What the documents look like
Copy drafting for a Tmall flagship store draws on four document groups: product publishing pages, brand authorization letters, delivery commitment records, and after-sales tickets. Compared with JD POP stores, Tmall flagship stores place more weight on brand authorization, flagship store identity, and Tmall points compensation rules. Authorization subject, authorization scope, and commitment wording that appear in drafted copy need to correspond to these materials. The source data updates daily and per order, so retrieval faces the same material at different points in time, and drafting has to establish which stage a cited material belongs to.
The retrieval step therefore works on mixed material: a publishing page that carries category qualifications, an authorization letter with a stated validity window, a delivery commitment record tied to a specific order, and an after-sales ticket with its own status. Copy sentences that mention authorization duration, authorized scope, or delivery promises must be traceable to one of these documents at a known stage.
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 with different vector models cannot be associated together. Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval; hybrid retrieval merges multiple result paths through RRF. ReRank performs a second ordering of candidate text. The citation cap is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reordering is used. Question optimization performs coreference resolution and question expansion. Output citations are an array whose length can be 0, so the downstream path continues even when nothing is retrieved.
Retrieval returns content only. It does not verify authorization, confirm flagship identity, or decide whether a drafted sentence may be published. A retrieved fragment shows that content matched; it does not show that an authorization is valid, that a commitment was met, or that a ticket has been approved.
Hypothetical paper exercise: the source material reads, "Authorizing party A, authorized party B, the authorized object is the home appliance flagship store operated by B on the Tmall platform, authorization validity from January 1, 2024 to December 31, 2024, authorization scope limited to the sale and display of A-brand goods within that store." The drafted sentence under review reads, "All A-brand goods sold in this store carry long-term authorization from the brand owner and can be purchased with confidence." The comparison checks the authorization duration wording in the drafted sentence against the start and end dates stated in the source material, and checks whether the authorized object matches the store named in the draft. The source material states a fixed one-year window and does not contain the phrase "long-term." The drafted sentence is therefore not supported by this fragment for the duration wording. This exercise uses no real customer, company, or identity number.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the product copy base and the authorization letter base that share one vector model | Compare the vector model identifier of each base and exclude any base whose identifier differs |
| Search mode | Hybrid retrieval | Compare against semantic retrieval alone and full-text retrieval alone on the same test question |
| Citation cap | Set to a token amount that covers one complete authorization letter paragraph | Compare against a shorter cap and observe whether the validity-window sentence is truncated |
| Minimum relevance | Begin at a middle value | Compare against higher and lower values and observe whether unrelated fragments decrease |
| Question optimization | Enable for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the authorization subject was altered |
| ReRank | Enable when candidates are numerous | Compare ordering with and without it and observe whether the authorization sentence moves forward |
These values are hypothetical starting points for this exercise. They are not official defaults.
Three common errors
- Treating retrieval output as authorization verification. Input the drafted sentence and the retrieved fragment, then compare whether the fragment states the authorization subject, object, and validity period. A retrieved fragment indicates that content was matched; it does not indicate that authorization holds.
- Associating knowledge bases across vector models. Input the vector model identifier of each base and compare whether the identifiers agree. Bases with differing identifiers are excluded, and cross-model retrieval is not executed.
- Writing a pending status as already effective. Input the current status field of the ticket or authorization letter and compare whether the drafted sentence claims effectiveness. When the status is pending review, the effectiveness wording in the draft has no support.
How to verify the setup
- Input the source material and the drafted sentence from the hypothetical exercise above, then compare the authorization duration wording in the drafted sentence against the start and end dates stated in the source. The wording is supported only when the subject, the field, and the value all match; otherwise record it as unsupported or pending verification. The source states a fixed window from January 1, 2024 to December 31, 2024, and the drafted sentence claims long-term authorization, so the duration wording is recorded as unsupported.
- Input the vector model identifier of each knowledge base and compare whether the identifiers agree. Association is completed only when the identifiers agree. Bases with differing identifiers are excluded, and the actual observation is recorded. No cross-model retrieval is run to compare missed recalls.
- Input an explicit empty citation array, written as [], and check whether the downstream node continues execution on the empty result. This check verifies only the empty-result continuation design. It does not infer that a specific retrieval rule will return an empty array.