What the documents look like
Rebate calculation happens after sales, and its centre is policy wording and financial treatment. Brand finance, channel sales and dealer finance work from four document families: rebate policies, sales achievement tables, reconciliation statements and red-letter invoices. The recurring work is reconciling achievement definitions, deducting returned-goods amounts and offsetting invoicing. In store service and internal training, front-line staff need to retrieve policy definitions, reconciliation explanations and invoicing offset rules from the knowledge base, and the retrieved content directly shapes the consistency of outward answers. Errors in this area produce rebate disputes, channel conflict and incorrect revenue recognition, so retrieval is expected to keep subject, period and field aligned.
The documents carry policy clauses and reconciliation notes that are often long within a single record. That length matters for the retrieval step because the reference limit is measured in tokens rather than in a fixed number of items. The update frequency given for this category is monthly or quarterly; it does not change how retrieval is configured, so it is not developed here.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; the check is to compare model identifiers and exclude a base whose identifier does not match, then record what was observed. No claim is made about how a cross-model association would behave, because that association is not performed.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. They are compared per test question rather than assumed to suit all material. The reference limit is expressed in tokens. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. Question optimisation performs coreference resolution and question expansion; the original question and the optimised question are compared to check whether the subject has been altered. Hybrid retrieval merges multiple result paths through RRF. Re-ranking requires a model and must be enabled in the settings. Image retrieval requires image input and the corresponding model capability. The optional paths shown in the overall retrieval diagram do not mean every path runs on every request.
Retrieval and similarity scoring do not carry permission or entity-isolation duties. Retrieval is not authorisation. The node outputs cited content as an array, and its length can be zero.
Hypothetical paper exercise sample: the source text is "Brand X 2025 second-quarter dealer rebate policy, clause 4: the quarterly achievement rate is calculated on net sales after deduction of returns, and a return is dated by the date the dealer signs for and receives the returned goods into stock." The draft sentence to be assessed is "This quarter's rebate is calculated on the dealer's declared sales, and returns are offset in the original sales month." The draft differs from the source on two fields, the calculation basis and the period to which returns belong. It is recorded as pending verification, and the evidence gap is preserved. No authorisation or effective clause is invented.
Configuration and implementation checks
| Configuration item | Value or check relevant to this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Rebate policy base and reconciliation and invoicing notes base; compare their vector model identifiers for consistency | The node can associate multiple knowledge bases that use the same vector model |
| Search mode | Compare semantic, full-text and hybrid results separately for each test question | No single mode is assumed to suit all material |
| Reference limit | For this exercise, the starting value is set to 2000 tokens, because policy clauses and reconciliation notes are long per record; compare against a 1000-token result for completeness of content | Unit is tokens; the starting value is hypothetical, not an official default |
| Minimum relevance | Takes effect only under semantic retrieval or result re-ranking; label relevant and irrelevant samples before adjusting | This value is a filter threshold, not a ranking rule |
| Question optimisation | Compare the original question and the optimised question in multi-turn follow-ups, checking whether the rebate subject and period were altered | Used for coreference resolution and question expansion |
| Result re-ranking | Where a re-ranking model exists and is enabled in the settings, filter using the re-ranking score | Re-ranking requires a model and must be enabled in the settings |
Three common errors
- Associating knowledge bases that use different vector models. The input is the vector model identifier of each base; the comparison is whether the two identifiers match; the observable result is that a base with a non-matching identifier is excluded, and no cross-model retrieval is run to compare missed recalls.
- Treating minimum relevance as a permission or entity filter. The input is a query about unrelated business; the comparison is whether the returned content falls within the target rebate policy scope; the observable result is that a similarity score does not exclude unauthorised records, and permission requires a separate system implementation.
- Adopting an optimised question whose subject was altered. The input is the original question containing a reference and the optimised question; the comparison covers the rebate subject, period and field; the observable result is that a replaced subject is recorded as pending verification and does not enter the answer.
How to verify the setup
- Enter the hypothetical sample source text and the draft sentence, compare the calculation basis and the period to which returns belong, and observe that when the two values differ the item is judged pending verification rather than contradiction or denial.
- Enter a query of the exact-number type, compare the content returned by full-text retrieval and by semantic retrieval, and record the number match as actually observed, without asserting a first-position hit or complete recall.
- Enter an empty array as cited content, check how the downstream node is designed to handle a zero-length citation path, and confirm the downstream behaviour against the actual configuration rather than claiming that a fictional rule search always returns an empty array.