What the documents look like
Tmall flagship store daily operations draw on four document families for store service and internal training: product publishing pages, brand authorization letters, delivery commitment records and after-sales work orders. Compared with JD POP stores, Tmall flagship stores place greater emphasis on brand authorization and flagship identity, which means any training statement about qualifications must be traceable to the authorization document itself. Typical operating entities include brand e-commerce departments, Tmall store managers and agency operations companies. Daily data sources are the Tmall merchant backend, the Tmall rule center, ERP order tables and customer service chat records, with updates at daily and per-order frequency. The practical consequence for this page: when training material states a delivery deadline or an after-sales responsibility, it must point to the specific document and version that supports it, and a retrieval result cannot stand in for an authorization judgment.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, for vector search. Its output is a citation array, and the length may be zero, meaning the output path runs through even when nothing is retrieved. Knowledge bases with different vector models cannot be associated together; the identifier must be checked first and non-matching bases excluded, without running cross-model retrieval to compare missed recalls. Retrieval output must be handed to the answering step, where the answer is checked line by line against the cited source text. A thinking process or a fluent answer does not replace source evidence. The minimum relevance value is a filter threshold that takes effect only in semantic retrieval or when result reranking is used, and it does not perform permission or entity isolation. RRF is a fusion process for mixed results, not a switch that can be turned on. Image retrieval requires corresponding image input and model capability; reranking requires a rerank model and must be enabled in settings. Full-text retrieval and semantic retrieval each suit different query types, and mixed retrieval fuses results through RRF, but the final citations are not necessarily ordered strictly by a single vector similarity score.
Configuration and implementation checks
The following items concern citation tracing in store service and internal training, and list only documented settings or implementation checks tied to this page's documents.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select two knowledge bases whose vector model identifiers match | Check whether the vector model identifiers are the same; exclude any that differ; do not run cross-model retrieval |
| Search mode | Semantic retrieval | Used for natural-language questions; compare against full-text retrieval for numbered queries |
| Minimum relevance | Use the value currently shown in the interface | This value takes effect only in semantic retrieval or reranking; compare recalled item counts after adjustment |
| Citation limit | Use the value currently shown in the interface | Controls cited content volume in tokens; compare citation counts across different values |
| Question optimization | Leave off for now | When the question is already clear, enabling it adds one model call and latency; compare response time |
| Result reranking | Leave off for now | Takes effect only when a rerank model exists and reranking is enabled; compare rerank scores and citation order |
Three common errors
- Treating retrieval output as an authorization conclusion. Input: the citation array returned by the knowledge base search node. Compare: the authorizing entity in the cited source text against the entity in the material to be cited. Observable result: when the cited content contains no authorization scope or validity period field, that citation cannot support an authorization judgment and must be recorded as an evidence gap.
- Treating minimum relevance as a permission filter. Input: an after-sales work order that is similar to the question but belongs to another business line. Compare: the business ownership of that work order against the current training audience. Observable result: a high similarity score does not mean the record may be used for this training; business ownership must be checked manually, and when it is missing the record stays unsupported.
- Treating an empty citation array as a process failure. Input: a citation array of length zero output by the search node. Compare: whether the downstream answering step is designed to accept an empty array. Observable result: an empty array runs through the output path and does not equal a workflow error or termination; check whether the downstream step handles empty citations, and do not assert that retrieval rules necessarily return empty.
How to verify the setup
Hypothetical exercise sample: an agency operations team for a Tmall flagship store prepares store service training material and needs to cite a brand authorization letter. The source text reads: "Authorizing party A Brand Management Co., Ltd. authorizes B E-Commerce Co., Ltd. to operate the A brand flagship store on the Tmall platform, with an authorization period from January 1, 2025 to December 31, 2025." The draft sentence to be judged is: "Company B may operate the A brand flagship store on the Tmall platform long term." The three checks below all use this sample; the expected results do not claim to have been tested.
- Input: the citation array returned by the knowledge base search node, which contains the authorization letter source text above. Compare: "long term" in the draft sentence against "January 1, 2025 to December 31, 2025" in the source text. Observable result: the source states only a one-year period, so "long term" in the draft is unsupported and must be marked as pending verification; it cannot be judged as a contradiction or a negation.
- Input: an after-sales work order request awaiting handover, whose content is "add this work order to the training case library." Compare: whether the request carries an authorization approval record. Observable result: a request in pending review status does not demonstrate a lack of operating permission, and it does not equal approval; the authorization status must be recorded as pending verification, with the evidence gap preserved.
- Input: a paper input of an empty citation array []. Compare: whether the downstream answering step is designed to explicitly handle an empty array. Observable result: the downstream step either continues with empty citations or clearly states that no cited content exists; both satisfy the acceptance condition. Do not claim that a fictional rule necessarily returns [], and do not assert that an empty array necessarily causes an error.