What the documents look like
Copy drafting for a Tmall flagship store draws on four document families: product publishing pages, brand authorization letters, delivery commitment records, and after-sales work orders. Brand authorization letters and flagship-store identity requirements mean that any statement about brand ownership or the scope of authorization in drafted copy needs a checkable source text before it can be used. Delivery commitment records update per order, and after-sales work orders are generated per order, so both can serve as sources for timing or service statements in copy. Among the daily documents, verifying category qualifications and maintaining the 48-hour delivery commitment consume the most time. The copy step inherits facts already stated in these materials rather than creating new ones.
Because the category places weight on brand authorization and flagship identity, a drafted sentence that names a brand, a series, an authorization period, or a delivery promise must trace back to a specific material. Retrieval output is a candidate citation; whether it is adopted still depends on checking against the effective version of the material.
Constraints at the citation tracing step
Each promotional sentence in the copy needs to correspond sentence by sentence to source text. The knowledge base search node can associate one or more knowledge bases that share the same vector model, and the reference content is output as an array whose length can be zero, meaning the output path still runs even when nothing is retrieved. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. Multiple recall paths are merged through RRF, reranking is enabled according to settings, and image retrieval requires corresponding image input and model capability. These optional paths do not mean every request executes all of them.
Hypothetical drill sample: Source text one reads "The A-brand X-series insulated cup sold by this store, authorization period from January 1, 2025 to December 31, 2025." Source text two reads "A-brand X-series insulated cup order, committed to ship within 48 hours after payment." The drafted sentence to be judged reads "A-brand X-series insulated cup authorized for the full year, ships within 48 hours after ordering." This sentence contains four fact types: brand, series, authorization period, and delivery commitment, and each must be compared against the period and fields in the two source texts.
Configuration and implementation checks
| Configuration or check item | Hypothetical trial starting point for this page | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Only one product materials base with the same vector model is selected | Knowledge bases with different vector models cannot be associated together; verify the identifier first and exclude non-matching ones |
| Search mode | Hybrid retrieval | The copy contains model numbers and proper nouns, so both semantic and keyword hits need to be observed |
| Citation limit | Hypothetical trial starting point of 2000 tokens for this page | This value is a trial starting point used to compare against the default citation limit and observe changes in the number of recalled items |
| Minimum relevance | Hypothetical trial starting point of 0.5 for this page | Trial starting point compared against a higher threshold to observe whether irrelevant fragments decrease |
| Result reranking | Enabled according to settings | Reranking requires a model and is enabled according to settings; compare candidate order before and after reranking |
| Answer and citation check | Compare answer sentences against cited source text one by one | Manual check action to verify whether citations cover each field of the drafted sentence |
Three common errors
- Treating "full year authorization" in the drafted sentence as an already-verified fact. The input is the drafted sentence and source text one; compare the authorization period field. When the start and end dates in the source text do not cover the full year, the statement is unsupported.
- Using retrieval similarity in place of source verification. The input is the citation array returned by retrieval; compare the cited source text against the fields of the drafted sentence. A similarity score does not constitute evidence of authorization scope or delivery commitment.
- Treating authorization material under review as already-effective authorization. The input is material marked with a pending-review status; compare its status field. Pending review does not demonstrate that authorization has taken effect, and it does not demonstrate the absence of operating permission.
How to verify the setup
- Input source text one and the drafted sentence from the hypothetical sample, and compare the authorization period field. The condition is satisfied only when the period in the drafted sentence falls within the start and end dates of the source text; when it exceeds them, record it as unsupported.
- Input source text two and the drafted sentence, and compare the delivery commitment field. The condition is satisfied only when the 48 hours after payment stated in the source text matches the drafted sentence; when it is missing or different, mark it as pending verification.
- Use a paper input of an empty citation array to check the downstream design, and observe whether the downstream continues along the output path with the empty array. Do not claim that a fictional rule retrieval will necessarily return an empty array.
Knowledge bases with different vector models cannot be associated together. Verify the identifier only, exclude non-matching ones, and record the actual observation. For material timeliness checks, use the current version and historical version documents to test the source; whether a citation is adopted still requires review against the effective version.