What the documents look like
Near-expiry markdown work concerns shelf inventory batches rather than order-by-order weighing. The data updates daily or by shift, and it comes from POS stock age, store inspections, ERP inventory and quality inspection records. The working documents are the near-expiry list, the markdown approval, the write-off form and the promotional price tag. The most time-consuming steps are identifying near-expiry batches, approving markdowns and segregating unsellable goods. When promotional copy is drafted, these materials determine which claims the copy can carry: a discount rate has to correspond to an approval record, a shelf-life statement has to correspond to quality inspection and stock age, and price tag wording has to correspond to the actual tag version in use. The drafting task is to trace each proposed sentence back to a specific source document and to keep unsupported statements visibly unsupported.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs references as an array whose length may be zero. A zero-length array still lets the output path run through. A retrieval pass may combine question optimization, semantic or full-text retrieval, RRF fusion, reranking, similarity filtering and a reference cap before references are produced. The minimum relevance value applies only in semantic retrieval or when result reranking is used; it is a filtering threshold, not a permission or entity isolation mechanism. Reranking requires a rerank model enabled by configuration, and image retrieval requires image input plus the corresponding model capability. These optional paths are not executed on every request. Knowledge bases with different vector models cannot be associated together; the model identifier must be checked first and non-matching bases excluded. The reference cap is measured in tokens, and thinking budgets belong to thinking configuration; the two must not be mixed.
Hypothetical paper exercise: a store manager plans a promotional sentence, and the sentence under review is “This batch of near-expiry yogurt is marked down by half, with a shelf life through the 20th of this month.” The source materials provided for checking are: the near-expiry list states “Object: brand X plain yogurt, batch A, status: near-expiry, suggested markdown rate: half”; the quality inspection record states “Batch A, inspection date the 10th of this month, conclusion: qualified, shelf life through the 20th of this month”; the markdown approval form states “Batch A, requested markdown half, status is pending review.” The materials do not state that approval has passed, and they do not state whether the batch has been segregated.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases containing the near-expiry list, markdown approval, write-off form and promotional price tag | Check each base's vector model identifier first and exclude bases with a different model |
| Search mode | Hybrid retrieval | Compared with semantic retrieval alone, this also covers exact terms such as batch numbers |
| Minimum relevance | Hypothetical starting point 0.5 | Compared with a lower value, to observe whether unrelated fragments enter the references |
| Reference cap | Hypothetical starting point 2000 tokens | Compared with a smaller value, to observe whether source text is truncated |
| Result reranking | Enable when a rerank model is available | Compared with the result when it is not enabled, to check whether approval status is ordered correctly |
| Reference handover | Pass retrieval output to the copy drafting step | Check each drafted sentence against the quoted source text |
These numbers are starting points for this hypothetical exercise only, not official default values.
Three common errors
- Treating “pending review” as “approved.” The approval form states “pending review.” If the drafted sentence says the markdown is already in effect, the input is the approval form text, the comparison is between “pending review” and “approved,” and the observable result is that the two do not match, so the drafted sentence is not supported.
- Treating a retrieval score as a permission filter. Minimum relevance only filters low-relevance results; it does not exclude unauthorized records. The input is the search parameter and the target base scope, the comparison is between the score threshold and access permission settings, and the observable result is that a higher or lower score does not change the permission determination.
- Forcing association across different vector models. The input is the vector model identifiers of two knowledge bases, the comparison is whether the identifiers are the same, and the observable result is that different identifiers must not be associated together; the non-matching base is excluded and handled separately.
How to verify the setup
- Check the drafted sentence against the hypothetical sample above. The input is the near-expiry list, the quality inspection record, the markdown approval form and the drafted sentence; the comparison is between the markdown rate, shelf life and approval status in the drafted sentence and the values stated in the materials; the observable result is that all three match and the approval status is not written as approved. The approval status item remains unverified.
- Check downstream handover with an explicit paper input of an empty array. The input is references
[]; the comparison is whether the downstream copy drafting step still runs through and records the gap; the observable result is that the path does not break because of the empty array, and no assertion is made that retrieval necessarily returns empty. - Check source timeliness with three versions of the same batch: list, approval and price tag. The input is the current and historical versions of the materials; the comparison is whether batch number, status field and tag version share the same subject, the same period and the same field; the observable result is that a contradiction is found only when the same field has explicitly mutually exclusive values, while an unstated value or evidence from a related stage alone is recorded as unsupported or pending verification.