What the documents look like
Commission settlement and deduction for creators happens after a sale closes. The records to reconcile cover valid promotion orders, refund-related commission deductions, settlement statements and invoices. The main sources are the selected alliance backend, order after-sales data, the finance system and creator contracts, with weekly or monthly updates. These materials carry different scopes and layered statuses, so the retrieval library has to hold text that does not share one format. Commission setting records describe rates and effective scope. Settlement statements record what is payable and what has been paid for a period. Refund deduction details record after-sales adjustments to commission. Invoices record billing and payment evidence. Questions in order fulfilment and after-sales support often carry an order number, an after-sales number, a creator name and a period, and a retrieved result has to point back to specific fields before it can support an answer. Updates arrive weekly or monthly; this page does not derive batch synchronisation, real-time triggers or transaction-linked actions from that cadence.
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. During reconciliation, check the vector model identifier first and exclude any that do not match; do not run cross-model retrieval to compare missed recall. A single retrieval combines semantic search, full-text search, hybrid search, question optimisation and result re-ranking according to the input and the search parameters. The optional paths shown in the overall flow diagram do not mean every request executes all of them: image retrieval requires a corresponding image input and model capability, and re-ranking requires a model and must be enabled by configuration. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic search or when result re-ranking is used. Similarity and relevance are ranking and filtering inputs; they do not perform permission, entity or business-ownership isolation. A retrieved commission record does not mean that record may be disclosed externally or has been approved. The node outputs cited content as an array, and the length can be 0, which means the downstream chain still runs when nothing is retrieved. RRF is the fusion process for hybrid retrieval results and is not written as an on/off switch. Question optimisation performs coreference resolution and question expansion and adds one model call.
Hypothetical paper test sample: the source text supplied for reconciliation reads, "Creator A, March 2025 settlement statement: commission payable 4,800 yuan, of which order A refund deduction is 300 yuan, actual payment 4,500 yuan; the statement status is pending finance review." A separate draft sentence to be judged reads, "Creator A's March 2025 commission has been settled and the invoice has been issued." This draft sentence concerns the subject Creator A, the object being the March 2025 settlement statement, the status field and the invoice field.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting value for this trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate the four libraries for commission setting records, settlement statements, refund deduction details and invoices | Check whether the vector model identifiers match across libraries first; exclude any that do not match and do not perform cross-model association |
| Search mode | Run semantic search and full-text search once each, then run hybrid search | Compare the three returned sets with the same test question and observe whether order-number questions depend on full-text hits |
| Citation limit | Hypothetical starting value 3,000 tokens | Test with statement and detail fragments of different lengths to see whether content is truncated, and compare with 2,000 and 4,000 for completeness |
| Minimum relevance | Hypothetical starting value 0.5 | Label relevant and irrelevant samples first, then compare with 0.4 and 0.6 for false recall and missed recall |
| Question optimisation | Enabled for multi-turn follow-up scenarios, disabled for single-turn explicit questions | Compare the original question with the optimised question and check whether the subject or period has been altered incorrectly |
| Result re-ranking | Enabled when candidates are numerous and the question is complete | Compare ordering before and after enabling, and observe whether re-ranking score filtering changes the selected content |
The values above are hypothetical starting points for this trial and are not official defaults. The configuration table lists only the settings or check items relevant to this feature and to order fulfilment and after-sales support; database internals are not treated as a required item on every page.
Three common errors
- Treating a retrieval hit as authorisation. After entering "Creator A March 2025 settlement statement" and recalling the source text above, outputting the draft sentence "settled and invoice issued" directly. The source status field is pending finance review and the invoice field is not stated. After comparing the object and fields, this should be recorded as unsupported or pending verification.
- Associating libraries that use different vector models. With the same test question as input, if any associated library has a mismatched vector model identifier, check the identifier first and exclude it, then record the actual observation. Do not infer missed recall or redundant results from the joint association.
- Treating minimum relevance as a permission or entity filter. Enter an unrelated query and adjust minimum relevance, observe the change in the number of returned items, and do not conclude from that change that unrelated business or unauthorised records have been excluded.
How to verify the setup
- Enter the hypothetical source text and draft sentence above, and compare the subject, period, status field and invoice field. The source text states only pending finance review and does not state the invoice. The draft sentence's "settled and invoice issued" does not meet the conditions, so the output should be unsupported and the evidence gap retained. If the source text instead states the status as reviewed and the invoice field as issued, the same field values then form a checkable basis.
- Enter an explicit paper empty result of [], check whether the downstream chain continues to run with a citation length of 0, and observe whether an abnormal interruption appears. Do not claim that a fictional rule retrieval necessarily returns [].
- Enter an exact query containing an order number, run full-text search and semantic search separately, and compare whether the returned content includes that order number. Observe the full-text path hit situation and do not guarantee a first-position hit or full recall.
This feature has verified public technical facts for identifying related content in the documents; the body is written independently and does not copy them as a fixed five-row table. The final checklist is applied in the body above and is not reproduced here. This page addresses only creator commission settlement and deduction, order fulfilment and after-sales support, and knowledge base retrieval and recall. Which materials, operations and requests are in which authorisation state during an operations handover cannot be treated as authorised execution through retrieval or a form.