Tmall flagship operations: citation tracing for procurement

Review citation tracing for procurement in Tmall flagship operations. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

Tmall flagship store procurement and supplier records usually combine brand authorization letters, product listing pages, delivery commitment records, after-sales tickets, and ERP order tables. Brand authorization and flagship store identity affect whether a product can stay listed. Product delisting, delayed-shipment compensation, and store rating decline are common consequences when these records are wrong. The records come from the Tmall merchant backend, the Tmall rules center, ERP order tables, and customer service chat logs. Update frequency is daily or per order.

For citation tracing, the practical question is which version of which document a statement came from. A supplier authorization letter states a subject, an authorized category, and a validity period. A quotation states a price and a quotation validity period. These are different fields. A quotation validity period does not state a supply period. A retrieved authorization fragment shows that a document was recalled, not that the authorization is currently effective.

Constraints at the citation tracing step

The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. Citation content is output as an array, and its length can be zero. An empty array does not mean the workflow has errored or stopped.

Retrieval results must be passed to the answering step, and the answer must then be checked sentence by sentence against the cited source text. The minimum relevance value is a filtering threshold. It does not perform permission filtering or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to external systems.

Full-text retrieval suits precise queries such as numbers and model codes. Hybrid retrieval merges multiple retrieval paths through RRF. Rerank reorders candidates after it is enabled and a model is available. Image retrieval requires image input and the corresponding model capability. Rerank requires a model and must be enabled by configuration. The overall retrieval flow lists optional paths; not every request executes all of them.

Hypothetical paper exercise sample: a brand e-commerce team is preparing supplier records on 10 March 2025. The source text for checking contains three sentences. First sentence: "Supplier A obtained brand authorization from the brand owner on 1 January 2025, with an authorized scope of the Tmall flagship store home appliance category." Second sentence: "The authorization letter is valid until 31 December 2025." Third sentence: "Supplier A submitted a quotation on 20 February 2025, with a quotation validity period of 30 days." The draft sentence to be judged is: "Supplier A currently holds Tmall flagship store home appliance category authorization, and the quotation can be used for March procurement." Checking requires comparing the draft against the source on subject, authorized scope, and quotation validity period. The quotation validity period is not a supply period. The draft's statement about March procurement lacks supply-period evidence.

Configuration and implementation checks

Configuration itemHypothetical starting value for this exerciseReason and comparison method
Associated knowledge basesSelect the supplier records base and the product listing page base that use the same vector modelCompare the vector model identifiers of the two bases; exclude any base whose identifier does not match
Search modeHybrid retrievalSupplier names and authorization numbers need keyword hits; authorized scope descriptions need semantic recall
Citation limitSet to 3000 tokensCompare against the chunk length of a single supplier record; covers the authorization letter and the quotation as two source texts
Minimum relevanceSet to 0.5Compare against semantic retrieval and rerank scores; filters low-relevance fragments; does not perform permission filtering
Result rerankEnabledWhen candidates are numerous, reorder by rerank score and compare against the unreranked order
Citation handoffPass the citation array to the answering stepCheck each answer sentence against the cited source text; when the array is empty, check whether the downstream path still runs

These values are hypothetical starting points for this exercise. They are based on the sample record chunk length and candidate count. They are not official defaults.

Three common errors

  1. Treating a retrieval hit as authorization confirmation. The input is the draft sentence "Supplier A currently holds authorization." The comparison object is the subject and scope in the authorization letter source text. The observable result is whether the draft uses the same subject and the same scope as the source. Retrieving an authorization fragment only shows the document was recalled. Whether the authorization is effective still requires checking the validity period in the source text.
  2. Treating the quotation validity period as a supply period. The input is the quotation source text "quotation validity period of 30 days." The comparison object is the draft sentence "the quotation can be used for March procurement." The observable result is whether the source states a supply period. When the source does not state a supply period, that statement is unsupported.
  3. Treating an empty citation array as a workflow failure. The input is an empty array output by the search node. The comparison object is the design of the downstream answering step. The observable result is whether the downstream path still runs and produces an answer without citations. An empty array is a documented allowed output length. It does not mean the node errored.

How to verify the setup

  1. Input the three source sentences of the hypothetical sample and the draft sentence. Compare the draft against the source on subject, authorized scope, and quotation validity period. Observe whether the draft uses only information stated in the source. If the draft presents the quotation validity period as a supply period, judge it unsupported.
  2. Input a retrieval question containing only a supplier name. Compare the returned citations against the supplier record source text. Observe whether the citations come from associated knowledge bases that use the same vector model. If results come from a base with a different vector model, record that the association configuration does not meet the requirement.
  3. Input a clearly defined empty citation array. Compare the output of the downstream answering step. Observe whether the downstream path still runs and preserves the no-citation state. If the downstream path stops because of the empty array, record that the handoff design needs adjustment.