What the documents look like
Creator commission settlement happens after a sale closes. The objects being reconciled are valid promotion orders and after-sales commission deductions, so the materials carry both a period attribute and a status attribute. Commission setting records state a rate and an effective scope. Settlement statements state a period and an amount. Refund deduction details state the deducted object and the reason. Invoices state the issuing entity and the amount. These materials come from the selected-alliance backend, order after-sales data, the finance system and creator contracts, and they are summarized weekly or monthly. Operating reports and exception reviews ask whether the same entity, the same period and the same field carry consistent values. The retrieval step only recalls relevant fragments. It does not judge authorization, approval or posting.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases whose vector model identifiers differ cannot be associated together. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Exact fields such as identifiers, rates and periods suit full-text retrieval, natural-language descriptions suit semantic retrieval, hybrid retrieval fuses multiple result paths through RRF, and ReRank performs a second ordering of candidates. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires image input and the corresponding model capability. Reranking requires an available model and must be enabled in the settings. Retrieval and similarity scores carry no permission or entity-isolation duty, so recalled fragments still need manual checking of entity and status.
Hypothetical drill sample: a knowledge base contains three source sentences. The first is "Settlement statement for March 2025: entity A, period March 2025, commission rate 5%, status confirmed." The second is "Refund deduction detail: entity A, original order date 12 March 2025, refund completion date 2 April 2025, deducted commission amount 200 yuan." The third is "Commission setting record: entity A, commission rate 5% from 1 March 2025, adjusted to 4% from 1 April 2025." One draft sentence to be judged is "Entity A's March 2025 commission is settled at a rate of 4%." Checking compares the draft sentence with the third source sentence. March 2025 falls inside the effective scope of 5%, and 4% takes effect from 1 April 2025. The draft sentence differs from the source on the same entity, the same period and the same field, so it is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only commission settlement bases whose vector model identifiers match | Check each base's vector model identifier first, exclude mismatches, and do not run cross-model retrieval |
| Search mode | Use full-text retrieval for fields such as "March 2025" and "5%", and semantic retrieval for deduction reason descriptions | Compare recalled fragments across the three modes per test question; do not assume one mode suits all materials |
| Reference limit | Start at a token amount that covers one complete settlement statement sentence | Test with statements and details of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Start at a middle value after labeling relevant and irrelevant samples | Takes effect only under semantic retrieval or result reranking; compare filtered results against the labeled samples |
| Question optimization | Enable for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether entity or period was altered |
| Result reranking | Enable when candidates are numerous and a rerank model is available | Compare ordering before and after reranking; the score is used for filtering, not as a permission judgment |
Three common errors
- Associating knowledge bases that use different vector models. The input is two bases with mismatched vector model identifiers, the comparison is against the node's association condition, and the observable result is that the combination does not meet the shared-association requirement. Check identifiers first and exclude mismatches; do not infer post-run recall behavior.
- Treating a retrieval hit as settlement authorization. The input is a recalled commission setting record fragment, the comparison is between the fragment content and the settlement action, and the observable result is that the fragment supplies only a rate value. Payment and posting need downstream implementation, and retrieval does not constitute authorization.
- Treating the order date in a refund deduction detail as the deduction period. The input is the second source sentence, the comparison is between the original order date and the refund completion date, and the observable result is that the deduction falls in the period containing the refund completion date. Grouping by order date conflicts with the source.
How to verify the setup
- Input the third source sentence and the draft sentence "Entity A's March 2025 commission is settled at a rate of 4%", compare entity, period and rate fields, and observe that the draft differs from the source, so it is unsupported. If the draft instead settles at 5%, it agrees with the source.
- Input the second source sentence and a draft sentence "Entity A's refund deduction for the March 2025 order is settled in March", compare the original order date with the refund completion date, and observe that the deduction corresponds to 2 April. The draft is unsupported and the evidence gap should be recorded.
- Input a paper empty result array and compare whether the downstream node continues to execute with a zero-length reference. The observable result is that the empty-reference path can proceed. This check addresses downstream handling design and does not assert that retrieval under any invented rule necessarily returns an empty array.