What the documents look like
Fruit-store procurement for order fulfilment and after-sales questions rests on four document families: purchase orders, acceptance records, pesticide-residue test reports, and loss records. These are produced daily or per batch, so the version in force at the moment of a question matters. When a customer dispute concerns short delivery, bruising, or a food-safety concern, the reply must point to the specific document text, and the source must be traceable before the store treats the answer as settled.
The difference from convenience-store prepared-food ordering is that fruit procurement depends on wholesale-market conditions and grade-based acceptance rather than a head-office standard meal. The people involved are procurement staff, store fresh-produce supervisors, and wholesale-market stalls. Errors in this area produce spoilage losses, food-safety complaints, and stock-outs, which is why the citation chain carries operational weight rather than being a formatting preference.
Constraints at the citation tracing step
The knowledge-base search node can select one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length can be zero; an empty array still lets the output path run. The citation cap controls how many tokens are referenced in one search. The minimum relevance value applies only under semantic retrieval or result re-ranking, and it is a filter threshold; it does not perform permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF, and re-ranking requires a model and must be enabled by configuration. Whether a retrieved result is actually adopted still needs human review against the version in force.
A retrieval result is not an authorization. The search scope, the citation array, and the downstream reply are separate stages, and each needs its own check.
Hypothetical paper exercise. A store fresh-produce supervisor submits three source sentences: “The purchase order dated 10 June 2025 records 80 boxes of Red Fuji apples, supplier: Stall A.” “The acceptance record dated 10 June 2025 records 78 boxes actually received, of which 2 boxes were crushed, grade: second.” “The pesticide-residue test report dated 11 June 2025 records that the batch passed.” A draft sentence to be assessed: “The supplier has confirmed a full refund for the crushed apples in this batch.” The request to be checked: “Please generate an after-sales reply based on the above materials.” The subject, object, date, quantity, grade, and status needed for the judgment are present in the source text; the refund fact does not appear.
Configuration and implementation checks
The configuration below is tied to the paper exercise above. It lists only documented settings or implementation checks, and the starting values are hypothetical test starting points rather than official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the store procurement and acceptance knowledge bases that share the same vector model | Only knowledge bases with the same vector model can be linked together; verify the identifier first and exclude those that do not match |
| Search mode | Hybrid retrieval | The sample contains dates and grades, so semantic and keyword coverage are both needed |
| Citation cap | Set to a token amount that covers the three source sentences | Compare against the length of the sample text and confirm all three sentences can be cited |
| Minimum relevance | Keep the default and do not adjust for now | It applies only under semantic retrieval or re-ranking; first observe whether the default citation includes the source text |
| Result re-ranking | Enable by configuration when a re-ranking model is available | Secondary ordering when candidates are numerous; compare citation order with the disabled state |
| Citation output | Check whether the output array contains the three source sentences | An empty array still lets the path run; compare with the downstream handling design |
Three common errors
- Treating “full refund” in the draft sentence as an established fact. The source text contains only the purchase, acceptance, and test sentences; the refund subject, amount, and confirmation status do not appear, so the sentence is unsupported.
- Treating minimum relevance as a permission filter. The value is a filter threshold and does not exclude unauthorized records; retrieval by itself does not constitute authorization.
- Describing RRF as a switch that can be turned on. RRF is the result-fusion process of hybrid retrieval, not an independent configuration item.
How to verify the setup
- Input the three source sentences and the draft sentence, compare whether the draft sentence contains a refund fact not stated in the source, and observe whether the output is marked as pending verification rather than confirmed. The expected result is a pending-verification marking; this page does not claim the check has been run.
- Input the sample request, compare whether the citation array contains the three source sentences, and observe whether each citation corresponds to the source text. When the array is empty, record the gap instead of inferring a business outcome.
- Input the sample date and grade fields, compare the quantity difference between the acceptance record and the purchase order, and observe whether the citation preserves “78 boxes, 2 boxes crushed, grade second.” Do not add authorization or effective clauses that the source does not contain.