What the documents look like
Dealer rebate calculation happens after sales close, so the work centres on policy interpretation and financial treatment rather than on stock movement. The main data comes from ERP sales records, dealer remittances, the contract ledger and the finance system. The documents handled day to day are rebate policies, sales achievement tables, reconciliation statements and red-letter invoices. The most time-consuming steps are reconciling the achievement basis, deducting returned goods amounts and offsetting through invoicing.
For operating reports and exception reviews, these characteristics mean that every cited item has to point back to a specific document, a specific position in that document and a specific version. Summary figures and verbal confirmations do not carry the same traceability. A rebate conclusion that cannot be traced to the underlying policy clause, achievement figure or deduction line remains open.
Constraints at the citation tracing step
Citation tracing requires that each cited item can be traced to the original document's location and version. The knowledge base search node outputs citations as an array, and the length may be zero. Even when no content is found, the output path still runs. This means an empty citation result needs human assessment: it may indicate missing data, or it may indicate that the search scope did not cover the target document. An empty array does not by itself mean there is no rebate obligation.
Hypothetical paper exercise sample: a brand and a dealer sign a 2025 annual rebate policy stating that a quarterly sales achievement rate of no less than 90% is required to qualify for a rebate. The sales department submits a 2025 second-quarter sales achievement table stating: "Dealer A's sales achievement rate this quarter is 92%, and the returned goods amount of RMB 150,000 has been deducted." The finance department drafts the sentence: "Dealer A meets the 2025 second-quarter rebate conditions and a rebate should be accrued." The question to decide is whether this draft sentence matches the sales achievement table wording, and whether the returned goods deduction basis matches the calculation method stated in the rebate policy.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Rebate policy base, sales achievement table base | Check that both bases use the same vector model; bases with different vector models cannot be associated together |
| Search mode | Hybrid search | Identifier and model-number queries rely on full-text search; natural-language questions rely on semantic search |
| Citation limit | Set to 3000 tokens as the hypothetical trial starting point | Chosen after comparing against the average length of a single reconciliation statement; this is not an official default |
| Minimum relevance | Set to 0.5 as the hypothetical trial starting point | Chosen after comparing against historical reconciliation check results; observe whether filtering omits key clauses |
| Question optimisation | Enabled | For follow-up questions about rebate clauses, coreference resolution can improve retrieval completeness |
| Result reranking | Enabled | When candidate reconciliation statements are numerous, rerank scores can be used for secondary ordering |
Three common errors
- Treating retrieval output as an authorisation basis. The knowledge base search node outputs a citation array, and that array does not constitute rebate approval or payment authorisation. When checking, distinguish retrieval results from business approval records. A pending review status does not demonstrate that authorisation has been granted.
- Associating knowledge bases that use different vector models. Before associating knowledge bases, check the vector model identifier and exclude those that do not match. Do not run cross-model retrieval to compare missed recalls.
- Judging an empty citation array as meaning there is no rebate obligation. The search node output length may be zero. An empty result needs to be assessed together with whether data training is complete and whether the retrieval scope covers the target document. A negative conclusion cannot be drawn directly.
How to verify the setup
- Input the sales achievement table wording and the draft sentence from the hypothetical exercise sample, and compare whether the draft sentence matches the original on the same entity, the same period and the same field. If the original states "the returned goods amount of RMB 150,000 has been deducted" and the draft sentence does not mention the returned goods deduction, the draft is unsupported and the returned goods deduction basis needs to be checked and supplemented.
- Input the vector model identifiers for the rebate policy base and the sales achievement table base, and compare whether they are identical. If the identifiers differ, exclude joint association, record the actually observed model identifiers, and do not proceed with cross-model retrieval.
- Input an empty citation array [] and check whether the downstream answer step is designed to indicate "no cited content retrieved". If the downstream step directly outputs a rebate conclusion, a human check step needs to be added to the workflow to define how empty results are handled.