What the documents look like
Near-expiry markdown handling follows the shelf-stock lifecycle rather than per-order weighing and sorting. The working papers for this category are the near-expiry list, the markdown approval, the damage write-off form, and the promotional price tag. These materials draw on POS stock age, store inspection, ERP inventory, and quality inspection records, and they update by day or by shift. For the contract and transaction material comparison task, the job is to align the markdown approval against the damage write-off form and the promotional price tag on subject, batch, quantity, price, and validity scope, and to confirm that each cited source can be traced back to a specific material. A mismatch between an approval status and a write-off record can lead to expired sales, expanded loss, or a food-safety complaint, so the citation chain matters as much as the values themselves.
Constraints at the citation tracing step
The knowledge base search node can be associated with one or more knowledge bases that share the same vector model, and it outputs references as an array whose length may be zero. When the array is empty, the output link still runs through. The reference content then passes to the answering step, where each answer sentence is checked against the cited source text. The minimum relevance value is a filter threshold that applies only under semantic retrieval or when result re-ranking is used; it does not act as a permission or entity isolation control. Hybrid retrieval merges multiple result paths through RRF, re-ranking requires a model and is enabled according to settings, and image retrieval requires image input plus the corresponding model capability. Full-text retrieval suits exact queries such as identifiers, model numbers, and proper nouns.
Hypothetical drill sample: at a fresh produce store, on the shift of 10 March 2025, the store manager submits a markdown approval whose source text reads "Batch A-0310, leafy vegetables, inbound date 8 March 2025, shelf life to 12 March 2025, requested markdown 30%, approval status is pending review." A separate damage write-off form reads "Batch A-0310, leafy vegetables, write-off quantity 5 kg, write-off reason: visibly wilted, write-off date 10 March 2025." The draft sentence to be judged is "Batch A-0310 has completed markdown approval and executed the write-off." The draft treats a pending-review status as completed approval and treats a write-off record as an executed write-off; neither claim has supporting evidence in the two source texts.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the knowledge bases holding the near-expiry list, markdown approval, damage write-off form, and promotional price tag | Check whether the vector model identifiers match; exclude any base whose identifier differs from the shared association |
| Search mode | Hybrid retrieval | Batch identifiers and dates suit full-text retrieval, material semantics suit semantic retrieval, and the two paths merge through RRF |
| Minimum relevance | Hypothetical starting value 0.3 | Compare against each returned reference to observe whether low-relevance fragments are filtered; this value is for this trial comparison only |
| Reference limit | Hypothetical starting value 2000 tokens | Compare against the actual text length of the near-expiry list and the damage write-off form to observe whether references are truncated |
| Result re-ranking | Enabled | When candidates are numerous, re-rank the text a second time and use the re-rank score for filtering |
| Reference output handoff | Pass the reference array to the answering step | Check each answer sentence against the cited source text |
Three common errors
- Treating a pending-review status as completed approval. The markdown approval source marks "approval status is pending review," while the draft writes "has completed markdown approval." The input is the approval status field, the comparison is between pending review and completed, and the observable result is that the two values do not match.
- Treating a write-off record as an executed write-off. The damage write-off form records quantity and reason but states no execution confirmation, while the draft writes "executed the write-off." The input is the write-off execution status field, the comparison is between a record and an execution confirmation, and the observable result is that execution confirmation is absent.
- Associating knowledge bases with different vector models. The input is the vector model identifier of each knowledge base, the comparison is whether the identifiers match, and the observable result is that a base with a differing identifier is wrongly included in the shared association.
How to verify the setup
- Input the markdown approval source text from the hypothetical drill sample and the draft sentence "Batch A-0310 has completed markdown approval and executed the write-off." Compare the value of the approval status field and observe whether the draft presents pending review as completed. The expected result is that this mismatch is identified and the pending-verification status is retained.
- Input the damage write-off form source text from the hypothetical drill sample and the same draft sentence. Compare whether the write-off form states an execution confirmation and observe whether the draft presents the record as executed. The expected result is that the missing execution confirmation is identified and the evidence gap is recorded.
- Input the vector model identifiers of the knowledge bases. Compare whether the identifiers match and observe whether a base with a differing identifier has been included in the shared association. The expected result is that the differing base is excluded and the actual observation is recorded, without running a cross-model retrieval to compare missed recalls.
A paper input of an empty reference array can be used to check the downstream design: the search node output length may be zero while the output link still runs through, so the answering step must be reviewed for how it handles an empty array. This is a design check on paper, not a claim that any retrieval rule returns an empty array.
The optional paths listed in the overall retrieval flow do not all execute on every request. Image retrieval requires corresponding image input and model capability, and re-ranking requires a model and is enabled according to settings. Knowledge bases with different vector models cannot be associated together; the check is to verify identifiers, exclude non-matching bases, and record the actual observation. Retrieval is not authorization, and the search node does not grant permission to execute a markdown, a write-off, or a price change. A pending approval is not an approval, an offered option is not execution, and a recorded write-off is not a completed write-off. Where a source text does not state a field, that field remains unknown.