What the documents look like
Creator commission settlement happens after a sale closes. The verification focus is valid promotion and post-sale commission deductions. Source materials come from the selected alliance backend, order after-sales data, the finance system, and creator contracts. Everyday documents include commission setting records, settlement statements, refund deduction details, and invoices.
For retrieval, one settlement period involves several materials at once. A contract states the commission rate and the conditions for deduction. A statement lists the settlement amount. A refund detail record shows the after-sales offset. The subject, object, and period across these three material types must align before retrieval can supply usable evidence for contract-and-transaction comparison. The update frequency is weekly or monthly; retrieval configuration follows the existing document scope and does not imply an automatic synchronization mechanism.
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 using different vector models cannot be associated together. Check the vector model identifier first, exclude those that do not match, and record what is observed.
Retrieval recalls candidate content through semantic search, full-text search, or hybrid search. Hybrid search merges multiple result paths through RRF. ReRank performs a second ordering of candidate text. The reference cap is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reranking is used. A similarity score does not carry permission or entity isolation. The retrieval library scope does not equal the authorization scope.
Hypothetical paper exercise sample: a creator contract states, "The settlement period is the calendar month; the commission rate is fifteen percent of the valid order amount; commission corresponding to refunded orders is offset in the following month's statement." The settlement statement for the same period states, "Commission on valid orders for this period totals RMB 8,000; refund offset is RMB 1,200; actual payment is RMB 6,800." The draft sentence to be assessed is, "This creator's commission for the period is calculated at fifteen percent; refund commission has been offset; the actual payment amount matches the statement." During comparison, confirm that the contract subject, settlement period, and statement subject align, and that the refund offset amount matches the refund deduction details. The draft sentence does not state a statement number or a refund detail source, so those points remain to be verified.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | One contract base, one statement base, one refund deduction detail base | All three must use the same vector model; verify the identifier before associating |
| Search mode | Hybrid search | Contract clauses contain natural language; statements contain amounts and numbers; compare semantic and full-text recall differences |
| Reference cap | Hypothetical starting point of 3000 tokens | Test whether content stays complete against actual statement and refund detail lengths; compare with a smaller value |
| Minimum relevance | Hypothetical starting point of 0.3 | Label relevant and irrelevant samples first, then compare filtering effects |
| Question optimization | Enabled | Used for coreference resolution and question expansion; compare whether the subject is altered between the original and optimized question |
| Result reranking | Enabled according to settings | Reranking requires a model; compare candidate ordering before and after reranking |
Three common errors
- Treating semantic similarity as a permission filter. A recalled statement record may belong to another creator; the similarity score does not distinguish subjects. Check method: input the same question, compare the subject field of recalled content against the target creator identifier, and observe whether records for other subjects are mixed in.
- Associating knowledge bases that use different vector models. The documentation states that only knowledge bases with the same vector model can be associated. Check method: verify the vector model identifier of each base, exclude those that do not match, and record what is observed. Do not run cross-model retrieval to compare missed recalls.
- Treating a pending-review status as approved. The draft sentence involves a refund offset; when the refund detail is pending review, the deduction cannot be judged complete. Check method: compare the refund detail status field against the statement offset amount, and observe whether a pending-review status is still counted in the actual payment.
How to verify the setup
- Input the contract text and statement text from the hypothetical sample. Compare whether the contract subject, settlement period, and statement subject align. Observe whether the retrieval result returns both materials together with matching subject fields.
- Input the draft sentence "This creator's commission for the period is calculated at fifteen percent; refund commission has been offset; the actual payment amount matches the statement." Compare the draft sentence against the corresponding fields in the contract, statement, and refund details. Observe whether the statement number and refund detail source not stated in the draft sentence are marked as pending verification.
- Input an explicit paper empty result []. Check how the downstream design handles an empty reference. Observe whether the output path runs through. Do not claim that retrieval under a fictional rule necessarily returns an empty array.