What the documents look like
Dealer rebate calculation happens after sales, and its center is policy interpretation and financial treatment. The working materials are rebate policies, sales attainment tables, reconciliation statements and red-letter invoices, drawn from ERP sales, dealer remittances, contract ledgers and the finance system. Rebate disputes, channel conflict and revenue recognition errors are the recorded consequences of getting this wrong. When product data and selection questions cite these materials, each answer needs to land on a specific document, a specific period and a specific field so that later reconciliation has something to check against.
Two features of this category shape the citation tracing step. First, the rebate policy text carries the calculation rule, while the sales attainment table carries the actual quantities; an answer that cites one without the other is incomplete. Second, a draft sentence can look numerically correct while omitting a deduction that the policy requires, so tracing has to compare field against field rather than judge whether the sentence reads fluently.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model, and it outputs citations as an array whose length may be zero. A zero-length array still lets the output path run through. Whether the retrieval result is passed to the answering step, and whether the answer maps item by item to the cited original text, requires separate checking.
The minimum relevance value is a filter threshold on a 0-1 scale and applies only in semantic retrieval or when result reranking is used. It does not perform permission or entity isolation, and retrieval by itself does not constitute authorization. Hybrid retrieval combines multiple result paths through RRF, which is a fusion process rather than a switch. Reranking requires a rerank model and is enabled according to settings. Image retrieval requires corresponding image input and model capability. The optional paths in the overall retrieval flow do not all execute on every request.
Hypothetical paper exercise: a brand finance reviewer checks dealer A's quarterly rebate in July 2025. The rebate policy text states: "When the quarterly sales attainment rate is not lower than 90%, the rebate is accrued at 2% of the purchase amount; the return amount is deducted from the numerator of the attainment rate." The sales attainment table records dealer A's second-quarter purchase amount of 1,200,000 and returns of 80,000. A draft sentence reads: "Dealer A's second-quarter attainment rate exceeds 90%, and the rebate is accrued at 2%." The question to decide is whether this draft sentence agrees with the policy wording and the attainment table fields.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Rebate policy base, sales attainment base | First check whether the two bases share the same vector model identifier; exclude any that differ |
| Search mode | Hybrid retrieval | Policy clauses contain codes and proper nouns, used together with semantic recall |
| Minimum relevance | 0.5 | Hypothetical starting point for this exercise; compare against the number of recalled items when no threshold is set |
| Citation limit | 2000 tokens | Hypothetical starting point; compare citation completeness against 1000 tokens |
| Result reranking | Enabled | Second ordering when candidates are numerous; compare ordering against the disabled state |
| Citation output | Array, zero allowed | When the array is empty, check whether the downstream path still runs through |
Three common errors
- Treating minimum relevance as a permission or entity filter. Relevance is a numeric threshold and does not exclude unrelated businesses or unauthorized records. The check is to compare the selected knowledge base scope and the document version instead.
- Describing RRF as a switch. RRF is the fusion process for hybrid retrieval results and has no independent toggle. The check is to compare the search mode and the rerank setting.
- Substituting thinking output or answer fluency for source evidence. The check is to compare the answer against the cited original text item by item and confirm that subject, period and field agree.
How to verify the setup
- Enter the policy text and attainment table fields from the hypothetical exercise above, and compare the attainment rate in the draft sentence against the policy wording. If the draft sentence does not deduct the return amount, the observable result is that it disagrees with the policy original. If the draft sentence deducts the return amount and states the resulting rate, the observable result is agreement on that field.
- Enter a paper citation array of [] and compare whether the downstream answering step still runs through. If the downstream step continues processing on the empty citation and records the gap, the condition is met. If the downstream step halts, the condition is not met.
- Enter two knowledge bases whose vector model identifiers differ and compare whether the mismatched base is excluded. If only the same-model base remains and the excluded item is recorded, the condition is met. Cross-model retrieval is not performed to compare missed recall.
Missing evidence stays missing. A pending review does not show that an operation permission is absent. Signature, cooperation reached and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period. Where the policy or the attainment table does not state a field, the result is unsupported or pending verification rather than a contradiction.