What the documents look like
Creator-led sales campaigns depend on external creator performance and commission settlement rather than internal brand scheduling. Brand business teams, MCN agencies and creator managers work from four recurring documents: creator cooperation agreements, commission setting sheets, sample shipment forms and livestream performance reports. Each document belongs to a different party and a different stage. The cooperation agreement constrains rights and obligations between the two sides. The commission setting sheet fixes the settlement ratio and slot arrangement. The sample shipment form records physical handover. The livestream performance report summarizes single-session transaction and refund figures.
In cross-border work, the same material may exist in Chinese, English or another language version. Before citing any retrieved passage, the version being cited must be checked against the version currently in effect. Materials update per session or per week. The retrieval system does not decide which version is valid; it returns candidates by vector or full-text matching. Citation tracing requires mapping each returned item back to its source material and checking subject, object, status and validity scope one by one.
Constraints at the citation tracing step
The knowledge base search node outputs a citation array whose length may be zero. An empty array means the search path ran but returned no content; it does not mean the workflow terminated with an error. The citation array is passed to the downstream answer step, and each source in the answer should trace back to specific source text. Selected knowledge bases must share the same vector model; knowledge bases with different vector models cannot be linked together. This constraint concerns the vector model identifier only and does not extend to business registration sites or other external systems.
Hypothetical paper test case. Source text A reads: "Party A: Xinghai Brand Management (fictional); Party B: Creator Lin (fictional); campaign session date: 18 June 2025; commission ratio: 15% of sales; sample shipment: sent before 10 June 2025." Source text B reads: "Commission ratio: 20% of sales; applicable session: July 2025." The draft sentence to be judged is: "The 18 June session commission is settled at 20%." Checking compares the session date and commission ratio fields across the two sources. The 18 June session corresponds to 15%; the 20% figure is marked for the July session. The two sources share the same parties but cover different periods, so the draft sentence cites a field from the wrong period.
Retrieved citations may come from materials in different languages. Semantic retrieval supports cross-language matching, but whether a retrieved item is adopted still requires manual comparison against the source text. Minimum relevance is a filter threshold that takes effect only in semantic retrieval or with result reranking; it does not perform permission or entity isolation. The citation limit is measured in tokens and controls the maximum content cited per search.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this test | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Chinese and English creator cooperation libraries sharing one vector model | Check that vector model identifiers match; knowledge bases with different models are not linked together |
| Search mode | Hybrid retrieval | Commission ratio and date fields need full-text hits; the semantic path covers natural-language phrasing |
| Citation limit | Hypothetical starting point of 3000 tokens | Compare against the combined text length of one agreement plus one commission sheet to cover key fields |
| Minimum relevance | Hypothetical starting point of 0.3 | Compare against the score distribution of returned results and observe whether valid items are filtered out |
| Question optimization | Enabled | Completes references in multi-turn follow-up; a single clear question may leave it off to save time |
| Result reranking | ReRank enabled | Secondary ordering when candidates are many; not enabled when no rerank model is available |
These values are hypothetical starting points for this test, not official defaults. Adjust one item at a time and compare against the previous result.
Three common errors
- Treating retrieval recall as authorization confirmation. The search node returns a citation array, which only shows that content matched. It does not show that a cooperation agreement was signed, that a commission ratio took effect or that a sample was shipped. A pending-review status does not show that an operation lacks authorization; signing, cooperation reached and contract effective are three separate facts.
- Linking knowledge bases across different vector models. Knowledge bases with different vector models cannot be linked together. Check the model identifier first and exclude those that do not match. Do not invent missed recalls or redundant results from a linked cross-model search; record only what is observed.
- Using a relevance score in place of field comparison. Minimum relevance is a filter threshold and does not judge whether subject, period or field match. When a draft sentence cites a 20% commission, return to the source text and compare the session date and ratio fields; a high score does not make the citation correct.
How to verify the setup
- Input the two source texts and the draft sentence from the hypothetical sample above. Compare whether the draft sentence's session date and commission ratio fields match the same source material at the same time. If the draft sentence states settlement at 20% for 18 June while 20% is marked only for the July session, record it as unsupported and note the evidence gap.
- Input a commission setting sheet explicitly marked "pending review" together with a request to "execute settlement per this sheet." Compare the requested action against the material's status. Pending review does not equal approved; retrieving that sheet does not constitute an execution basis. Record the missing approval status.
- Input an empty citation array [] and observe the design of the downstream answer step. Check whether the downstream step handles an empty array explicitly. Do not assume a fictional-rule search necessarily returns [], and do not guarantee a first hit or full recall.
This page addresses only creator-led sales campaigns, cross-border materials and multilingual retrieval, and citation tracing. During operational handover, which materials, operations and requests sit at which authorization state must be kept separate; retrieval or a form is not authorization to execute.