What the documents look like
TikTok Shop US operations differ from shelf-based Douyin store operations because the US storefront is shaped by local return windows, return grading, and overseas logistics constraints. Copy drafting draws on four recurring materials: product listings, return and refund records, return grading evidence, and settlement statements. These materials update daily or per order, so a sentence written today may rest on a listing version that is already superseded.
For citation tracing, the practical consequence is that each proposed sentence must be traceable to a specific original material, and the material version must be identified. A promotional claim that lacks material support can contribute to platform automatic refunds, higher return fees, or failed appeals. Retrieval scope is limited to the selected knowledge bases. The search node outputs cited content as an array, and the array length may be zero.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model. Knowledge bases built on different vector models cannot be associated together. The vector model identifier is checked first, and non-matching bases are excluded. Cross-model retrieval is not executed in order to compare missed recalls.
Search output is a citation array whose length may be zero. An empty array means the output path still runs through; it does not mean the workflow has errored or terminated. Citation content is checked line by line against the original text. Thinking process and answer fluency do not substitute for source evidence. Material timeliness is tested with current and historical document versions, and whether a citation is adopted still depends on review against the effective version. Minimum relevance is a filter threshold; it does not perform permission or entity isolation, and retrieval itself is not authorization.
Hypothetical paper exercise sample: a merchant holds a product listing draft. The material original text reads: "This product is a pure cotton T-shirt; washing water temperature must not exceed 30 degrees Celsius; orders placed between March 1, 2025 and March 31, 2025 qualify for free shipping." The proposed sentence reads: "This T-shirt is machine washable and ships free year-round." The object of comparison is the product listing draft, the effective scope is March 2025, and the authorization boundary is internal review of that listing only. The material does not state "year-round free shipping" or "machine washable." The proposed sentence exceeds the material scope and is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Comparison method |
|---|---|---|
| Associated knowledge bases | Select only knowledge bases sharing the vector model of the product listing | Check the vector model identifier and exclude bases on different models |
| Search mode | Semantic retrieval | Observe whether recalled content relates to the proposed sentence topic |
| Minimum relevance | 0.5 | Compare with 0.3 and 0.7, observe changes in entries after filtering |
| Citation cap | 2000 tokens | Compare with 1000 and 4000, observe cited content length |
| Question optimization | Off | Compare with the on state, observe whether the query terms are rewritten |
| Citation output | Array, length may be zero | Check whether downstream still runs through on an empty array |
The values 0.5 and 2000 are hypothetical trial starting points for this page. They were chosen to make filtering and truncation effects observable, and they are compared against 0.3, 0.7, 1000, and 4000 with actual observations recorded. These values are not official defaults.
Three common errors
- Treating the proposed sentence as the material original text. Input is the proposed sentence "This T-shirt is machine washable and ships free year-round." The comparison object is the material original text "washing water temperature must not exceed 30 degrees Celsius; orders placed between March 1, 2025 and March 31, 2025 qualify for free shipping." The observable result is that the proposed sentence contains "year-round free shipping" and "machine washable," which the material does not state, so it is judged unsupported.
- Treating an empty citation array as a workflow error. Input is the search node output
[]. The comparison object is the documented statement that the length may be zero and the output path still runs through. The observable result is that downstream nodes still receive the empty array and no error termination is triggered. - Treating minimum relevance as a permission filter. Input is minimum relevance 0.5. The comparison object is the documented statement that it filters low-relevance results only. The observable result is that this setting does not exclude unrelated businesses or unauthorized records, and retrieved results still require manual review of the authorization scope.
How to verify the setup
- Input the hypothetical material original text and the proposed sentence. Compare whether the proposed sentence contains "year-round free shipping" and "machine washable," which the material does not state. The observable result is that the proposed sentence is marked unsupported, and the "March 2025" scope in the material original text is retained.
- Input the search node output
[]. Compare whether downstream nodes still run through. The observable result is that downstream nodes receive the empty array without error, and the citation output path remains connected. - Input knowledge base identifiers for the same vector model and for a different vector model. Compare whether they can be associated together. The observable result is that the same-model base is selected and the different-model base is excluded, and cross-model retrieval is not executed.