What the documents look like
Dealer rebate calculation happens after sales close, so the centre of the work is policy wording and financial treatment rather than stock movement. In the procurement and supplier material task, the documents serve as source evidence for a rebate amount: purchase-in records, supplier contract ledgers, rebate policy text, reconciliation statements, and red-letter invoices can each be cited as the basis for a figure. These materials update monthly or quarterly, and any citation has to be checked against the version period so that a purchase quantity or policy clause from the wrong period is not used. Brand finance, channel sales, and dealer finance all work from the same rebate, and the source material each side cites for one rebate needs to agree.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model, for vector search. Cited content is output as an array, and its length may be zero, which means the output path still runs when nothing is retrieved. The retrieval output needs to be passed to the answering step, where each answer statement is checked against the cited original text. Minimum relevance takes effect only under semantic retrieval or when result reranking is used; it is a filtering threshold and does not carry permission or entity isolation. The same-vector-model requirement for associated knowledge bases constrains knowledge base association only, and does not extend to business-registration sites or other external systems needing the same model.
Hypothetical exercise sample: a brand finance reviewer checking a dealer rebate for the third quarter of 2025 holds three materials. Material one is an excerpt from a supplier contract ledger, stating "Party A is a brand, Party B is a dealer, the contract period runs from 1 January 2025 to 31 December 2025, and the rebate is settled quarterly." Material two is rebate policy text, stating "The third-quarter rebate rate is 2% of the purchase amount, and the return amount must be deducted from the purchase amount." Material three is a reconciliation statement, stating "Third-quarter purchase amount 1,000,000 yuan, return amount 50,000 yuan, status pending dealer confirmation." The draft sentence to be judged is: "The dealer's third-quarter rebate amount is 19,000 yuan, confirmed." The word "confirmed" conflicts with the statement status "pending dealer confirmation," and the sentence needs to be marked as pending verification.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | One rebate policy base and one contract ledger base | Both bases use the same vector model; check the model identifier before associating |
| Search mode | Hybrid retrieval | Policy clauses contain codes and rates; hybrid retrieval covers both semantics and keywords |
| Minimum relevance | Hypothetical starting point 0.5 | Compare with actual retrieval scores and observe whether low-scoring fragments are filtered |
| Citation limit | Hypothetical starting point 2,000 tokens | Compare with the length of a single citation to avoid exceeding what the answering step can carry |
| Result reranking | Enabled | Second-pass ordering when candidates are numerous; rerank scores are used for filtering |
| Question optimisation | On | Completes references in multi-turn dialogue; check the completed result against the original question |
Three common errors
- Writing a pending status as confirmed. The reconciliation statement says "status pending dealer confirmation," and the draft sentence says "confirmed." The two values are mutually exclusive on the same field, which is a contradiction. Check: input the reconciliation statement text and the draft sentence, compare the status field values, and mark a contradiction when "pending confirmation" and "confirmed" appear together.
- Using the wrong period's purchase amount. The contract ledger states the contract period covers all of 2025, and the rebate policy states a third-quarter rate. Citing the full-year purchase amount to compute a third-quarter rebate does not match the periods. Check: input the contract period and the policy period, compare whether both cover the same quarter, and mark pending verification when the periods differ.
- Treating retrieval results as an authorisation basis. The retrieval output citation array may have length zero; an empty array does not mean the workflow terminated with an error, and it does not mean the rebate was approved. Check: input the empty array and the downstream handoff request, compare whether the downstream reads the empty array as "no rebate," and record an unsupported conclusion when it does.
How to verify the setup
- Citation source check. Input the material two text from the hypothetical sample, "The third-quarter rebate rate is 2% of the purchase amount, and the return amount must be deducted from the purchase amount," and compare the calculation process behind the 19,000 yuan figure in the draft sentence, confirming that the 50,000 yuan return amount was deducted from the 1,000,000 yuan. Expected observation: the calculation process matches the deduction rule in the original text; mark a calculation error when the deduction is absent.
- Status field check. Input the reconciliation statement text "status pending dealer confirmation" and the draft sentence "confirmed," and compare the values on the same status field. Expected observation: when the two values differ, the draft sentence must not be output as a confirmed fact and must retain a pending-verification mark.
- Empty citation handoff check. Input the paper array
[]and the design description of the downstream answering step, and compare whether the downstream treats the empty array as "no cited content retrieved." Expected observation: the downstream retains the empty-citation mark and continues execution; when the downstream turns the empty array into a negative conclusion, record a handoff design gap.
Retrieval checks describe observation and failure branches only: when the first result does not hit the target clause, record the actual returned content and score, without guaranteeing a first-hit or full recall. Knowledge bases with different vector models cannot be associated together; check the vector model identifier first and exclude those that do not match, rather than running a cross-model retrieval to compare missed recall. The optional paths shown in the overall document flow diagram do not mean every request executes all of them: image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled in settings. Thinking budget belongs to model thinking configuration and is not used as the citation token limit.