What the documents look like
Dealer rebate calculation occurs after sales are complete, and it centres on policy interpretation and financial treatment. The working materials come from ERP sales records, dealer remittances, the contract ledger and the finance system. The routine documents are rebate policies, volume achievement tables, reconciliation statements and red-letter invoices. The update rhythm is monthly or quarterly, which differs from the immediate transaction pace of inventory replenishment. Disputes in rebate calculation typically arise from three steps: checking the achievement basis, deducting returned goods, and offsetting by invoice. Time gaps and definition gaps between materials need to be checked item by item.
Citation tracing requires that a conclusion be traced back to specific document text. A rebate policy states the calculation rule; a volume achievement table records the quantity for a period; a reconciliation statement confirms the mutual account between the two sides; a red-letter invoice corresponds to an offsetting action. The subject, object, period and status of the four material types must agree before a traceable citation can be supported.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model, and outputs citations as an array whose length may be zero. An empty array means the search chain completed without obtaining citation content, and the downstream design must be able to accept that situation. After citation content is passed to the answering step, the correspondence between each answer and the cited source text is checked one by one. Retrieval by itself does not constitute authorization, and it does not constitute business approval.
The retrieval scope is determined by the selected knowledge bases and the document version. Knowledge bases with different vector models cannot be associated together; the identifiers are checked first and non-conforming ones are excluded. The minimum relevance value is a filter threshold that takes effect only in semantic retrieval or when result reranking is used; it does not perform permission filtering or entity isolation. Hybrid retrieval merges multiple result paths through RRF, which is part of the retrieval process and not a separate switch that can be turned on. Image retrieval requires image input and the corresponding model capability; reranking requires a model and is enabled according to settings. Optional paths shown in the document flow diagram do not mean that every request executes all of them.
Hypothetical paper exercise sample: a brand and a dealer sign a 2025 annual rebate agreement. The agreement text states: "Annual purchase volume reaching 5,000,000 yuan is granted a 2% rebate; returned goods amounts are deducted from the purchase volume." A July 2025 reconciliation statement shows first-half purchase volume of 2,600,000 yuan and returns of 150,000 yuan in the same period. A draft sentence states: "The dealer has reached the 5,000,000 yuan target in the first half and a rebate can be accrued." The check inputs are the agreement text above, the reconciliation figures and the draft sentence. The comparison object is the achievement condition in the agreement against the actual basis in the reconciliation statement. The observed result is that the amount and the period in the draft sentence do not correspond to the agreement condition. This sample is a paper exercise and does not represent a real customer or a business action that has occurred.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Rebate policy base and contract ledger base | Check whether the two vector model identifiers are the same; exclude if they differ |
| Search mode | Hybrid retrieval | Policy clauses contain numbers and natural language; compare semantic recall and full-text recall results |
| Citation limit | Measured in tokens, starting point 3000 | Reconciliation paragraphs are long; compare 3000 against lower values to see whether citations cover key clauses |
| Minimum relevance | Starting point 0.5 | Compare against 0.3 and 0.7, and observe whether low-relevance fragments enter the citations |
| Result reranking | Enabled | When candidates are numerous, compare ordering before and after reranking, and confirm the rerank score is used for filtering |
| Citation content handoff | Check whether the output array is passed to the answering step | Use an empty array and a populated array separately to check downstream display |
The values in the table are starting points for this hypothetical exercise, chosen from the paragraph length of the materials and the number of candidates, and are to be adjusted after comparison with adjacent values. The citation limit uses the token unit and is unrelated to the thinking budget.
Three common errors
- Treating a retrieval hit as completed authorization or approval. Retrieval only returns citation content; rebate accrual, invoice offsetting and payment require downstream business system implementation. During checking, compare whether a retrieval output and an authorization record both exist, and record the gap when one is missing.
- Associating knowledge bases that use different vector models. The input is the model identifiers of the two knowledge bases; compare whether the two are identical, exclude when they are not, and do not run a cross-model retrieval to compare recall differences.
- Mixing the amount in the draft sentence with the agreement condition. The input is the agreement text and the reconciliation figures; compare whether the subject, period and field agree. When the amount or the period does not correspond, judge the sentence as unsupported, and do not judge it as contradictory.
How to verify the setup
- Use the agreement text, the reconciliation figures and the draft sentence from the hypothetical sample as input. Compare whether the amount and the period in the draft sentence correspond to the agreement condition. The observed result is that the draft sentence is unsupported. If the draft sentence is changed to "First-half purchase volume was 2,600,000 yuan, returns were 150,000 yuan, and the 5,000,000 yuan level has not yet been reached," it agrees with the reconciliation basis.
- Input two knowledge bases whose vector model identifiers differ. Compare whether the identifiers are identical. The observed result is that the system does not accept the joint association, or associates only the conforming one. Record the actual observation, and do not infer missed recall or redundant results.
- Input an explicit paper empty array
[]. Compare whether the downstream answering step accepts it normally. The observed result is that the downstream step does not error and does not generate a conclusion without a source. This check verifies the downstream design and does not mean that a retrieval against invented rules will necessarily return an empty array.