-office calendar, and the rectification has been completed." The judgment requires the same subject, the same period and the same field: the calendar and the price tag list both concern the price field of Product A in March 2025, and their values are mutually exclusive, which makes a contradiction. The rectification notice status is awaiting store confirmation and does not state completion, so the "rectification has been completed" part of the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated libraries | Price tag list library, display standard library, rectification notice library | All three must use the same vector model to be associated together; check the vector model identifier first, exclude non-matching libraries and record the observation |
| Search mode | Hybrid retrieval | Calendar numbers and product names are exact lookups while semantic questions also need coverage; compare against semantic-only and full-text-only results |
| Citation limit | 2000 tokens as a hypothetical starting point for this exercise | This value is a hypothetical starting point, not an official default; compare against 1000 and 4000 tokens to see whether the cited content is complete |
| Minimum relevance | 0.4 as a hypothetical starting point for this exercise | This value is a hypothetical starting point; label relevant and irrelevant samples first, then compare filtering against 0.2 and 0.6 |
| Question optimisation | Enabled | Used for reference resolution of "this store, this product"; compare the original question with the optimised question and check whether the subject was altered |
| Result re-ranking | Enabled according to settings | A re-ranking model is required; compare candidate ordering before and after re-ranking, and use the re-ranking score for filtering |
Three common errors
- Treating minimum relevance as a store or permission filter. Incorrect input: using a similarity threshold to exclude records from other stores. Compare the store identifier in the retrieved results against the target store and observe whether records from non-target stores are still returned; relevance does not perform entity isolation.
- Treating a rectification notice awaiting store confirmation as completed. Incorrect input: the draft sentence "the rectification has been completed". Compare the status field stated in the rectification notice and observe whether completion is stated; awaiting review does not demonstrate a lack of operating permission, and awaiting confirmation does not constitute completion.
- Associating libraries that use different vector models. Incorrect input: associating the price tag list library and the display standard library when they use different vector models. Check the vector model identifier first, exclude non-matching libraries and record the actual observation; do not run a cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the hypothetical source text above together with the draft sentence, compare the price field of Product A in March 2025 between the calendar and the price tag list, and observe whether they share the same subject, period and field with mutually exclusive values; if so, the contradiction holds and the completion part of the rectification remains to be verified.
- Input a paper empty result array [] and compare how the downstream node is designed to handle a citation content length of 0, observing whether the chain runs through; do not claim that a fictional rule retrieval necessarily returns [].
- Input a retrieval question containing a store identifier, compare the store identifiers in the returned citations against the target store, and observe whether records from non-target stores are excluded; retrieval performs no authorisation, and unsupported or pending items must remain as evidence gaps.