What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, so store service and internal training do not run through franchise contracts. Citation tracing in this setting draws on four source systems: the POS system, member CRM, ERP inventory, and store inspections. The daily documents are daily sales reports, shift schedules, display photos, and inventory transfer orders, with updates at daily and weekly frequency.
For citation tracing, this means the sales figures, shift arrangements, display requirements, and transfer records that appear in training material each have a separate original source. Verification has to return to the matching document; one source cannot stand in for another. Display photos are image material, and whether they can be retrieved depends on whether an image description index or image vector index was built at ingestion time, and on whether a vision model or an image-capable vector model is configured.
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 with different vector models cannot be associated together. The node outputs cited content as an array, and the length can be zero, which means the output path still completes even when nothing is retrieved.
Semantic search, full-text search, hybrid search, question optimization, result reranking, and image retrieval are combinable paths; they do not all execute on every request. Image retrieval requires image input and model capability. Reranking requires a rerank model and must be enabled by configuration. RRF is a fusion process for mixed results and is not written as a switch. Minimum relevance is a filter threshold that takes effect only under semantic search or result reranking, and it does not perform permission or entity isolation. Retrieval itself does not constitute authorization. The citation limit is measured in tokens and is unrelated to thinking budget.
Hypothetical drill sample: the shift schedule of a brand-owned store for 10 March 2025 reads "Morning shift guide Li Moumou, on-duty period 09:00 to 18:00, responsible for display maintenance in Zone A." The daily sales report of the same store for 10 March reads "Zone A sales for the day RMB 8,600, average transaction value RMB 215." A draft sentence awaiting judgment reads "Li Moumou was responsible for Zone A display on 10 March, and Zone A sales that day were RMB 8,600, which shows that the display work directly produced this sale." The draft places a display duty next to a sales result. The shift schedule states only the on-duty period and the responsible zone; the daily sales report states only the amount and the average transaction value. Neither document states a causal relation between display work and the sales result, so the causal claim is unsupported. The draft also merges the shift schedule and the sales report into one citation, and the two sources need to be labelled separately.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the store training material base | First check the selected knowledge bases and material versions, confirm the vector model identifier is the same, and exclude bases with different vector models |
| Search mode | Hybrid search | Training material contains both codes and natural-language description; compare recall coverage against pure semantic search |
| Citation limit | A starting value set in tokens | Compare returned citation length before and after adjustment, and observe whether key original text is truncated |
| Minimum relevance | Set only when semantic search or reranking is enabled | Compare against results with the filter off, and confirm the threshold does not perform permission judgment |
| Result reranking | Enabled by configuration when a rerank model is available | Compare candidate order against the un-reranked order, and observe the 0-1 score change |
| Citation display check | Compare the answer against cited original text item by item | Use the shift schedule and daily sales report text to check the draft sentence, and mark unsupported causal wording |
Three common errors
- Treating a retrieval hit as authorized execution. The shift schedule states an on-duty period; it does not mean the employee has been approved to execute a transfer or change a display standard. Retrieval outputs cited content and does not constitute any business authorization. Approval, database writes, and external synchronization need separate downstream implementation.
- Merging evidence from different fields into one conclusion. The responsible zone in the shift schedule and the sales amount in the daily sales report are different fields. Placing them side by side cannot produce a causal relation, and the claim remains unsupported.
- Treating an optional path as executed on every request. Image retrieval and result reranking need the corresponding input and model capability. When they are not configured, they do not execute, and their non-execution cannot be read as retrieval failure or missing data.
How to verify the setup
- Input the shift schedule text and the daily sales report text from the hypothetical drill sample, and compare the causal wording in the draft sentence against the field scope of the two originals. The citation tracing requirement is met only when the draft states solely the on-duty period, responsible zone, sales amount, and average transaction value as recorded in the originals. A causal conclusion does not meet it.
- Input a draft sentence that contains only the on-duty period from the shift schedule, and compare whether its citation label points to the shift schedule. The source traceability requirement is met only when the citation label matches the original source.
- Input an explicit paper value of an empty array [], and compare whether the downstream answer step is designed to continue through an empty citation. The empty-result handling requirement is met only when the downstream design allows an empty citation and keeps a record of the gap. This check does not claim that a fictional rule search will necessarily return an empty array.