What the documents look like
Creator sales campaigns depend on external creator performance and commission settlement rather than internal brand scheduling. Brand business teams, MCN agencies and creator managers work across four document types: creator cooperation agreements, commission setting sheets, sample dispatch forms and livestream reports. Inventory procurement and supplier materials cross these documents. An agreement may assign stock preparation responsibility, a commission setting sheet may tie into procurement cost definitions, a sample dispatch form records where materials went, and a livestream report allows reverse inference about inventory consumption. Screening creators, negotiating commission and slot fees, and reconciling after-sales commission deductions are the most time-consuming steps. Source material is scattered across selected alliance platforms, creator marketing platforms, creator historical data and store transaction details.
These documents differ in subject, object, status and validity scope. Vector retrieval recalls candidate original text by semantic similarity. It does not determine whether an agreement has taken effect or whether a procurement request has been approved.
Constraints at the vector indexes step
Knowledge bases associated for retrieval must use the same vector model. Check the actual model and input type first. One original data block can be linked to multiple indexes; indexes and original content are checked separately. Index content should be complete and focused. Retrieval coverage and index length are weighed against samples. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion time; checking only the query side is insufficient. After model or index changes, compare cited original text using the same question set. Do not promise outcomes based on vector scores alone.
Semantic recall and relevance scores do not handle permission or entity isolation. A similarity score does not confirm supplier qualification, authorization status or a cooperation relationship. The same-vector-model requirement for retrieval knowledge bases does not extend to business registration sites or other external systems. Knowledge bases with different vector models cannot be associated together. Check identifiers and exclude non-matching ones. Record only actual observations; do not infer runtime results. RRF is a fusion process for mixed results, not a standalone toggle. The citation limit uses tokens as its unit. Minimum relevance applies only during semantic retrieval or result reranking. Empty reference output can have length zero, and that path still executes.
Configuration and implementation checks
Hypothetical paper exercise sample: a brand business team retains a paper record. The original text reads: "On March 1, 2025, Party A and Party B signed a cooperation intent for creator campaign inventory procurement, agreeing that Party B would provide a sample dispatch form number before March 10, and Party A would start internal approval after receiving it. As of March 5, Party A had not received that number." A draft sentence reads: "Party B has completed sample dispatch, and Party A's inventory procurement has been approved." The task is to determine whether the draft sentence is supported by the original text.
| Configuration item | Setting for this page | Verification method |
|---|---|---|
| Associated knowledge bases | Select only knowledge bases using the same vector model as the inventory procurement and supplier materials | Check model identifiers; exclude those with different vector models |
| Index content | Build separate indexes for agreements, commission setting sheets, sample dispatch forms and livestream reports | Compare index original text against the draft sentence for subject, date and status |
| Citation limit | Hypothetical trial starting point of 800 tokens, because the sample material is short and complete clause sentences need to be retained | Compare against displayed value; observe whether citations are truncated |
| Minimum relevance | Hypothetical trial starting point of 0.4, because generic fragments need filtering | Compare against retrieval result scores; confirm the threshold applies only to semantic retrieval or reranking |
| Question optimization | Decide whether to enable based on actual results; short questions and multi-turn follow-ups may benefit | Check the completion result in conversation details |
| Result reranking | Takes effect when a ReRank model is available and enabled per settings | Check whether rerank scores and cited original text are consistent |
Three common errors
- Treating retrieval recall as authorization confirmation. Input "Has Party B received procurement approval?" The system recalls the agreement original text. That text records only intent and the sample dispatch form number requirement; it does not record completed approval. The error is equating recalled content with an approval fact.
- Writing a pending review status as completed. The draft sentence states "has completed sample dispatch," while the original text states only that Party A had not received the number. The error is treating unsupported stage evidence as a completed fact.
- Associating knowledge bases with different vector models together. A supplier material query is input, and the system associates a knowledge base with a different vector model. The error is ignoring the same-model requirement and executing retrieval without checking identifiers.
How to verify the setup
- Input the hypothetical original text and the draft sentence "Party B has completed sample dispatch, and Party A's inventory procurement has been approved." Compare the draft sentence against the original text on subject, date and status fields. The original text states Party A had not received the number; the draft sentence states completion. These are mutually exclusive values on the same field, so the draft sentence is judged unsupported. If retrieval returns only the original text and no approval completion record, the pending-verification condition is met.
- Input "sample dispatch form number" as the query term. Compare whether the returned citation includes the original sentence "provide a sample dispatch form number before March 10." If the citation limit is set to 800 tokens and the returned content retains that sentence completely, the citation-not-truncated condition is met. If the returned content omits that sentence, record it as pending review.
- Input an explicit paper empty result []. Check whether the downstream design continues execution along the empty-reference path. Do not claim that any fictional rule retrieval necessarily returns []. If the downstream node receives an array of length zero and continues, the empty-result handling condition is met. If the downstream does not define empty-array handling, record it as pending implementation.