What the documents look like
Near-expiry markdown handling follows the shelf-stock lifecycle. Records arrive from POS stock-age data, store inspections, ERP inventory and quality-check records, and the working documents are near-expiry lists, markdown approvals, damage reports and promotional price tags. These materials update daily or by shift. In cross-border work, the same near-expiry batch can appear in a Chinese list and a foreign-language price tag at the same time, and field names, date formats and batch-number conventions differ between sources. Document parsing and chunking must turn these materials into retrievable fragments that can be traced back to the source text, while keeping batch, store, date range and authorization status as constraints on each fragment.
Constraints at the document chunking step
The batch number on a near-expiry list and the approval object on a markdown approval form must be fully covered by the same chunk. If chunking splits the applicable store from the markdown rate into different fragments, a retrieval hit cannot show whether that markdown covers the target store. Multilingual retrieval depends on semantic recall judging similarity across languages, and semantic similarity does not represent permission or entity separation; retrieved results still require manual checking of the subject and the authorization boundary. One set of source data can map to multiple vector indexes, so a batch can be recalled through a Chinese product name, a foreign-language product name and an internal number at the same time. Adding indexes does not change the source data, and it does not complete training or authorization.
Configuration and implementation checks
The following configuration is organized around one hypothetical drill sample. The source text for checking contains three sentences: "During the inspection on 10 March 2025, store A found that the leafy vegetables in batch L260310 had a remaining shelf life of less than 24 hours; the markdown approval form for this batch is numbered ZK-2025-0310-07 and its approval status is pending review; no damage report has been generated, and the promotional price tag has been printed but not posted." The draft sentence to be judged is: "Batch L260310 has completed markdown approval and can be sold at a 30 percent discount." The draft sentence conflicts with the source approval status of pending review, and the source does not state a markdown rate.
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Batch numbers and foreign-language product names coexist, so keyword matching and semantically similar content both need coverage |
| Citation limit | Specific tokens value to be filled in | Compare with the average length of candidate fragments and observe whether complete approval conditions are covered |
| Minimum relevance | 0-1 value to be filled in | Compare with semantic retrieval or rerank scores and observe whether low-relevance fragments are filtered |
| Question optimization | Decide per conversation turn whether to enable | Observe retrieval differences between a single clear question and a multi-turn follow-up |
| Chunk boundary | Treat the approval form as one complete unit | Compare whether batch number, store and status fall in the same fragment |
| Post-update review | Check training and index status after material changes | Record upload success and retrieval freshness separately, and compare actual recalled content |
Three common errors
- Treating "pending review" as already approved. In the sample, the approval status is pending review, while the draft sentence states that markdown approval is complete. These are mutually exclusive values on the same field, which forms a contradiction. For the check, input the sample source text and the draft sentence, compare the approval status field, and observe whether the draft sentence matches the source.
- Treating a retrieval hit as authorization. Recalling a markdown approval fragment only shows that the fragment was retrieved; it does not show that approval passed or that the price tag may be posted. For the check, input the retrieved fragment, compare the approval status inside it with the status claimed by the draft sentence, and observe whether an unsupported conclusion appears.
- Treating a successful upload as an updated index. After a material change, a successful upload still requires separate checking of training and index status. For the check, input the changed material and a search term, compare whether the returned fragment is the changed content, and observe whether an old version is still recalled.
How to verify the setup
- Input the three sample source sentences and the draft sentence "Batch L260310 has completed markdown approval and can be sold at a 30 percent discount," compare the value of the approval status field, and if the draft sentence conflicts with the source value of pending review, judge that the draft sentence is not supported by the source.
- Input the search terms "L260310" and "leafy vegetables near-expiry," compare whether the returned fragment contains the batch number, store A and the approval status together, and if the fragment lacks the approval status, judge that chunking did not preserve the complete constraint conditions.
- Input an empty-result follow-on scenario, use a paper input of an empty citation list to check the downstream display design, and observe whether the downstream handles it as an empty result. This check does not claim that any retrieval rule necessarily returns an empty list.