What the documents look like
Dealer rebate calculation happens after the sale. Its center is the policy basis and the financial treatment, which separates it from inventory replenishment work. The material set for one rebate cycle normally includes the rebate policy, the volume achievement table, the reconciliation statement, and the red-letter invoice used for offsetting. These materials draw on ERP sales records, dealer remittances, the contract ledger, and the finance system. Update rhythm is monthly or quarterly, and that rhythm does not by itself change how retrieval is configured on this page.
In a cross-border setting, the same batch of material can exist as a Chinese policy original and a foreign-language reconciliation statement. Retrieval has to locate the corresponding clause across languages. The rebate policy states the achievement condition, the calculation base, and the applicable period. The volume achievement table records actual dealer volume for a period. The reconciliation statement is checked between brand finance and dealer finance. The red-letter invoice handles the invoicing offset. For a retrieved result to be worth checking, the subject, the object, the period, and the status across these four material types must agree. A retrieval hit is not a verification, an approval, or an authorization to act.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The practical step is to check the vector model identifier first, exclude the non-matching bases, and then search within the matching set. Semantic retrieval carries cross-language understanding, so a Chinese question can match a foreign-language knowledge point. That ability depends on how well the vector model was trained, and its precision is not stable.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Full-text retrieval suits exact lookups such as numbers, model codes, and proper nouns. Hybrid retrieval uses vector search and full-text search together and merges the two result sets through RRF. RRF is a fusion process, not a separate switch that can be turned on. The minimum relevance value is a filter threshold, and it takes effect only under semantic retrieval or when result reranking is used. It does not perform permission or entity isolation. The reference output is an array whose length can be zero, which means the output path still completes when nothing is retrieved.
A hypothetical paper exercise sample: a brand finance staff member holds a rebate policy original reading, "A dealer that completes the volume target for the first quarter of 2025 accrues a rebate at three percent of the quarterly remittance amount; a dealer that does not complete the target accrues at one percent." A foreign-language reconciliation statement for the same dealer and the same quarter states the remittance amount. The draft sentence to be judged is, "The dealer's first-quarter 2025 rebate is accrued at three percent." Judging this requires evidence for the same subject, the same period, and the same field. The volume achievement status is not stated in the sample, so the draft sentence remains unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the rebate policy base and the reconciliation material base that use the same vector model | Check the vector model identifier first, exclude bases on other models, then run the search |
| Search mode | Test semantic retrieval and full-text retrieval separately on the same question | Compare how each mode recalls policy clauses and foreign-language reconciliation content |
| Reference limit | Use a hypothetical starting value to cap the content length of a single reference | Test with policy passages of different lengths to check content completeness, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set the filter value | The value takes effect only under semantic retrieval or result reranking; compare against the labeled set |
| Question optimization | Enable for multi-turn follow-ups that contain references; disable for complete questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Use when a rerank model is available and enabled by configuration | The rerank score runs from 0 to 1 and is used for filtering, not for permission decisions |
Three common errors
- Treating the minimum relevance value as a permission or entity isolation control. The value is a filter threshold. It does not separate dealer subjects and does not exclude unauthorized records. A retrieval hit does not mean access is authorized; authorization requires a downstream implementation.
- Forcing knowledge bases on different vector models into one association. The documents state that knowledge bases using the same vector model can be associated. Bases on different models must first be checked by identifier and excluded. No claim should be made about missed recall, redundant returns, or unpredictable behavior from such an association, because those outcomes are not documented.
- Treating the rebate rate in the draft sentence as a verified conclusion. The sample original states only the policy accrual rule. The volume achievement status is not stated, and evidence for the same period and same field is missing, so the rate remains unsupported.
How to verify the setup
- Input the rebate policy original and the foreign-language reconciliation statement above, and search for "first-quarter rebate accrual rate." Compare whether the recalled content covers both the policy clause and the reconciliation data, and observe whether any subject or period mismatch appears. The expectation is that the policy original is located while the volume achievement field still shows as missing.
- Input the draft sentence "The dealer's first-quarter 2025 rebate is accrued at three percent" and compare it field by field against the policy original. The subject and period agree. The accrual rate field lacks volume achievement evidence in the sample. The rate can gain support only after material stating the achievement status is added; the current record stays pending verification.
- Input a clearly empty-result scenario and check whether the downstream handles it as an empty array. The documents state that the reference content length can be zero and that the empty-result path completes. Observe whether the downstream misreads the empty result as a retrieval failure or as an authorization pass. Both outcomes should be recorded as unconfirmed.