What the documents look like
Creator commission settlement and deduction happen after a sale closes. The objects staff reconcile are valid promotion orders and after-sales commission deductions. In store service and internal training, the working materials are commission setting records, settlement statements, refund deduction details and invoices. These materials update weekly or monthly, and retrieval needs cluster around two questions: which setting a given deduction rests on, and which period a given statement covers.
The form of the material shapes the retrieval constraint. Commission setting records commonly carry a rate, an effective scope and a creator identifier. Refund deduction details carry an order, a refund status and a deducted amount. Invoices carry an issuing entity and an amount. Field names and conventions are not unified across the three material types, so the same transaction can be described differently in different materials. The retrieval step only returns relevant fragments as reference content. It does not decide which material carries settlement effect.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases on different vector models cannot be associated together. At implementation time, check the vector model identifier first, exclude any that do not match, and record the actual observation.
A single retrieval combines question optimization, semantic retrieval, full-text retrieval or hybrid retrieval, image retrieval, RRF fusion and ReRank according to the input content and search parameters, then produces reference content after similarity filtering and reference limit trimming. These paths are optional combinations. Image retrieval requires image input and the corresponding model capability. ReRank requires a model and must be enabled by configuration. Not every request executes all of them.
The reference limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Similarity is a relevance measure and does not perform permission or entity isolation. Retrieval results do not constitute authorization. Reference content is output as an array and its length can be 0. When nothing is retrieved, that output path still completes.
Hypothetical drill sample: a store training material, "Commission Setting Verification Note", states: "Creator A has a commission rate of 8% for the period 1 March 2025 to 31 March 2025, applicable to the apparel category. This setting takes effect from 1 March 2025." In the same batch, a "Refund Commission Deduction Detail" records: "Order A was refunded on 20 March 2025, with a commission deduction of 40 yuan." The handover request to be assessed is: "Please confirm Creator A's commission rate for Order A in March 2025 and execute the commission deduction." That request asks for a business action. Retrieval and forms do not constitute authorization for that action.
Configuration and implementation checks
| Configuration item | Relevant value or check for this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Associate the bases holding commission setting records and refund deduction details; check that vector model identifiers match | The two material types in the sample can be associated together only under the same vector model |
| Search mode | Compare semantic, full-text and hybrid retrieval against test questions; do not assume one mode suits all materials | Rate and date queries can be compared across full-text and hybrid recall |
| Reference limit | Unit is tokens; test with actual materials of different lengths before adjusting | The sample setting note is short; check whether the effective scope sentence is recalled in full |
| Minimum relevance | Takes effect only in semantic retrieval or when result reranking is used; label relevant and irrelevant samples before tuning the threshold | Used to filter fragments weakly related to the deduction question |
| Question optimization | Used for coreference resolution and question expansion; compare the original and optimized question and check whether the subject was altered | In a follow-up such as "how much is deducted on that one", check whether the subject remains Order A |
| Result reranking | Requires a ReRank model enabled by configuration; the score is used for filtering | When candidates are numerous, compare ordering before and after reranking |
Hypothetical starting value for this drill: the reference limit is set to 2000 tokens. The reason is that a single setting note in the sample runs to roughly a hundred characters, and both the effective scope and the date range need to be held in full. The comparison objects are 1000 tokens and 4000 tokens, checked for recall completeness on the same question. This number is a starting point for this drill, not an official default.
Three common errors
- Treating a retrieval hit as authorization to deduct commission. In the sample, the request to be assessed asks to "execute the commission deduction". Retrieval only returns reference content. Deduction is a business action that needs a separate system implementation and the corresponding authorization. Retrieval results do not constitute that authorization.
- Forcing knowledge bases on different vector models into one association. If the two material types in the sample have mismatched vector model identifiers, the correct action is to check the identifiers, exclude the mismatched one and record the observation. No conclusion about joint operation can be drawn afterwards.
- Judging an unstated fact as a conflict. The sample setting note does not record a specific deduction amount for Order A, while the refund detail records a 40 yuan deduction. The fields differ, so this is unsupported or pending verification, not a contradiction.
How to verify the setup
- Input the sample source text with the question "What is Creator A's commission rate for March 2025?" and compare whether the recalled fragment contains both "8%" and "1 March 2025 to 31 March 2025". The condition is met when the reference content contains both the rate and the period. If only the rate is recalled and the period is missing, record it as incomplete recall. This is the expectation; it was not tested.
- Input the handover request to be assessed and compare whether the output presents "execute the commission deduction" as completed or approved. The condition is met when the output only presents the reference materials and marks the request as pending verification. If the output states the deduction has been executed, record it as an overreach in wording. This is the expectation; it was not tested.
- Use a paper input of an empty array to check the downstream design, confirming that the path still completes when the reference length is 0 and that downstream does not conclude "no commission deduction obligation" from it. The condition is met when downstream marks the empty result as pending verification. If downstream treats the empty result as a negative conclusion, record it as a design gap. This is the expectation; it was not tested.