What the documents look like
Near-expiry markdown handling works on shelf-stock batches. The records that matter are near-expiry lists, markdown approvals, damage write-off forms, and promotional price tags. These materials update daily or by shift, and the main sources behind them are store inspections, ERP inventory, quality-check records, and POS stock age. In a cross-border, multilingual setting, one batch can appear under a Chinese product name, a foreign-language label, and price tags in several languages. Retrieval has to place these different language expressions of the same subject inside one knowledge-base scope. The knowledge-base search node can associate one or more knowledge bases that use the same vector model; bases built on different vector models cannot be associated together.
The materials often arrive as short phrases, identifiers, and status words: batch numbers, markdown percentages, write-off reasons. Semantic retrieval relies on vector distance and suits natural-language questions. Full-text retrieval relies on keyword hits and suits identifiers and proper nouns. Hybrid retrieval uses both and merges the results through RRF. The minimum relevance value is a filter threshold that applies only under semantic retrieval or when result reranking is used; it does not perform permission or entity isolation. The reference cap is measured in tokens and controls how much content one search may cite. Question optimization handles coreference removal and question expansion and adds one model call. Image retrieval requires image input and matching model capability, and reranking requires a model enabled by configuration. The optional paths shown in the overall flow diagram do not mean every request executes all of them. Retrieval by itself is not authorization; approval status and signature status need separate checking.
Hypothetical paper exercise: the following materials are invented for this paper check and come from no real customer or system record.
Material text one: "Near-expiry batch A-2024-0315, yogurt 200 g, best-before 2024-03-20, store inspection recommends a 30% markdown, markdown approval form status is pending review."
Material text two: "Damage write-off form BL-2024-0316, same batch A-2024-0315, write-off reason recorded as 'packaging damaged', quantity 12 units, quality-check record not attached."
Draft handover request to be judged: "Search the knowledge base, confirm that the markdown approval for batch A-2024-0315 has passed, and generate the promotional price tag on that basis."
Points to be judged: the markdown approval form shows pending review and no pass; the write-off reason and the markdown recommendation sit in different fields; the quality-check record is not attached. The request treats pending review as passed, which is an unsupported judgment.
Constraints at the retrieval step
The near-expiry materials are short and status-heavy, so the choice of search mode changes what can be found. Identifier-style questions such as a batch number compare better under full-text or hybrid retrieval, while natural-language questions compare better under semantic retrieval. Hybrid retrieval merges semantic and full-text candidates through RRF, and the merged range is wide enough that reranking is commonly used to reorder and then filter by rerank score. The minimum relevance threshold applies only under semantic retrieval or reranking, so it cannot be relied on to exclude unrelated businesses or unauthorized records. The reference cap is a token amount, not a top-k item count; the documented reason is that chunk lengths differ greatly in mixed knowledge bases, which makes top-k unstable.
Question optimization runs before retrieval and adds a model call, so it helps multi-turn and short questions while increasing overall time. Image retrieval depends on whether the vector model accepts image input or a vision model is configured, and on whether image indexes were built at ingestion. Reranking depends on a rerank model being available and enabled. Retrieval returns cited content as an array whose length may be zero, so the downstream path still runs when nothing is found. Retrieval does not approve anything, and an approval form status of pending review does not by itself prove that an operator lacks permission to act.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only near-expiry bases that share one vector model | The document requires the same vector model; compare base identifiers and exclude those whose models differ |
| Search mode | Run semantic retrieval and full-text retrieval separately first | Compare full-text hits for identifier questions and semantic hits for natural-language questions |
| Reference cap | Set one hypothetical starting value that covers a complete near-expiry list | Compare whether content is truncated across short and long materials, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set a threshold | The value applies only under semantic retrieval or reranking; compare results before and after filtering |
| Question optimization | Enable for multi-turn and short-question scenarios | Compare the original question with the optimized question and check whether the subject was changed |
| RRF and reranking | Observe fusion and rerank scores after hybrid retrieval | RRF is a fusion process, not a standalone switch; reranking needs a model and must be enabled by configuration |
Three common errors
- Treating pending review as passed. In the sample, the markdown approval form status is pending review, while the draft sentence asks to generate a promotional price tag on that basis. The input is the approval status field; the comparison is between the mutually exclusive values "pending review" and "passed"; the observable result is that the two do not match, so the judgment is unsupported.
- Merging different fields into one fact. In the sample, the write-off form reason reads packaging damaged, the markdown recommendation reads 30%, and the quality-check record is not attached. The input is the write-off reason field and the markdown recommendation field; the comparison covers the same batch and the same period; the observable result is that the two fields are independent and cannot support each other.
- Treating retrieval results as authorization. The draft sentence asks to generate a promotional price tag directly after retrieval. The input is the retrieval request; the comparison is between the retrieval output and approval authorization; the observable result is that retrieval returns cited content only, authorization status needs separate checking, and a gap is recorded when it is missing.
How to verify the setup
- For material text one, input batch number A-2024-0315 and "markdown approval status". Compare the approval status field returned by retrieval with the original text "pending review". The condition is met when the returned content matches the original and the words "passed" do not appear; if "passed" is returned or the result is empty, record it as pending verification.
- For material text two, input write-off form BL-2024-0316 and "write-off reason". Compare the returned write-off reason with the original text "packaging damaged". The condition is met when the reason field and the missing quality-check record are both presented; if the write-off reason is used as markdown grounds, record it as field mixing.
- For the draft handover request, input the full request text and compare the retrieval output with the approval authorization boundary. The condition is met when retrieval produces no authorization action and promotional price tag generation is shown to need a separate downstream implementation; if the output claims the approval has passed, record it as an overreach judgment.
An empty-result handoff can be checked against downstream design using the explicit paper input []. A cross-model association is not permitted, so verify identifiers first and exclude non-matching bases rather than running a cross-model search to compare missed recalls.