What the documents look like
Dealer rebate calculation happens after sales, and its core work is reconciling policy terms with financial treatment. The documents involved are rebate policies, sales achievement tables, reconciliation statements, and red-letter invoices. Their sources include ERP sales records, dealer remittances, contract ledgers, and financial systems. These materials circulate among brand finance, channel sales, and dealer finance in different languages and different versions. A rebate calculation error can produce rebate disputes, channel conflict, and incorrect revenue recognition, so any citation must trace back to a specific policy version and a specific reconciliation period. The materials update monthly and quarterly, and retrieval must separate the current version from historical versions.
For citation tracing, this means the knowledge base search node can associate one or more knowledge bases that share the same vector model. Its output is an array of citations, and the length can be zero, which means the output path still runs even when nothing is retrieved. Retrieval finds candidate source text. It does not perform permission checks and does not perform business verification. Semantic similarity and the minimum relevance score are filtering thresholds, and they do not act as entity isolation or authorization filters. Hybrid search merges multiple result paths through RRF. Reranking requires a rerank model and must be enabled by configuration. Image retrieval requires image input and corresponding model capability. These optional paths do not mean every request executes all of them. Knowledge bases with different vector models cannot be associated together; their identifiers must be checked first and non-matching ones excluded.
Hypothetical paper exercise sample: A brand finance team receives a dealer rebate reconciliation request. The source text available for checking is: "Version 3 of the 2024 annual rebate policy states that dealers with a quarterly achievement rate of no less than 90% receive a rebate of 2% of remittance amount. The 2024 second-quarter sales achievement table shows this dealer's achievement rate is 88%. The July 2024 reconciliation statement records that both parties have not yet confirmed the rebate amount." One draft sentence to be judged is: "This dealer's 2024 second-quarter rebate has been confirmed at 2% and a red-letter invoice can be issued to offset it." The retrieval scope is limited to the knowledge base containing the policy, achievement table, and reconciliation statement source text. The authorization boundary is internal brand finance checking, and it does not include outward invoice execution.
Constraints at the citation tracing step
The search node's citation output carries the source text used for an answer, and that output can be an empty array. An empty array is a valid output path, not a workflow failure. This matters because a missing citation must remain visibly missing rather than being filled by an invented source.
Retrieval and authorization are separate. A similarity score ranks or filters candidate text by relevance. It does not determine whether a user may see a record, whether a dealer is entitled to a rebate, or whether a rebate has been approved. The knowledge base association rule about the same vector model is a retrieval configuration constraint. It does not extend to business registries or other external systems.
The optional retrieval paths have their own preconditions. Image retrieval needs image input and a model that supports it. Reranking needs a rerank model and must be enabled. RRF is a fusion process for combining multiple retrieval paths, not a toggle. None of these paths should be assumed to run on every request.
The thinking budget settings in the AI configuration are reasoning budgets. They are not citation token limits. The citation limit is a separate search parameter measured in tokens.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only the knowledge base containing the rebate policy, achievement table, and reconciliation statement source text | Check each knowledge base's vector model identifier first and exclude those with different models |
| Search mode | Hybrid search | Policy clauses contain numbers and percentages, so keyword and semantic retrieval are both useful |
| Minimum relevance score | Hypothetical starting point 0.3 | Compare against leaving it unset to observe whether low-relevance fragments are filtered |
| Citation limit | Hypothetical starting point 3000 tokens | Compare against a smaller value to check whether policy source text is truncated |
| Reranking | Enable when a rerank model is available | Compare candidate ordering against the disabled state without assuming improvement |
| Citation output handling | Check whether the answer carries the cited source text | Compare against an answer with no citation to confirm traceability to the source sentence |
These values are hypothetical starting points for this exercise, not official defaults.
Three common errors
- Treating a retrieval hit as a confirmed rebate. In the sample, the reconciliation statement records that the rebate amount is not yet confirmed, while the draft sentence says it "has been confirmed at 2%." The input is the reconciliation statement source text and the draft sentence. The comparison is the confirmation status stated in each. The observable result is that the draft sentence's status does not match the source text.
- Treating an 88% achievement rate and a 90% policy threshold as the same measure. The input is version 3 of the policy and the achievement table. The comparison is the same achievement rate field for the same quarter. The observable result is that 88% is below 90%, and the draft sentence's 2% rebate lacks support.
- Forcing knowledge bases with different vector models into one association. The input is two knowledge bases with different vector model identifiers. The comparison is their model identifiers. The observable result is that the identifiers differ, so one must be excluded, and cross-model retrieval is not executed to compare missed recall.
How to verify the setup
- Input: the sample policy source text, achievement table source text, and draft sentence. Comparison: whether the confirmation status claimed in the draft sentence matches the reconciliation statement on the same subject, same period, and same field. Observation: the reconciliation statement records no confirmation while the draft sentence says confirmed. This is judged unsupported, and the evidence gap is recorded.
- Input: version 3 of the sample policy and the 2024 second-quarter achievement table. Comparison: the 88% achievement rate for the same quarter against the 90% policy threshold. Observation: 88% is below 90%, so the rebate accrual basis does not hold, and the case must be returned for a valid version or authorization basis.
- Input: a paper output of
[]and the downstream answer node design. Comparison: whether the empty citation array still runs through the output path. Observation: the downstream does not terminate with an error and does not fabricate a source, which satisfies the handling design. This does not claim that a retrieval for a fictional rule will always return an empty array.
This page addresses only dealer rebate calculation, cross-border materials and multilingual retrieval, and citation tracing. During operational handover, the authorization state of materials, operations, and requests must be kept separate from retrieval or form submission. Retrieval and forms are not authorization execution.