What the documents look like
Creator-hosted sales campaigns turn on external creator fulfillment and commission settlement. Brand business staff, MCN agencies and creator managers work from a common document set: creator cooperation agreements, commission setting sheets, sample shipment sheets and livestream reports. The underlying data comes from selected alliance records, creator platform data, creator history and store transaction details. Copy drafting for products and marketing depends on the subject, object, status and validity scope recorded in these materials. A cooperation agreement states the cooperation period and the authorization boundary. A commission setting sheet states the commission rate and settlement conditions. A sample shipment sheet states the shipped quantity and receipt status. A livestream report states the session transaction and refund basis. Any sentence about price, gift, commission or exclusivity needs to correspond item by item with those materials. Fields the materials do not state remain unknown, and evidence from a related stage supports only a pending conclusion.
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. Before retrieval, check the vector model identifier, exclude those that do not match, and record what was observed. Do not run cross-model retrieval to compare missed recalls.
Compare semantic retrieval, full-text retrieval and hybrid retrieval on the same test question. Do not assume one mode suits all material. The citation limit is measured in tokens; test with materials of different lengths to see whether content stays complete, then adjust the limit. Minimum relevance takes effect only under semantic retrieval or when result reranking is used; label relevant and irrelevant samples first, then adjust the threshold. Question optimization performs coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered. RRF is a fusion process for mixed results, not an on/off switch. Image retrieval requires a corresponding image input and model capability. Reranking requires a rerank model and must be enabled through its setting. The retrieval step does not handle permission or entity isolation, and a similarity score is not a permission filter.
Hypothetical paper exercise sample: the source material reads, "Party A commissions Party B to promote product A in Party B's livestream from June 1, 2025 to June 30, 2025, with a commission rate of fifteen percent of transaction value; two samples were received on May 20. This agreement takes effect from the date both parties sign." The draft sentence to be assessed reads, "This cooperation has been signed and is effective, commission settles at fifteen percent of transaction value, and all samples have been received." Checkpoints: signing, cooperation formation and contract effectiveness are three distinct facts; the material states only the effectiveness condition, not the actual signing date or signing status. The draft sentence is unsupported on signing status and effectiveness. The sample receipt quantity matches the material, and the commission rate matches the material.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the creator cooperation agreement base and the commission setting sheet base that carry the same vector model identifier | Check the vector model identifier first, compare with non-matching bases, and exclude them |
| Search mode | Start from hybrid retrieval | Compare recall content on the same test question against semantic retrieval and full-text retrieval separately |
| Citation limit | Start from the token amount needed to cover the complete clauses of one agreement | Compare with a smaller limit and observe whether clauses are truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then set a starting value between 0 and 1 | Compare with adjacent values and observe whether irrelevant passages are included |
| Question optimization | Enable for multi-turn follow-up questions that contain references | Compare with the disabled state and check whether the subject was altered |
| Result reranking | Use when a rerank model is available and enabled through its setting | Compare with unranked results and observe ordering changes |
These values are hypothetical starting points for this exercise, not official defaults. The specific values for citation limit and minimum relevance depend on actual material length and labeled samples.
Three common errors
- Treating a retrieval score as a permission decision. Input an out-of-scope query, compare the returned content with the authorized scope of the target base, and observe whether unauthorized records appear. A similarity score does not perform permission filtering, and the retrieval step does not handle entity isolation.
- Treating a pending review or pending signature as already effective. Input material that states only an effectiveness condition and no actual signing status, compare the effectiveness wording in the draft sentence with the material fields, and observe whether a condition was written as a fact. Pending review does not prove a lack of operating permission, and an unsigned document does not constitute a completed signature.
- Associating knowledge bases across different vector models. Input two bases with different vector model identifiers, compare the identifier fields, and observe whether they are associated together. Cross-model bases cannot be associated together; check identifiers only, exclude non-matching ones, and record the observation without inferring runtime results.
How to verify the setup
- Input the agreement text and the draft sentence from the hypothetical sample, and compare whether the signing status, effectiveness fact and commission rate in the draft sentence correspond field by field with the source text. The source states only the effectiveness condition while the draft sentence writes it as already effective, so that sentence is assessed as unsupported; the commission rate matches the source, so it is assessed as supported. The condition is met when each field yields a supported, unsupported or pending conclusion.
- Input a multi-turn follow-up question that contains a reference, compare the retrieval terms before and after question optimization, and observe whether the subject was replaced or expanded incorrectly. The condition is met when the optimized retrieval terms retain the original subject and the expansion introduces no object absent from the source.
- Input the same test question through semantic retrieval, full-text retrieval and hybrid retrieval, and compare the three recall paths with the results after citation limit trimming. The condition is met when the actual return of each path and the empty-result branch are recorded, without using one path to infer the behavior of another. The empty-result handoff can be checked against the downstream design using the explicit paper input []; do not claim that a fictional rule retrieval necessarily returns [].