What the documents look like
Commission settlement and commission deduction for creators take place after a sale is completed. The review focus falls on valid promotion and post-sale deductions. The working materials are commission setting records, settlement statements, refund deduction details, and invoices. Their sources span the selected alliance backend, order after-sale data, the finance system, and creator contracts. These materials sit in separate systems, and their field definitions and update rhythms do not match.
Citation tracing answers one question for this category: which material, and which version of that material, supports a given settlement or deduction conclusion. A commission rate written into a contract and a commission rate actually applied in a statement belong to the same field. When the two values are mutually exclusive, that is a contradiction. A deduction amount stated in refund deduction details and a statement that records no deduction belong to different fields, so the result is unsupported and requires separate verification. A settlement statement pending review does not mean an operator lacks permission. A quotation validity period does not mean a supply period. Signing, reaching cooperation, and contract effectiveness are three distinct facts.
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Its output is an array of citations, and the length may be zero. An empty array does not mean the workflow terminated with an error. Retrieval and similarity scores carry no permission or entity isolation. A retrieval result is not an authorization. Knowledge bases built on different vector models cannot be associated together; verify the model identifier first, exclude those that do not match, and record what was actually observed.
Hypothetical drill sample: the source material reads, "Creator A and Brand B signed the Live Promotion Cooperation Confirmation on March 1, 2026, agreeing on a 15% commission rate for the March special session, with settlement based on valid orders in the selected alliance backend." A second source reads, "The March 2026 settlement statement shows a 12% commission rate for Creator A's special session, on a commission base of 100,000 yuan." The draft sentence under review reads, "Creator A's March special session commission settles at 15%, with 15,000 yuan payable after deduction." The 15% in the draft and the 12% in the statement are mutually exclusive values in the same field. The 15,000 yuan figure lacks support from deduction details. The conclusion is unsupported and must be verified against refund deduction details and invoices before any determination.
Constraints at the citation tracing step
Contract and transaction material comparison requires the same entity, the same period, and the same field. Retrieved content must be checked line by line against the cited source text. A citation shown in an answer is not evidence that the underlying business fact was approved, paid, or executed.
The search node output is an array of citations. That array may be empty. When the array is empty, the downstream answer step still runs through the chain, and no source assertion should be produced from it. Retrieval does not grant permission, and a relevance score is not a category, entity, or permission filter. A high similarity score does not exclude unrelated businesses or unauthorized records.
The overall retrieval flow described in the documentation combines several optional paths. Semantic retrieval, full-text retrieval, hybrid retrieval, question optimization, image retrieval, and rerank are not all executed on every request. Image retrieval requires an actual image input and a model capable of processing it. Rerank requires a rerank model and must be enabled in settings. RRF is a fusion process for combining results from multiple retrieval paths; it is not a toggle that a user switches on. Thinking budget is a reasoning budget and must not be repurposed as a citation token limit.
Citation upper limits are measured in tokens, not in a guessed item count. The minimum relevance value applies only under semantic retrieval or when result reranking is used. It is a filter threshold, not a final sort rule. After question optimization, hybrid retrieval, image retrieval, or reranking, final results may come from multi-path fusion and need not follow a single vector similarity order.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this drill | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Include the libraries holding contract, statement, and deduction detail materials | Check the vector model identifier library by library; only matching models may be associated together |
| Search mode | Hybrid retrieval | Number, rate, and amount fields rely on full-text hits; natural language questions rely on vector recall |
| Citation upper limit | Hypothetical starting point set to cover one passage from each of the three material types | Compare against the actual length of a single settlement statement, then adjust |
| Minimum relevance | Hypothetical starting point at a middle value | Compare against the actual relevance distribution of retrieved results and observe whether the three material types remain covered after filtering |
| Question optimization | Enabled only for multi-turn follow-up questions | Compare against retrieval results for a single clear question |
| Result reranking | Enabled when a rerank model is available | Compare against candidate ordering without reranking |
Three common errors
- Treating a retrieval hit as an authorization. Retrieving a settlement statement only means the material was recalled within the selected knowledge base scope. It does not mean the commission was approved or that payment was executed. The check input is the retrieval output array; the comparison object is whether an authorization record exists; the observable result is that the authorization record remains missing.
- Treating different fields as a contradiction. A contract rate and a statement commission rate belong to the same field and can be judged contradictory. A statement that records no deduction amount and deduction details that state an amount belong to different fields, so the result can only be recorded as unsupported. The check input is the source text of both materials; the comparison object is the field name and its value; the observable result is that the field names do not match.
- Forcing association across vector models. Knowledge bases built on different vector models cannot be associated together. Verify the model identifier first, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.
How to verify the setup
- For the hypothetical sample above, the input is the contract source text and the statement source text. The comparison object is the commission rate field. The observable result is that the two values differ, which is judged a contradiction, and the evidence gap is retained.
- For the draft sentence "Creator A's March special session commission settles at 15%, with 15,000 yuan payable after deduction," the input is that draft sentence and the deduction details. The comparison object is whether a deduction amount field exists. The observable result is that when the details are missing, the conclusion is judged unsupported.
- For empty citation handling, the input is a clearly stated paper output array
[]. The comparison object is whether the downstream answer step continues through the chain with zero citations. The observable result is that the chain runs through and produces no source assertion.
These checks describe expected observations. They do not claim that any test, deployment, approval, or payment has occurred. Where a material field is not provided, the value remains unknown.