What the documents look like
Dealer rebate calculation happens after sales close, and its core is policy wording and financial treatment. The working materials come from ERP sales records, dealer remittances, contract ledgers and the finance system. Day-to-day documents include rebate policies, sales attainment tables, reconciliation statements and red-letter invoices. This page addresses product and marketing copy drafting, which requires retrieving source text from these materials to support a proposed promotional sentence. The applicable period, attainment conditions and target scope inside a rebate policy directly determine whether a given sentence can be supported by the material. Data updates monthly or quarterly, so retrieval must check whether the document version matches the period the copy refers to.
Constraints at the retrieval step
Rebate materials contain many numbers, policy names and proper nouns, so semantic retrieval and full-text retrieval each have suitable uses. A knowledge base search node can associate one or more knowledge bases that use the same vector model. The citation limit is measured in tokens, and the minimum relevance score takes effect only under semantic retrieval or when result reranking is used. Question optimization handles coreference resolution and question expansion, and the subject must be checked for unintended changes. A similarity score does not perform permission or entity isolation, and retrieval itself does not constitute authorization.
Hypothetical paper exercise sample: the source material reads, "2024 East China dealer rebate policy: dealers whose annual remittance attainment rate is not lower than 95% accrue a rebate at 1.5% of remittance amount; the rebate offsets next-quarter payment by red-letter invoice." The proposed copy sentence reads, "East China dealers who complete the annual remittance target receive a rebate." The check compares whether the proposed sentence retains the 95% threshold, the accrual base and the offset method. The material does not state the immediacy implied by "receive," so that wording is not supported.
Configuration and implementation checks
| Item | Relevant setting or check for this page | Hypothetical trial starting point and comparison method |
|---|---|---|
| Associated knowledge bases | Select libraries containing rebate policies, attainment tables and reconciliation statements | Verify that each library's vector model identifier matches; exclude those that do not |
| Search mode | Compare semantic, full-text and hybrid retrieval against test questions | Use a policy name for full-text, a natural-language phrasing for semantic |
| Citation limit | Unit is tokens | Start at a value covering one complete policy clause, then compare against truncation |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking | Label relevant and irrelevant samples, then adjust and observe filtering changes |
| Question optimization | Used for coreference resolution and question expansion | Compare whether the subject in the original and optimized question stays consistent |
| Result reranking | Requires a model and must be enabled per settings | Compare candidate order before and after reranking, and observe the 0-1 score distribution |
Three common errors
- Treating the proposed sentence "dealers who complete the annual remittance target receive a rebate" as supported by the material. The material states a 95% threshold and red-letter invoice offset; the proposed sentence omits the threshold and adds immediacy, so it is not supported.
- Treating a similarity score as a permission filter. Retrieval returning a policy clause does not mean the current user is authorized to confirm a rebate on that basis; authorization must be determined by a downstream business system.
- Describing the RRF fusion in hybrid retrieval as a switch that can be toggled. RRF is the process of merging multiple result paths; the documentation provides no independent switch for it.
How to verify the setup
- Input the proposed sentence and the source material, then compare whether the proposed sentence retains the 95% threshold, the 1.5% accrual base and the red-letter invoice offset method. Expected observation: the check can point out that "receive" lacks material support.
- Input a rebate policy name for full-text retrieval and a natural-language phrasing for semantic retrieval, then compare the retrieved content. Expected observation: the identifier-style query hits the policy clause, and the natural-language query hits a semantically similar passage.
- Input a query unrelated to rebates and observe the returned result. Expected observation: when the cited content is an empty array, the downstream design handles the empty result; if content is returned, record the actual observation without concluding that permission or entity filtering has failed.