What the documents look like
Near-expiry markdown handling follows the shelf-stock lifecycle, not the piece-by-piece weighing and sorting rhythm of other fresh categories. The working materials are near-expiry lists, markdown approvals, damage write-off forms and promotional price tags. Their data comes from POS stock age, store inspection, ERP inventory and quality-check records. Store managers, loss administrators and finance auditors all draw on these materials.
Updates occur daily or per shift, so retrieval must select the version matching a specific date and shift scope. For copy drafting, the retrieval target is the product status, markdown conditions and approval conclusions recorded in those materials. The retrieval result supplies quoted content only. It does not perform approval, change inventory or synchronize any external system.
A near-expiry list may state a batch, a shelf-life end date and a status. A markdown approval may state a proposed range and a pending or signed approval state. These fields carry different meanings, and a drafting step must compare the proposed sentence against the exact field it depends on.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The vector model identifier should be checked first, non-matching bases excluded, and the actual observation recorded. This page does not assert any runtime outcome from associating mismatched models.
Search mode can be compared across semantic retrieval, full-text retrieval and hybrid retrieval using the same test question. No single mode is assumed suitable for all materials. Hybrid retrieval combines multiple result paths through RRF, which is a fusion process rather than a separate switch that can be turned on.
The citation limit is measured in tokens. Short approval forms and long inspection records should be used to observe whether content is truncated, and the limit adjusted from that observation. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Relevant and irrelevant samples should be labelled first, then the threshold compared against those scores. This value filters low-relevance results. It is not an authorization or entity filter, and retrieval does not decide whether a record is permitted for a given store or role.
Question optimization performs reference resolution and question expansion. The original question and the optimized question should be compared, and the subject checked for unintended change. Reranking requires a rerank model and must be enabled according to configuration. Image retrieval requires an image input and matching model capability. Quoted content is output as an array whose length can be 0, and the downstream path still runs when nothing is retrieved.
Hypothetical paper exercise sample, not a real record. Source text one: "2025-04-10 morning shift inspection record: A-zone low-temperature cabinet strawberry batch SY-0410, shelf life to 2025-04-12, status near-expiry, entered on the same day's near-expiry list." Source text two: "2025-04-10 markdown approval form: strawberry batch SY-0410, markdown range subject to store manager approval, approval status pending review, unsigned." Proposed draft sentence under review: "The store's strawberries have received markdown approval, and a near-expiry discount is available today."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only same-vector-model bases containing near-expiry lists and markdown approvals | Compare against the identifier of a base using another vector model and exclude the non-matching base |
| Search mode | Run semantic retrieval and full-text retrieval once each, then hybrid retrieval | Compare the quoted content returned by the three modes for the same test question |
| Citation limit | Hypothetical starting point of 800 tokens | Compare a short approval form against a long inspection record to observe truncation |
| Minimum relevance | Hypothetical starting point of 0.4 | Compare against scores of samples labelled relevant and irrelevant, then adjust |
| Question optimization | Enable for questions containing references, disable for questions with a complete subject | Compare whether the subject changes incorrectly between the original and optimized question |
| Result reranking | Enable when a rerank model is available | Compare candidate ordering and score changes before and after reranking |
Three common errors
- Treating "pending review" in source text two as markdown approval and drafting "has received markdown approval." The approval status field is pending review and unsigned, so the proposed sentence lacks a basis for approval.
- Treating the minimum relevance value as a permission or entity filter and assuming a higher value prevents retrieval of other stores or unauthorized records. That value filters low-relevance results only, and retrieval does not perform authorization decisions.
- Associating knowledge bases that use different vector models, or extending the same-vector-model requirement to external systems. The identifier should be checked and non-matching bases excluded, without running cross-model retrieval to compare missed recalls.
How to verify the setup
- Input source text one and source text two, then compare "has received markdown approval" in the proposed sentence against the approval form fields "pending review" and "unsigned." The sentence is supported only when the approval form shows an explicit approved status and a signature record. At present it is not supported.
- Input the proposed sentence "a near-expiry discount is available today" and compare it against the markdown range and effective-scope fields in the materials. The materials do not state a specific markdown range or applicable period, so the sentence remains unverified. The condition is met only when the approval form states both a range and a validity period.
- Input a test question with a complete subject, run semantic retrieval, full-text retrieval and hybrid retrieval separately, and compare whether the returned quoted content includes batch SY-0410 and the corresponding date. If a mode does not return the batch, record it as not recalled, without asserting that another mode must return it. An empty result can be checked against a downstream handling design using an explicit paper input of [], without claiming that a fictional rule search necessarily returns [].
This page addresses near-expiry markdown handling, product and marketing copy drafting, and knowledge base retrieval and recall. Whether a specific promotional sentence is supported by the materials remains a copy-drafting check and does not become an identity-certification procedure.