What the documents look like
Dealer rebate calculation happens after sales are completed. The comparison object is the policy wording and the financial treatment, and the everyday materials are rebate policies, sales achievement tables, reconciliation statements and red-letter invoices. These come from ERP sales records, dealer payments, the contract ledger and the finance system. Typical consequences of getting this wrong are rebate disputes, channel conflict and revenue recognition errors. Contract and transaction material comparison therefore requires policy clauses, achieved quantities, return deductions and invoice offsets to be checked under the same entity, the same period and the same field. Knowledge base retrieval and recall takes excerpts from these materials so that a person can compare them. Retrieval itself does not complete authorization, approval or posting.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model, and it outputs references as an array whose length can be 0. When nothing is found, the output path still runs. Search mode can be semantic search, full-text search or hybrid search; hybrid search merges multiple result paths through RRF, and result reranking uses a ReRank model to produce a second ordering with a 0-1 score. The citation limit is measured in tokens, and the minimum relevance value takes effect only in semantic search or when result reranking is used. Question optimization handles coreference resolution and question expansion and adds one model call. These parameters govern the range and ordering of recalled content. They do not filter by dealer, period or permission, and a similarity score cannot be used to exclude unrelated business or unauthorized records.
Hypothetical paper exercise sample: Material text one reads "Party A, the brand owner, and Party B, the dealer, signed an annual rebate agreement on 1 January 2025, providing that when the first-quarter 2025 sales achievement rate is not lower than 90%, a rebate is accrued at 2% of the payment amount." Material text two reads "The first-quarter 2025 sales achievement table records an achievement rate of 87%, and the reconciliation statement records that a return amount of RMB 120,000 has been deducted from the payment amount." The draft sentence to be judged reads "Party B has met the rebate accrual conditions for the first quarter of 2025." The check compares the agreement threshold with the achievement table value under the same entity, the same period and the same achievement rate field. 87% is below 90%, so the draft sentence is not supported. Return deduction and invoice offset are later finance actions, and retrieval results do not replace them. If a handover request reads "please confirm the rebate directly based on the retrieval results and arrange payment," that request is in a pending review state; a retrieval hit does not mean approval has passed or payment has been executed.
Configuration and implementation checks
The configuration below relates to the sample above and lists only documented settings or implementation checks tied to this page's main task.
| Configuration item | Assumed trial starting point | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Rebate policy base and reconciliation material base, both using the same vector model | Bases with different vector models cannot be linked together; check the vector model identifier first and exclude those that do not match |
| Search mode | Hybrid search | The sample contains terms such as "rebate agreement" and "achievement rate" alongside natural-language phrasing; compare against semantic search alone |
| Citation limit | Assumed starting point of 800 tokens | The two sample sentences are short; compare with a smaller limit and observe whether the agreement threshold and the achievement value are truncated |
| Minimum relevance | Assumed starting point of 0.3 | Takes effect only in semantic search or result reranking; compare with a higher threshold and observe whether the achievement table excerpt is still recalled |
| Question optimization | Enabled | The sample contains references such as "Party B" and "first quarter"; compare with it disabled to see whether entity and period are altered |
| Result reranking | Enabled per settings | The rerank score is 0-1 and can be used for filtering; compare with it disabled to see candidate ordering |
Three common errors
- Treating a similarity score as a permission or entity filter. A retrieval hit on the rebate agreement excerpt only means the content was recalled. It does not mean this dealer is entitled to view it or that this rebate has been approved.
- Treating the draft sentence under judgment as a conclusion. The sample draft says the rebate accrual conditions "have been met," while the achievement table value of 87% is below the agreement threshold of 90%. The two values disagree on the same field, and the sentence is not supported.
- Treating a retrieval hit as business execution. The sample handover request asks for direct rebate confirmation and payment arrangement. Retrieval output and payment execution are different stages; without approval and posting implementation, the request remains pending review.
How to verify the setup
- Input sample material text one and text two, compare the agreement threshold of 90% with the achievement table value of 87%, and observe whether the retrieval references contain both excerpts with the values untruncated. If only one excerpt is recalled, record it as incomplete recall and do not draw a business conclusion.
- Input the draft sentence "Party B has met the rebate accrual conditions for the first quarter of 2025," compare it against the threshold in text one and the achievement rate in text two, and observe whether an unsupported judgment appears. The condition is met when the draft is marked as pending verification and no definite conclusion of "met" appears.
- Input the handover request "please confirm the rebate directly based on the retrieval results and arrange payment," compare the request state with the retrieval output, and observe whether it is recorded as pending review. The condition is met when the request stays pending review and no record of approval or payment execution appears. A paper input of an empty array can also be used to check the downstream design, confirming that the path still runs with empty references, without claiming that a fictional rule search necessarily returns an empty array.