What the documents look like
Dealer rebate calculation happens after sales close. The working objects are policy definitions and financial treatment, and the data comes from ERP sales, dealer remittances, the contract ledger and the finance system. The daily documents are rebate policies, sales achievement tables, reconciliation statements and red-letter invoices, with monthly and quarterly update rhythms. This page is limited to the inventory procurement and supplier records task: retrieval must pull the original passages from rebate policies, sales achievement tables, reconciliation statements and red-letter invoices that carry the achievement basis, returned-goods amount and invoice offset. These documents are field-dense. One reconciliation statement can hold the counterparty, the period, the quantity and the amount in the same block, so a retrieved fragment has to preserve the field correspondence or the later reconciliation step cannot proceed.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; the vector model identifier has to be checked first and non-matching bases excluded. Search mode is compared across semantic retrieval, full-text retrieval and hybrid retrieval against the test question, and no single mode is presumed to suit all materials. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization handles coreference resolution and question expansion. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a switch. Result reranking requires a rerank model and must be enabled by configuration, and image retrieval requires the corresponding image input and model capability; these paths do not all execute on every request. Retrieval and recall carry no authorization decision and no entity isolation, and a similarity score is not a permission filter.
Hypothetical paper exercise sample: a brand finance staff member checks one reconciliation statement. The material text reads "Counterparty: one dealer in East China; Period: July 2024; Returned-goods amount: CNY 12,000; Rebate policy number: RB-2024-07." A draft sentence awaits judgment: "This dealer's returned-goods amount for July 2024 is zero, and the rebate is accrued in full." The retrieval input is "East China dealer July 2024 returned-goods amount." The comparison object is the counterparty, period and returned-goods amount fields in the recalled fragment. The source text states a returned-goods amount of CNY 12,000, the draft states zero, and the same field carries mutually exclusive values, which is a contradiction. If the recalled fragment shows only the period and not the returned-goods amount, the draft sentence is unsupported and the evidence gap is recorded.
Configuration and implementation checks
| Configuration item | Value or check on this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Select the bases holding rebate policies, sales achievement tables, reconciliation statements and red-letter invoices | The node can associate one or more knowledge bases using the same vector model |
| Search mode | Compare semantic, full-text and hybrid retrieval using "East China dealer July 2024 returned-goods amount" as the test question | No single mode is presumed to suit all materials |
| Citation limit | Hypothetical trial starting point of 2000 tokens, compared with a 500-token result for fragment completeness | This is a hypothetical starting point, not an official default; the documented unit is tokens |
| Minimum relevance | Label relevant and irrelevant samples first, then set the threshold from them | Takes effect only under semantic retrieval or when result reranking is used |
| Question optimization | Compare the original question with the optimized question and check whether the counterparty was altered | Used for coreference resolution and question expansion |
| Result reranking | Confirm whether a rerank model is available and enabled by configuration | Reranking requires the model and must be enabled by configuration |
Three common errors
- Treating the similarity score as a permission or entity filter. The input is "East China dealer July 2024 returned-goods amount." If the assumption is that raising the minimum relevance excludes other dealers' records, the comparison object is the score against record ownership, and the observed result is that a score change does not alter a record's business ownership; authorization still needs a downstream implementation.
- Treating RRF as a switch that can be turned on. The input is a hybrid retrieval request. The comparison object is the configuration item list against the documented description, and the observed result is that the documentation describes RRF only as merging multiple result paths and provides no separate switch.
- Treating a pending-review status as granted authorization. The input is a handover request under review. The comparison object is the request status against the authorization record, and the observed result is that pending review neither shows an absence of operating permission nor shows approval; signing, reaching cooperation and contract effectiveness are three distinct facts.
How to verify the setup
- Compare the hypothetical sample text "Counterparty: one dealer in East China; Period: July 2024; Returned-goods amount: CNY 12,000; Rebate policy number: RB-2024-07" with the draft sentence "This dealer's returned-goods amount for July 2024 is zero, and the rebate is accrued in full." The input is the retrieval question "East China dealer July 2024 returned-goods amount." Compare the counterparty, period and returned-goods amount fields in the recalled fragment. The observed result is mutually exclusive values of CNY 12,000 and zero in the same field, judged as a contradiction.
- Use a paper input of an empty array to check the downstream design. The input is an explicitly empty citation content. The comparison object is how the downstream node handles citation content of length zero. The observed result is that the output path runs through; no claim is made that retrieval of a fictional rule necessarily returns an empty array.
- Use one handover request under review to check the authorization boundary. The input is the request text and its status field. The comparison object is the request status against the authorization record. The observed result is that, where no authorization is stated, the item is recorded as pending verification, and it is not judged as approved or rejected.