What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, so store materials sit inside the brand's own systems rather than under a franchise contract. The daily materials for this category are daily sales reports, shift schedules, display photos and inventory transfer orders, refreshed daily or weekly. In cross-border work, the same set of materials can appear in Chinese, English or another language, and retrieval has to match across those languages.
Display photos are image material. Whether they can be retrieved depends on whether an image description index or an image vector index was produced at ingestion time, and on whether the current vector model accepts image input. The vector model and index step therefore cannot be reviewed only from the query side; it also has to confirm what index was built when the material was stored.
The knowledge base used for retrieval must use the same vector model, so the actual model identifier and the input type are checked first. One original data block can be associated with multiple indexes, and the index content is checked separately from the original text. Index content is best built from complete and focused information, with retrieval coverage and index length weighed against sample material.
Retrieval is a recall process. It does not carry permission or entity isolation, and the same-vector-model requirement does not extend to external systems also needing the same model. RRF is a fusion process for mixed results and is not written as a switch. The reference limit uses tokens as its unit, and the minimum relevance value takes effect only in semantic retrieval or when result reranking is used.
Constraints at the vector indexes step
A hypothetical paper exercise sample is given here for checking. A regional manager for a brand-owned store submits a paper handover request with this text: "From 1 March 2025 to 31 March 2025, the shift schedule for the Shanghai Nanjing West Road store (Chinese version) and the inventory transfer order (English version) have been uploaded to knowledge base A. Knowledge base A uses vector model M1. Request: use the Chinese question 'which goods were transferred in March' to retrieve these two materials." The draft sentence to be judged is: "This retrieval can return all transfer records." The subject needed for the judgment is knowledge base A, the objects are the two materials, the period is March 2025, and the authorization boundary is retrieval within knowledge base A only. This sample is hypothetical and was not actually executed.
The same-vector-model requirement applies to the knowledge bases being linked for retrieval. In the sample, knowledge base A carries model identifier M1. A second knowledge base B is compared on its model identifier; if the identifiers differ, the two are excluded from being linked together, and no cross-model retrieval is run to compare missed recalls.
Image retrieval depends on the ingestion step, not only on the query. If a display photo was stored with an image description index or an image vector index, and the model supports that input, the photo can take part in retrieval. If no such index was produced, the image index is recorded as missing, and no failure cause is inferred from that alone.
The reference limit is expressed in tokens per search, and the minimum relevance value is a 0-1 number that filters low-relevance results. That value applies only in semantic retrieval or when result reranking is used. It is a filter threshold, not a final ordering rule. After problem optimization, mixed retrieval, image retrieval or result reranking, the final result can come from several recall paths fused together, so it does not necessarily follow a single vector similarity order.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Vector model | Knowledge base A uses M1 | Check the actual model identifier and input type, then compare with the model identifier of knowledge base B; if they differ, exclude joint linking |
| Index content | Shift schedule segmented by date and shift; transfer order segmented by document number and line item | Compare with a single-segment index of the whole material, and observe whether the cited original text for the same question set is more focused |
| Image index | Display photos produce an image description index at ingestion | Compare with an ingestion method that produces no image index, and observe whether an image query hits |
| Search mode | Mixed retrieval | Compare with semantic retrieval alone, and observe hit differences for number and model-type queries |
| Reference limit | Set to 3000 tokens as the hypothetical starting point for this exercise | Compare with a lower value and observe whether cited content is truncated; this is an exercise starting point, not an official default |
| Minimum relevance | Set to 0.3 as the hypothetical starting point for this exercise | Compare with a higher value and observe whether low-relevance results are filtered; it takes effect only in semantic retrieval or reranking |
Three common errors
- Extending the same-vector-model requirement to external systems. Input is the model identifiers of knowledge base A and knowledge base B; the two are compared for consistency, and the observable result is that differing identifiers exclude joint linking. Treating an external business registration site as also needing the same model goes beyond what the documents state.
- Treating RRF as a switch that can be turned on. Input is the search parameter configuration; it is compared for the presence of an item named RRF, and the observable result is that the documents describe RRF as the fusion process after mixed retrieval and do not state that an independent switch exists.
- Treating retrieval output as an authorization basis. Input is the draft sentence "This retrieval can return all transfer records"; it is compared with the authorization boundary, and the observable result is that retrieval returns only recalled content inside knowledge base A. It does not show that all records exist or are accessible.
How to verify the setup
- Input the hypothetical sample text and the draft sentence. Compare the model identifier of knowledge base A with the language of the materials. The observable result is that when the model identifier is consistent and the materials have been ingested, retrieval can be executed; when the model identifiers are inconsistent, record that joint linking is not possible.
- Input the Chinese question "which goods were transferred in March". Compare whether the cited original text contains the document number and line items from the transfer order. The observable result is that the condition is met when the cited content matches the original text. If the citation is an empty array, check the downstream design against the documented node output, whose length can be 0, and do not claim that a rule-based retrieval necessarily returns empty.
- Input the display photo query. Compare whether an image description index or an image vector index was produced at ingestion. The observable result is that the image can take part in retrieval when the index was produced and the model supports image input; when it was not produced, record the missing image index and do not infer a retrieval failure cause.
This page addresses only brand-owned store operations, cross-border materials and multilingual retrieval, and vector models and indexes. In an operational handover, which materials, actions and requests sit at which authorization state cannot be treated as authorized execution through retrieval or a form.