What the documents look like
Near-expiry markdown handling follows the shelf-life cycle of stock already on display. It differs from weighing and sorting, where each order is weighed item by item. The working materials are near-expiry lists, markdown approvals, damage reports and promotional price tags. Their sources include POS stock age, store inspection, ERP inventory and quality inspection records. In an operations handover these materials decide whether a batch may be marked down, may stay on the shelf, and who signs off.
The parties involved are the fresh produce store manager, the loss administrator and the finance auditor. Materials update by day or by shift. At handover the authorization state of the same day, same store and same batch must be checked. A near-expiry list records batch and expiry date. A markdown approval records applicant, approver and approval status. A damage report records quantity and reason. A promotional price tag records the marked-down price. Each material carries a different fact and cannot substitute for another.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. Check the vector model identifier, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.
Cited content is output as an array and its length can be 0. Even when nothing is found, the output path still runs. Semantic retrieval, full-text retrieval and hybrid retrieval are compared against a test question. No single mode is presumed suitable for all materials. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process, not a switch to turn on. Result reranking requires a ReRank model and must be enabled in settings. Image retrieval requires a corresponding image input and model capability; the optional paths shown in the overall document flow diagram do not mean every path executes on every request. Minimum relevance takes effect only in semantic retrieval or when result reranking is used. That value is a filter threshold and does not perform permission or entity isolation. The citation limit is measured in tokens, and content completeness must be tested with materials of different lengths.
Hypothetical paper exercise
Material text for checking: the store inspection record states "10 June 2025, Store A, banana batch B-0610, expiry date 12 June 2025, inspection status near-expiry". The markdown approval states "batch B-0610, applicant Zhang San, approver Li Si, approval status pending review, markdown percentage not filled in". The damage report states "batch B-0610, damage quantity 0, reason not filled in".
Handover request sentence to be judged: "Batch B-0610 has completed markdown approval, may be put on the shelf at the approved price tag, handed to the next shift for execution."
The request involves subject, object and state: the subject is Store A, the object is batch B-0610, and the approval state is pending review. Pending review does not show that approval is complete, and does not show that operating permission is absent. The judgment requires comparison of the same subject, same period and same field. Fields not stated remain unknown.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the near-expiry handling base whose vector model identifier matches | Check the vector model identifier of each base, exclude non-matching ones and record the observation |
| Search mode | Run semantic retrieval and full-text retrieval once each, then run hybrid retrieval | Compare the cited content of the three modes against the same test question |
| Citation limit | Hypothetical starting point 800 tokens | Test with near-expiry lists and approval forms of different lengths for truncation, and compare with a shorter limit |
| Minimum relevance | Hypothetical starting point 0.5 | Label relevant and irrelevant samples first, then compare with 0.3 and 0.7 and observe the filtering change |
| Question optimization | Off first, then on once | Compare the original question with the optimized question and check whether subject and batch number were altered |
| Result reranking | Enable once when a ReRank model is available | Compare candidate ordering before and after enabling, and observe score and order changes |
These numeric values are hypothetical starting points for this exercise and do not come from official defaults. The configuration items cover only the retrieval task on this page and do not involve form submission, approval writes or external synchronization.
Three common errors
- Treating a pending markdown approval as completed approval. In the hypothetical sample the approval status is pending review. If a retrieval result or the handover sentence concludes from this that the batch may go on the shelf, the input is the approval text, the comparison object is the approval status field, and the observable error is that the handover sentence writes pending review as completed. The correct action is to record that field as to be verified.
- Treating a semantic recall score as a permission filter. In the hypothetical sample the minimum relevance is set to 0.5. If this is taken to mean that unrelated stores or unauthorized records are excluded, the input is the score and the record list, the comparison object is the store and authorization scope of each record, and the observable error is that records from a different store still appear in the cited content. Retrieval does not carry authorization duties.
- Associating bases that use different vector models. The input is the vector model identifier of the two bases, the comparison object is whether the identifiers match, and the observable error is that the retrieval scope cannot be confirmed after association. The correct action is to check the identifiers, exclude non-matching ones and record the actual observation, without inferring missed recalls or redundant results.
How to verify the setup
- Input the inspection record, markdown approval and damage report text from the hypothetical sample, together with the handover request sentence to be judged. Compare the approval state in the request sentence with the approval status field in the approval text. The condition is met only when both share the same subject, same period and same field and their values agree. In the sample the two disagree, so the request is judged unsupported.
- Input the same test question and run semantic retrieval, full-text retrieval and hybrid retrieval separately. Compare the cited content arrays returned by the three modes and observe whether batch number B-0610 is recalled and whether the cited content contains the approval status field. If a mode returns an empty array, use the explicit paper input [] to check the downstream design. Do not claim that a fabricated rule retrieval necessarily returns [].
- Input an association request for two bases with different vector model identifiers. Compare whether the identifiers match, exclude non-matching ones and record the actual observation. If the identifiers match, run one retrieval and observe the cited content. If the identifiers do not match, do not run a cross-model retrieval and keep that comparison result as unknown.
This page addresses only fresh near-expiry markdown handling, operations permission and handover management, and knowledge base retrieval and recall. At operations handover, which materials, operations and requests sit in which authorization state cannot be resolved by treating retrieval or a form as authorization execution.