What the documents look like
Rebate calculation happens after sales close. Its result depends on policy wording, volume attainment, and financial treatment, and these three kinds of fact sit in separate documents: the rebate policy, the volume attainment table, the reconciliation statement, and the red-letter invoice. When product and marketing copy cites these materials, each cited source must be a traceable original document, and every promotional claim, commitment, or figure in the copy must have a corresponding statement in the material. Data updates monthly or quarterly, so a citation must state the version and period of the material it relies on, and must not use a prior period's attainment result to describe the current period's policy.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. The search output is an array of references, and its length can be zero; an empty array still runs the output path and does not mean the workflow has reported an error and stopped. The reference content needs to be passed to the answering step, where the answer is checked against the cited original text item by item. Minimum relevance is a filter threshold, effective under semantic retrieval or result reranking, and it does not perform permission or entity isolation; it cannot be used to exclude unrelated business or unauthorized records. Hybrid retrieval merges multiple result paths through RRF, reranking requires a model and must be enabled by configuration, and image retrieval requires corresponding image input and model capability; these paths do not mean that every request executes all of them.
Hypothetical drill sample: a brand's March 2025 rebate policy original text states that "when a dealer's quarterly volume attainment rate is not lower than 90%, rebate is accrued at 1.5% of the quarterly payment amount; return amounts are deducted from the accrual base." The reconciliation statement for the same period states that "as of March 31, 2025, dealer A had quarterly payments of 1.2 million yuan, returns of 80,000 yuan, and an attainment rate of 92%." The draft sentence to be judged is "Dealer A's attainment rate this quarter exceeded ninety percent, and rebate is accrued on the full payment amount." The material does not state whether rebate accrual has been executed or whether an invoice offset has been issued, so the phrase "full accrual" lacks material support and remains pending verification.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this trial | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Select the bases holding the rebate policy and the volume attainment table | Check that vector model identifiers match, and exclude bases with different vector models |
| Search mode | Hybrid retrieval | Policy clauses contain numbers and proper nouns, so semantic and full-text paths each contribute |
| Minimum relevance | Hypothetical starting point 0.5 | Compare with the same sample when no threshold is set, and observe whether irrelevant fragments decrease |
| Reference limit | Hypothetical starting point 3,000 tokens | Compare with the total length of the material paragraphs, and confirm the policy original text can be cited in full |
| Question optimization | Enable for multi-turn follow-up questions | Compare with the disabled state, and check whether the retrieval query is complete after coreference resolution |
| Result reranking | Enable when candidates are numerous | Compare ranking positions with the merged result alone; a reranking model must be available |
These values are hypothetical starting points for this trial, not official default values. The comparison method is to change one setting at a time under the same sample and observe changes in the reference content.
Three common errors
- Associating knowledge bases that use different vector models. The check input is the vector model identifiers of two bases; compare whether the two identifiers match, exclude the base when they differ, and do not run cross-model retrieval to compare missed recalls.
- Treating minimum relevance as a permission or entity filter. The check input is one unauthorized record and one unrelated record; compare whether they are still recalled before and after filtering. The relevance threshold does not change authorization status, and the record may still appear in the reference array.
- Treating "full accrual" in the draft sentence as already supported by the material. The check input is the policy original text and the reconciliation statement above; compare whether "full accrual" has a corresponding field. The material states only the accrual rate and the deduction rule, and does not state the accrual execution status, so the wording remains pending verification.
How to verify the setup
- Use the policy original text and the reconciliation statement from the hypothetical sample as input. Compare whether "attainment rate exceeded ninety percent" in the draft sentence matches the 92% in the reconciliation statement on the same entity, the same period, and the same field. When the figure matches and the entity and period correspond, that clause is supported by the material.
- Use an unauthorized record as input. Compare whether that record appears in the reference array. If it appears, the relevance threshold has not performed an isolation function, and permission must be checked separately in the retrieval scope or in the manual input step.
- Use a paper input of an empty array to check the downstream design. Compare whether the answering step handles it as having no references. When the answering step does not stop because of the empty array and does not produce a conclusion without a source, the empty-result handling matches expectations. Do not claim that retrieval of a fictional rule necessarily returns an empty array.
This page addresses only dealer rebate calculation, product and marketing copy drafting, and citation tracing. Whether a specific promotional sentence in product copy is supported by the material must not be turned into an enterprise identity certification tutorial.