What the documents look like
Near-expiry markdown handling tracks shelf-stock batches rather than weighed orders. A batch carries a material identity, a shift or date range, and an approval state, and those three fields are what any citation must be able to point back to. The working documents are the near-expiry list, the markdown approval form, the write-off form, and the promotional price tag. Supporting records come from POS stock age, store inspection, ERP inventory, and quality inspection. In a cross-border, multilingual setting the same batch may exist as more than one language version, so a citation has to resolve to a specific version and a specific passage, not merely to the batch number. Document updates occur daily or per shift; that fact is treated here only as a source of version differences.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Knowledge bases built on different vector models cannot be associated together. The node outputs cited content as an array, and the array length may be zero; an empty array still lets the downstream path run, and it does not by itself mean the workflow has errored out. Retrieval performs recall only. It does not perform authorization. A similarity score is not a permission filter and not an entity filter, so whether a retrieved passage may be used still depends on a human check of the effective version and the authorization state.
Hypothetical drill sample. A store document library holds a Chinese Near-Expiry Markdown Approval Form whose text reads: "Object: leafy vegetable batch A-0417; status: markdown pending review; effective scope: shift of 2025-04-17; approver field unsigned." The same library holds an English Price Reduction Request reading: "Batch A-0417; status is pending review; shift 2025-04-17." A draft sentence under review reads: "Batch A-0417 has completed markdown approval and may be sold at the markdown price tag." The judgment requires the same subject, the same period, and the same field. Both source passages state that approval is pending, which conflicts with the draft sentence saying that approval is complete in the same field. Whether the goods may be sold at the reduced price still requires a separate check of business authorization; a retrieval match does not establish that authorization.
Configuration and implementation checks
The table below concerns this sample only. It lists settings already described in the official material, or implementation checks. Numeric entries are starting values for this hypothetical trial, with a stated reason and a comparison object; they are not official defaults.
| Configuration or check | Hypothetical starting value for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the Chinese and English libraries that share one vector model | Check the vector model identifier first, exclude non-matching libraries, and do not run a cross-model retrieval comparison |
| Search mode | Hybrid retrieval | Batch numbers are exact queries; compare hit behavior against pure semantic retrieval |
| Citation limit | Hypothetical start of 2000 tokens | Batch material carries approval fields; compare citation completeness against a 1000-token start |
| Minimum relevance | Hypothetical start of 0.5 | Compare against a 0.3 start and observe whether unrelated batch fragments enter the result |
| Question optimization | Enabled | Multilingual follow-ups often drop the subject; compare against the disabled state to see whether the query is completed |
| Citation handoff | Pass retrieval output item by item to the answering step | Check whether the answer sentence and the cited source share the same batch and the same status |
Result reranking requires a rerank model and must be enabled under its settings. RRF is the fusion process for hybrid retrieval results and is not a separate switch.
Three common errors
- Treating a retrieval hit as authorization. Input is the retrieved passage "batch A-0417 pending review." The comparison field is approval status, and the observable result is that the draft sentence "has completed markdown approval" conflicts with the source value, so the sentence is unsupported.
- Treating an empty citation array as an error. Input is a query that returns no content, and the output is []. Compare whether the downstream design still executes, and observe that the path runs; this observation does not establish that a retrieval rule always returns empty.
- Forcing an association across vector models. Input is two libraries with different vector model identifiers. Compare the identifiers, observe that non-matching libraries must be excluded, and record the actual observation rather than inventing missed recalls or redundant results.
How to verify the setup
- Input the two source passages and the draft sentence given above. Compare subject, period, and approval status. The draft sentence is satisfied only when a source explicitly records a completed signature and the period matches; the current sample is expected to remain pending verification.
- Input an empty citation array []. Compare whether the downstream node continues to execute and keeps a record of the gap. The path is expected to run and to produce no business action.
- Input two knowledge bases whose vector model identifiers differ. Compare the identifiers. Non-matching libraries are expected to be excluded and the observation recorded; no cross-model retrieval is run to compare missed recalls.
This page covers one scope only: near-expiry markdown handling, cross-border material and multilingual retrieval, citation source and tracing. Which materials, operations, or requests sit at which authorization state during an operations handover cannot be resolved by retrieval or by a form. The semantic recall and relevance settings do not carry permission or entity isolation. The same-vector-model requirement for associated libraries does not extend to any external system. No system error or automatic business action is asserted where the source text does not describe one.