What the documents look like
Commission settlement and commission deduction for creators happen after a sale closes. The review focus is valid promotion and post-sale deduction. Drafting product and marketing copy draws on commission setting records, settlement statements, refund deduction details and invoices. These materials update weekly or monthly.
A commission rate, settlement cycle or deduction condition that appears in copy needs to map to a definite field in one of those documents. A statement that a settlement statement has been generated does not confirm a specific deduction rate. The retrieval step has to surface the field that carries the value, and the drafting step has to keep the value attached to its subject and period.
The four document types serve different drafting needs. Commission setting records carry the agreed rate and scope. Settlement statements carry the period, the creator and the amounts. Refund deduction details carry which orders were reversed and how the related commission was handled. Invoices carry the billing basis. Copy that cites a rate without the matching period, or a deduction without the matching order scope, is not yet supported by the retrieved material.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Search modes are semantic search, full-text search and hybrid search. Hybrid search merges multiple retrieval paths through RRF, and ReRank performs a second ordering of candidate text. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result re-ranking is used. Question optimization handles coreference resolution and question expansion, and it adds one model call before retrieval.
Retrieval does not perform permission checks and does not separate records belonging to different subjects or different periods. A relevance score is a filtering threshold, not a category, entity or permission filter. A pending review item is not an approval.
Hypothetical paper exercise sample: the source material reads, "The March 2025 settlement statement records that creator A's commission rate on valid promotion orders is 8%, and the commission corresponding to refunded orders is deducted back in the month the refund is completed. Invoices are issued based on the actual settled amount." The draft sentence under review reads, "Creator A's commission rate is 8%, and refund commission is deducted back in the same month." The subject and the rate field match the source. The deduction timing needs a comparison between "the month the refund is completed" and "the same month" to see whether the two phrasings describe one period.
The sample is a paper exercise. No test, deployment, approval or settlement action has taken place.
Configuration and implementation checks
| Configuration or check item | Value or action on this page | Basis and notes |
|---|---|---|
| Linked knowledge bases | Link only commission settlement knowledge bases that share one vector model | The node requires the same vector model; check identifiers first and exclude non-matching bases |
| Search mode | Start with hybrid search, and compare full-text search for numbered fields | Semantic and full-text retrieval suit different material; no single mode is assumed for all documents |
| Reference limit | Hypothetical trial starting value set to cover the fields of one settlement statement | The starting value follows the field length in this sample; compare completeness against a longer statement sample |
| Minimum relevance | Label related and unrelated samples first, then adjust the threshold | The value takes effect only under semantic search or result re-ranking |
| Question optimization | Enable for multi-turn follow-ups, and check whether the subject was altered | Used for coreference resolution and question expansion, and it adds one model call |
| Empty-result handling | Use a paper input returning an empty array to check the downstream branch | Reference content length can be 0, and the output path still runs |
The reference limit uses tokens as its documented unit. The minimum relevance value is a filtering threshold. Neither setting performs subject isolation or permission control.
Three common errors
- Reading "the settlement statement has been generated" as confirmation of a commission rate. Generation of a statement and the presence of a rate field are two different facts. When the input is the statement text, the comparison object is the rate field itself, and the observable result is whether that field appears. A statement without the field leaves the rate unconfirmed.
- Treating the minimum relevance value as a permission or subject filter. The value is a filtering threshold. When the input is creator A's draft sentence, the comparison object is the relevance score against the threshold, and the observable result is whether low-scoring fragments are trimmed. Subject separation still requires a separate check, and retrieval itself does not enforce it.
- Treating a draft under review as approved and published. When the input is the draft sentence and the source material, the comparison object is whether both share the same subject, the same period and the same field. The observable result is that any part not stated in the material is recorded as pending verification. A pending item is not an approval.
How to verify the setup
- Input the source material and the draft sentence from the hypothetical sample above, and compare the rate field "8%" in each for the same subject and the same period. The sentence is supported by the material only when the field values match and the deduction timing phrasings can be mapped to one period. If the periods cannot be mapped, record the sentence as pending verification.
- Input the same source material, and compare "the commission corresponding to refunded orders is deducted back in the month the refund is completed" with the draft phrasing "refund commission is deducted back in the same month." When the two phrasings do not align, record the sentence as pending verification. This is not a negative finding, and it does not establish that the deduction does not occur.
- Input a paper query that contains no commission field at all, and compare whether the returned reference content length is 0. When an empty array appears, the downstream branch handles it as an empty result. This observation does not show that unrelated queries always return empty, and it does not show that retrieval performs permission filtering.
A cross-model knowledge base cannot be linked together with another model's base. Check identifiers first, exclude non-matching bases, and record the actual observation. Image retrieval requires image input and a capable model, and re-ranking requires a model and its configured activation. Neither path runs on every request.