What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, so store oversight runs through employment and internal reporting rather than a franchise contract. The working material for product questions and selection support includes daily sales reports, shift schedules, display photos and inventory transfer orders. Product questions and selection support must answer how a given item sold, what stock remained and how it was displayed at a given store during a given period, and any citation has to trace back to one of those records. The underlying data arrives from the POS system, member CRM, ERP inventory and store inspections. A single citation can touch both a sales figure and an inventory state at once, so checking a citation means confirming which document version and which time range it corresponds to.
Constraints at the citation tracing step
Retrieval recalls candidate content. It does not decide whether a store employee may view a given record, and it does not decide whether a transfer order has been approved. A knowledge base search node can be linked to one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. An empty array means the output path still runs; it does not mean the workflow errored out and stopped. The citation limit is measured in tokens, and the minimum relevance value is a filtering threshold that takes effect only under semantic retrieval or when result reranking is used. Neither of them performs permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF. Image retrieval requires an image input and a model with the corresponding capability, and reranking requires a configured model that is enabled by settings. These paths are not guaranteed to run on every request.
Hypothetical paper exercise sample: a daily sales report from a brand-owned store dated 10 March 2025 reads, "This store sold 12 units of item A coat on 10 March, 5 units remain in stock, and no transfer occurred that day." A display photo note from the same day reads, "Item A coat is displayed at the second stand on the right side of the entrance." A draft sentence to be judged reads, "Item A coat sold 12 units on 10 March, stock is sufficient, and replenishment was transferred in from another store." For that sentence, the 12 units sold matches the report, whether "stock is sufficient" conflicts with 5 units remaining depends on the stated basis for "sufficient," and the transferred-in replenishment has no support within the report's statement that no transfer occurred.
Configuration and implementation checks
The table below lists, around the sample above, official documented settings or implementation checks relevant to this feature and the primary task. The values are starting points for this hypothetical exercise, used for comparison against the sample text, and are not official defaults.
| Configuration item | Hypothetical exercise starting point | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Include only knowledge bases containing the daily sales report and the display photo | Check that the selected bases use the same vector model, and exclude any with a mismatched model |
| Search mode | Hybrid retrieval | The sample contains the proper noun "item A coat" plus natural-language description; compare against pure semantic retrieval results |
| Citation limit | Hypothetical starting point of 3000 tokens | The sample text is short; compare the number of cited passages against a run without this limit |
| Minimum relevance | Hypothetical starting point of 0.5 | Compare against 0.3 and 0.7 to observe whether the sample text is still recalled |
| Result reranking | Enabled | Compare ordering against reranking disabled, and confirm whether the sample text position changes |
| Citation content output | Passed to the answering step | Check each answer sentence against the cited source text; when the array is empty, check whether downstream still executes |
Three common errors
- Treating a retrieval hit as authorization. Input is a retrieval request over the sample above; compare the retrieval output against the transfer order's approval state. If the draft sentence writes "replenishment was transferred in" while the report states only that no transfer occurred, that wording lacks support. Retrieval itself does not produce transfer authorization.
- Treating an empty citation array as a workflow failure. Input is a retrieval that does not recall the sample text; compare the node output against the downstream handling design. When the output is [], downstream should still execute. Record the actual observation of that retrieval; do not assert a cause for the miss.
- Treating minimum relevance as an entity or permission filter. Input is a query over a mixed knowledge base containing item A coat and item B coat; compare the recalled set before and after filtering. A low relevance threshold affects relevance ordering only, and does not exclude other stores or unauthorized records.
How to verify the setup
- Input the sample source text above together with the draft sentence. Compare the draft's "sold 12 units" against the report text; when the two agree, that clause is supported. Compare "stock is sufficient" against "5 units remain"; when the basis for "sufficient" is not stated, record it as pending verification. Compare "replenishment was transferred in" against "no transfer occurred" and record it as unsupported.
- Input a paper retrieval result that explicitly returns []. Compare whether the downstream answering step still executes and produces a notice. If downstream continues as designed, the empty-array handling condition is met.
- Input the sample source text and the identifier of another knowledge base that uses a different vector model. Compare the two vector model identifiers. When the identifiers differ, exclude joint linking, record the actual observation, and do not run a cross-model retrieval to infer recall differences.