What the documents look like
Daily operations for a Tmall flagship store handle four document types in cross-border and multilingual retrieval: product listing pages, brand authorization letters, shipping commitment records, and after-sales tickets. These materials commonly mix Chinese and English. A brand authorization letter may be issued by an overseas brand owner, while category qualification fields on a product listing page and compensation records on an after-sales ticket each carry their own identifiers. Retrieval has to separate natural-language questions from exact-identifier questions. Semantic retrieval suits a question such as whether a product's authorization has expired, while full-text retrieval suits an exact field such as an authorization letter number.
Two industry facts shape this task. First, compared with JD POP stores, Tmall flagship stores place more emphasis on brand authorization, flagship store identity, and Tmall points compensation rules, so authorization documents and shipping commitment records carry operational weight. Second, the typical consequences of error are product delisting, delayed-shipping compensation, and a falling store rating, which makes the accuracy of retrieved evidence matter. Document updates occur daily and per order, but this page addresses only the retrieval and recall step. It does not extend into batch synchronization or transaction-triggered mechanisms.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases using different vector models cannot be associated together. The vector model identifier must be checked first, non-matching bases excluded, and the actual observation recorded. Search mode, reference limit, minimum relevance, and question optimization are configurable fields. The reference limit uses tokens as its unit. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch. Image retrieval requires image input and corresponding model capability. Reranking requires an available model and must be enabled according to settings. Empty results are output as an array whose length can be zero; even when nothing is found, the output path still completes.
Hypothetical paper exercise sample: a Tmall flagship store operator prepares to hand over a brand authorization letter. The source text reads: "Authorizing party: Hypothetical Brand Ltd.; authorized party: Hypothetical Flagship Store; authorized product category: home appliances; authorization validity period: January 1, 2025 to December 31, 2025; authorization letter number: HBL-2025-001." A draft sentence awaiting assessment reads: "Hypothetical Flagship Store has obtained authorization from Hypothetical Brand Ltd. to sell home appliances for the full year 2025." Assessing this draft sentence requires checking four fields against the source: subject, object, period, and number.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge base | Associate only one base using the same vector model | Check the vector model identifier first, exclude bases on different models, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Compare recalled content for the same question against semantic retrieval and full-text retrieval separately |
| Reference limit | Set to 2000 tokens as the hypothetical starting point | Test whether content is complete using materials of different lengths, then compare against a smaller value |
| Minimum relevance | Set to 0.3 as the hypothetical starting point | Label relevant and irrelevant samples first, then compare filtering effects against a higher value |
| Question optimization | Off as the hypothetical starting point | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Off as the hypothetical starting point | When a reranking model is available, compare ordering differences between enabled and disabled states |
Three common errors
- Treating a retrieval result as authorization status. Input a retrieval request and compare the returned reference content against the authorization letter source. If the reference contains only the authorizing party name and lacks the validity period field, authorization cannot be judged valid. The observed result is incomplete reference content, and the evidence gap must be recorded.
- Treating minimum relevance as a permission filter. Input a question involving an unauthorized category and compare returned results against the category field in the authorization letter. If returned content includes materials from other categories, the relevance threshold does not perform entity isolation. The observed result is that unrelated categories may still be recalled.
- Treating a pending review as approved. Input a handover request and compare the status field in the request against the source. If the source states only "submitted for review" with no approval record, authorization cannot be judged effective. The observed result is that the status field is unsupported.
How to verify the setup
- Input the authorization letter source and the draft sentence from the hypothetical exercise sample, then compare whether the draft sentence's subject, object, period, and number match the source. If all four fields match, the draft sentence has source support. If the period field is missing, record it as pending verification and do not judge it as contradictory.
- Input a query containing the exact number HBL-2025-001 and compare the returned results from full-text retrieval and semantic retrieval. If full-text retrieval returns content corresponding to that number, exact query capability is available. If semantic retrieval does not return it, record that mode's limitation for number-type queries.
- Input an empty array as reference content and check whether the downstream path continues executing. If the downstream node handles the empty array as designed, empty-result continuation works. If the downstream node errors, record that the node does not handle empty results.
This page addresses only Tmall flagship daily operations, cross-border materials and multilingual retrieval, and knowledge base retrieval and recall. During operational handover, the authorization status of materials, operations, and requests cannot be established by retrieval or a form. Semantic recall and relevance do not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to business registration sites or other external systems. Any system error or automatic business action not described in the source documents must not be asserted.