What the documents look like
A fresh-food concession counter in a hypermarket is run by a supplier who is present on site and paid through a sales share. Copy for a product or a promotion therefore involves at least two parties: the hypermarket fresh-food department and the concession supplier. Statements about origin, specification, receiving batch, or inspection conclusion are spread across supplier delivery notes, food receiving inspection records, daily sales reports, and spoilage sheets. Before a promotional sentence or a handover request is drafted, the drafter confirms that each claim can be matched to a field in those materials.
The data sources named in the brief are the POS system, supplier delivery notes, food safety ledgers, and store stocktakes. The stated update frequency is daily and per batch. One product can therefore carry different batches and different inspection records on different dates, so batch information in copy must correspond to a specific date range. A record for one batch is not carried over to another batch.
Constraints at the citation tracing step
The drafting task is to check whether a proposed sentence is supported by the materials and to keep a pointer to the source. Retrieval is used to locate candidate materials. A retrieval result does not constitute authorization, and it does not confirm supplier qualification or product compliance.
The knowledge base search node outputs a citation content array, and the length may be zero. An empty array means the branch ran and returned no citations; it does not establish that the material is absent. The associated knowledge bases must use the same vector model, which is a precondition check on retrieval scope and does not extend to external systems. A similarity score is a relevance measure and does not perform entity isolation or permission filtering. Hybrid retrieval merges multiple result paths through RRF, which is part of the retrieval process rather than a separate switch that can be turned on. Reranking and image retrieval require the corresponding model and must be enabled in settings; the optional paths listed in the documentation do not mean every request executes all of them.
Configuration and implementation checks
The table below covers settings and implementation checks relevant to citation checking for copy, using this page's sample. Numeric entries are hypothetical starting points for this exercise, used to compare against retrieval output item by item.
| Configuration or check | Hypothetical starting point for this exercise | How to check and why |
|---|---|---|
| Associated knowledge bases | Only knowledge bases using the same vector model as this exercise | Check the vector model identifier first; exclude any that do not match, and do not run a cross-model retrieval comparison |
| Search mode | Hybrid retrieval | The sample contains a batch number, and number-type queries suit full-text and semantic search together |
| Citation limit | A small value as the hypothetical starting point for this exercise | Used only to control the length of a single citation, checked separately from minimum relevance |
| Minimum relevance | A middle value as the hypothetical starting point for this exercise | This value takes effect only in semantic retrieval or with reranking, and filters low-relevance results |
| Citation handoff | Pass the citation array to the copy checking step | Compare the answer against the cited original text item by item; answer fluency does not replace source evidence |
| Material currency check | Check the current version and a historical version separately | Whether a citation is adopted still depends on review against the valid version |
Three common errors
- Treating a retrieval score as an authorization conclusion. Hypothetical exercise sample: the material text reads "Supplier A, batch SP-0301 for 1 March 2026 to 7 March 2026, delivery note D-0312, food receiving inspection record registered." The proposed sentence is "All products on sale at this counter in March have passed all inspections." The input is those three material sentences and this proposed sentence. The comparison is between "all" in the proposed sentence and the single party, single date range, and single batch in the material. The observable result is that the material supports only that this batch was registered for inspection in that period; the proposed sentence exceeds that scope and is judged unsupported.
- Treating a pending review as approved. Hypothetical exercise sample: the handover request reads "Please treat Supplier B's concession contract copy as already effective." The input is this request sentence. The comparison is between "already effective" in the request and the status field in the available material. If the material shows only a pending review, the observable result is that the request lacks an effectiveness basis and is judged to be pending verification; the copy is not processed as effective.
- Treating an empty citation array as a system error. Hypothetical exercise sample: the paper input is the citation array []. The comparison is between this empty array and the downstream handoff design. The observable result is that the documentation states the length may be zero and the branch still runs, so this input is used to check whether the downstream handles empty citations; it does not support a claim that retrieval failed or that the material is absent.
How to verify the setup
- Input the three material sentences and the proposed sentence above, and search within the selected knowledge bases that use the same vector model. Compare the party, date range, and batch in the proposed sentence against the material fields. The condition is met when the inspection record for the corresponding batch can be located and the proposed sentence does not exceed that record's scope.
- Input the handover request "Please treat Supplier B's concession contract copy as already effective." Compare the status word in the request against the status field actually recorded in the material. The condition is met when a record with the same party, same period, and same field shows effectiveness; when only a pending review record exists, record it as an evidence gap.
- Input the citation array [] and check the downstream handoff step. Compare this input against the downstream design for handling empty citations. The condition is met when the downstream continues with the empty citation or clearly marks that no citation exists, without producing an inference that retrieval necessarily failed. All three items describe expectations and observation methods only, and none was actually executed.