What the documents look like
Dealer rebate calculation happens after the sale. The work centres on policy wording and financial treatment, and the documents involved are rebate policies, sales achievement tables, reconciliation statements and red-letter invoices. The main data comes from ERP sales records, dealer payments, contract ledgers and the finance system, with monthly and quarterly update cycles.
For retrieval, this means the material carries a definite period and a definite owning party. A rebate policy states which dealers it covers and which date range it applies to, so an answer that cannot be tied to a period and a dealer cannot support reconciliation. Product information and selection questions in this setting are queries against policy clauses, achievement-caliber notes and reconciliation format requirements.
Constraints at the retrieval step
The monthly and quarterly rhythm means several period versions of the same policy can sit in one knowledge base. When a search returns clauses from more than one period, the cited content mixes periods. The knowledge base search node can associate one or more knowledge bases that use the same vector model, and that association sets the library boundary for retrieval.
Search mode is compared across semantic, full-text and hybrid retrieval for the test question; no single mode is assumed to suit all material. The citation limit is measured in tokens and controls how much content one search returns, so material of different lengths needs separate checks for completeness. Minimum relevance takes effect only under semantic retrieval or when result re-ranking is used, and it is a filter threshold. Question optimisation handles coreference removal and question expansion, and it adds one model call before the actual retrieval.
Hypothetical paper exercise: a brand finance staff member needs to check the scope of a dealer rebate policy. The knowledge base holds a document titled "Q3 2024 rebate policy", which states: "This policy applies to the calculation of purchase rebates for contracted dealers in the East China region during the period 1 July 2024 to 30 September 2024." A second document is titled "Q4 2024 rebate policy" and states: "This policy applies to the calculation of purchase rebates for contracted dealers nationwide during the period 1 October 2024 to 31 December 2024." The draft sentence to be judged is: "A purchase by an East China dealer in August 2024 falls under the Q4 rebate policy." The check compares the party, the period and the stated scope of application in the draft sentence against the policy documents.
Configuration and implementation checks
The items below relate to this exercise and to the named task. The values are starting points for this hypothetical trial, used to compare retrieval behaviour at different settings; they are not official defaults.
| Configuration item | Hypothetical starting value for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only the library holding rebate policy documents | That library uses one vector model; compare whether cited content covers the policy text when this library is associated and when it is not |
| Search mode | Hybrid retrieval | Policy titles contain quarter numbers, so full-text retrieval can hit the number while semantic retrieval can match the phrasing; compare the candidates each mode returns |
| Citation limit | 800 tokens | Policy clauses typically run from tens to hundreds of characters; compare 400 and 800 for whether cited content truncates a clause |
| Minimum relevance | 0.35 | Label relevant and irrelevant samples first, then adjust the threshold; compare 0.2 and 0.5 for whether irrelevant fragments are filtered out |
| Question optimisation | Off | The current question already names a definite party and period, with no complex coreference; compare whether the party is altered when this is on and off |
| Result re-ranking | Not enabled | The candidate set is limited; compare whether ordering changes when this is on and off |
Three common errors
- Treating a retrieval result as an authorisation basis. Cited content only shows that a related document exists in the knowledge base. It does not show that the document is in force, approved, or applicable to the current dealer. Checking authorisation status requires the contract ledger or approval records, and retrieval by itself does not complete that confirmation.
- Associating libraries that use different vector models. The knowledge base search node requires the associated knowledge bases to use the same vector model. Where libraries with different vector models would be associated, check each library's vector model identifier first, exclude those that do not match, and record what is actually observed. Do not infer a run result.
- Reading minimum relevance as an entity or permission filter. Minimum relevance is a similarity filter threshold. It does not separate dealer entities and does not judge record ownership. Retrieved content may include material belonging to other dealers, so the party has to be checked manually.
How to verify the setup
- Enter the draft sentence from the hypothetical exercise, "A purchase by an East China dealer in August 2024 falls under the Q4 rebate policy", and compare the cited content returned against the scope of application in the two policy documents. If the cited content includes the Q3 policy and does not include the Q4 policy, the draft sentence does not agree with the scope stated in the Q3 document. If the cited content includes both policies, the period attribution needs manual judgement and cannot be decided from the retrieval result alone. The expected observation is that cited content can be located to a specific policy document; a first-position hit is not guaranteed.
- Enter the policy text from the exercise, "This policy applies to the calculation of purchase rebates for contracted dealers in the East China region during the period 1 July 2024 to 30 September 2024", and compare whether the cited content returned contains the party, period and scope of application in full. If the cited content is truncated or returns only part of the clause, adjust the citation limit and compare again. The expected observation is that cited content covers the full clause; full recall is not guaranteed.
- Use the explicit paper input
[]to check the downstream design. When the cited content length is 0, the output path still runs, and the downstream node has to handle an empty array. Compare whether the downstream node continues as designed when the result is empty. Do not claim that a search over a hypothetical rule necessarily returns an empty array. The expected observation is that the downstream node receives empty cited content and continues under its stated logic.
The public technical facts checked for this feature are: the knowledge base search node can associate one or more knowledge bases that use the same vector model; search mode is compared across semantic, full-text and hybrid retrieval for the test question, with no single mode assumed to suit all material; the citation limit is measured in tokens, and material of different lengths is used to test content completeness before the limit is adjusted; minimum relevance takes effect only under semantic retrieval or when result re-ranking is used, and relevant and irrelevant samples are labelled before the threshold is adjusted; question optimisation handles coreference removal and question expansion, and the original question is compared with the optimised question to check whether the party has been altered.