What the documents look like
Creator special-event sales partnerships depend on external creator fulfillment, and commission settlement plus post-sale commission deductions are the main sources of dispute. Product material and selection Q&A must work with four document types: creator cooperation agreements, commission setting sheets, sample shipment sheets, and live-stream battle reports. These materials update on a per-event and weekly rhythm, and version differences between materials for a single event are common. The selection Q&A step must answer whether a given item in a given material can be cited as a source, and whether that citation can be traced back to the original text.
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length may be zero. Whether cited content is accepted depends on the document version and its effective scope, not on how fluent the answer reads. Retrieval output must be handed to the answering step, and each answer sentence must then be checked against the cited original text.
Constraints at the citation tracing step
The knowledge base search node accepts one or more knowledge bases built on the same vector model. Its output is a cited-content array, and a length of zero is permitted, meaning the downstream path still runs even when nothing was retrieved. Retrieval is not authorization: a hit only returns cited content and does not show that a sample was shipped, that commission was settled, or that an agreement took effect. The cited-content output must be passed to the answering step, and answer sentences must be compared line by line against the cited original text. The relevance score is a filtering threshold, not a category, entity, or permission filter, and it does not exclude unrelated businesses or unauthorized records.
Hypothetical paper exercise sample: the material original text reads, "Party A commissions Party B to promote product A during the special-event live stream on 18 June 2025, with commission settled at 8% of actual transaction value. The sample shipment sheet states that 2 units of product A sample were shipped on 10 June 2025. The live-stream battle report shows the transaction value for that event was 0 yuan." The draft sentence to be judged is: "The commission for this event has been settled at 8%." This draft sentence differs from the material original text in subject, period, and field: the material records only the commission rate and does not record settlement status.
Configuration and implementation checks
| Check item | Hypothetical starting point for this exercise | Comparison object and reason |
|---|---|---|
| Associated knowledge bases | Select only the creator cooperation agreement base | Check separately against the commission setting sheet base and confirm the vector model identifier matches |
| Search mode | Hybrid retrieval | Compare with full-text retrieval results; identifier-type fields are checked first for full-text hits |
| Citation limit | Hypothetical starting point of 800 tokens for this exercise | Compare with the 600-token result and observe changes in the number of cited items |
| Minimum relevance | Hypothetical starting point of 0.4 for this exercise | Compare with the 0.3 result and observe whether low-scoring fragments enter the citations |
| Question optimization | Not enabled for now | Compare with the enabled result and observe whether coreference completion changes retrieval |
| Output handoff | Citation array passed to the answering node | Check that answer sentences correspond line by line to the cited original text |
These starting values are exercise assumptions, not official defaults; each is chosen so that a single value change can be compared against a neighbouring value. The knowledge base search node's cited-content output is an array that may be empty, and an empty array does not mean the workflow has errored out. Whether a cited item is adopted still requires checking the effective version of the material. The same-vector-model requirement applies only to associating knowledge bases within this node; it must not be extended to require the same model for business-registration sites or other external systems.
Three common errors
- Treating a retrieval hit as authorized execution. Retrieval returns only cited content and does not show that a sample was shipped, that commission was settled, or that an agreement took effect. A pending review state is not evidence that an operation lacks permission; pending review states only the current status.
- Treating a quotation validity period as a supply period. These are different fields and cannot substitute for each other. Likewise, signing, reaching cooperation, and contract effectiveness are three distinct facts.
- Treating an unsigned termination as a completed termination, or an offered option as execution. An unprovided document field remains unknown, and missing evidence must stay missing rather than being filled in with an assumed authorization or effectiveness clause.
How to verify the setup
- Input the material original text and the draft sentence from the hypothetical sample above. Compare the draft sentence against the original text on subject, period, and field, and observe whether the same field shows mutually exclusive values. If the original text does not record settlement status, the draft sentence is judged unsupported. Only a same-subject, same-period, same-field explicit conflict counts as a contradiction; a merely related stage is unsupported or pending verification, not a negation or conflict.
- Input the phrase "has been settled at 8%" from the draft sentence. Compare it against the fact that the material original text records only the rate, and observe whether the answer rewrites the rate into a settlement status. If such a rewrite appears, the citation is judged inconsistent with the original text.
- Input an empty citation array
[]. Compare it against the downstream answering node's handoff design and observe whether the empty array is handled as a normal empty result. If the downstream treats the empty array as an error, record that design difference. This paper input checks the downstream design only; it does not claim that a fictional rule makes retrieval return[].
For each check, state what is compared and what remains unknown. Without a documented error code or an actual observation, record the gap rather than asserting a deterministic failure cause or result. Optional paths listed in the overall document flow, such as image retrieval or reranking, are not executed on every request; image retrieval requires a corresponding image input and model capability, and reranking requires a rerank model enabled by configuration. When knowledge bases use different vector models they cannot be associated together; check the identifier, exclude non-matching bases, and record what is actually observed rather than inferring runtime outcomes.