What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, and store managers and regional managers act on procurement and supplier material under internal brand authorization rather than a franchise contract. The daily documents in this category include daily sales reports, shift schedules, display photographs and inventory transfer orders. Transfer orders connect directly to procurement requests, and supplier material sits alongside them. Data arrives from the point-of-sale system, the membership CRM, ERP inventory and store inspections, with daily or weekly update cycles. For retrieval and recall, the practical consequence is that transfer orders and supplier records need fields that let a search locate a subject, a store and a period, and that retrieval returns quoted content only. It does not move any approval into a new state.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; the vector model identifier is checked first, non-matching bases are excluded, and the actual observation is recorded without running a cross-model search to compare missed recalls. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval, and the same test question is compared across the three rather than assuming one mode suits all material. The reference cap is measured in tokens, so material of different lengths is used to check whether content stays complete before the cap is adjusted. Minimum relevance takes effect only in semantic retrieval or when result reranking is used, so relevant and irrelevant samples are labelled before the threshold is tuned. Question optimization performs coreference resolution and question expansion; the original question and the optimized question are compared to check whether a subject, store or date was altered. Image retrieval requires a corresponding image input and model capability, and reranking requires a rerank model enabled by configuration. The optional paths shown in the overall flow diagram do not all execute on every request. Quoted content is output as an array whose length can be 0, so the output path still runs when nothing is found. Retrieval is a recall step and does not perform authorization checks; forms collect values, and business verification, approval, database writes and external synchronization each need their own downstream implementation.
Hypothetical paper exercise sample: a store manager at a brand-owned store submits an inventory transfer request. The material text reads: "Requesting store: Store A (directly operated); request date: 10 March 2025; item: spring knitwear, size M; quantity: 30 pieces; request status is pending regional manager review." The draft sentence to be assessed is: "This transfer request has been approved by the regional manager and may be executed." The assessment needs evidence from the same subject, the same period and the same field. The source text states only "pending regional manager review" and records no approval entry.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only store operations bases that use the same vector model | Check the vector model identifier first, exclude bases whose identifier does not match, and do not run a cross-model search |
| Search mode | Test the same question separately with semantic, full-text and hybrid retrieval | Compare the quoted content returned by the three modes and observe differences between identifier-style questions and natural-language questions |
| Reference cap | Hypothetical starting point of 800 tokens | Test transfer orders and supplier material of different lengths for completeness, and compare against a longer or shorter starting point |
| Minimum relevance | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first, then compare how many irrelevant passages enter the quoted content at different thresholds |
| Question optimization | Enable for follow-up questions that contain references, disable for explicit questions | Compare the original question with the optimized question and check whether subject, store or date was altered |
| Result reranking | Use when a rerank model is available and enabled by configuration | Compare candidate text ordering before and after reranking; the rerank score can be used for filtering |
Three common errors
- Treating "pending regional manager review" as approved. The input is the sample text above, the field compared is request status, and the source states only pending review with no approval entry. The finding is unsupported and pending verification.
- Treating a retrieval result as authorization to execute. The input is the quoted transfer-order content returned by retrieval, and the comparison is between that quoted content and the approval status field. Retrieval returns quotations only and does not change approval status, so a transfer cannot be executed on that basis.
- Associating knowledge bases that use different vector models. The input is the vector model identifier of each base, the identifiers are compared, and a mismatch leads to exclusion. A cross-model search is not run to compare missed recalls.
How to verify the setup
- Ask the sample text "What is the status of the transfer request for Store A on 10 March 2025?" The input is the material text above, and the comparison is the request status field inside the returned quoted content. A return of "pending regional manager review" with no approval wording counts as correct recall; any approval wording is recorded as inconsistent.
- Check the draft sentence "This transfer request has been approved by the regional manager and may be executed." The input is the draft sentence together with the material text, and the comparison is whether the approval fact appears in the source. The source records no approval entry, so the finding is unsupported and pending verification.
- Perform a paper check of the empty-result path. The input is an explicit paper input of an empty array, and the comparison is whether the downstream design accepts a quoted array of length 0. The output path running through without producing a business action matches the documented description; where the downstream design cannot be confirmed, record it as a gap.