What the documents look like
Rebate calculation happens after sales, and the object of reconciliation is the policy basis together with the financial treatment result. Source material comes from ERP sales, dealer remittances, contract ledgers and the finance system. The day-to-day documents in this category are rebate policies, volume achievement tables, reconciliation statements and red-letter invoices. In the operating review and exception review stage, the question is what state a given achievement, offset or invoice issuance occupies in the documents, and whether retrieval can bring back the corresponding original text. Rebate disputes, channel conflict and revenue recognition errors commonly start from inconsistent citation of the policy basis, so retrieval accuracy directly determines whether a review conclusion is usable. Material updates on a monthly and quarterly cadence do not change the retrieval configuration itself.
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 identifier is checked first, non-matching bases are excluded, and the actual observation is recorded. A cross-model retrieval is not run to compare missed recalls. Retrieval results output cited content as an array, and the length can be zero, which means the output path still completes even when nothing is found. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, hybrid retrieval merges the two paths through RRF, and ReRank performs a second ordering of candidate text. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used; it is a filter threshold and does not perform permission or entity isolation. The citation limit is measured in tokens and controls how much content a single search may cite. Question optimization performs coreference resolution and question expansion and adds one model call.
Hypothetical paper exercise sample: a brand finance reviewer is reviewing March 2026 rebates, with the retrieval scope limited to the rebate policy base and the reconciliation materials base. The source text reads: "For March 2026, the East China dealer quarterly rebate is calculated in tiers by achievement rate; the achievement rate is based on net sales confirmed by finance, and return amounts are deducted from the achievement basis." and "As of 31 March 2026, this dealer's reconciliation statement status is pending dealer confirmation, and the red-letter invoice has not been issued." The draft sentence to be assessed is: "This dealer's March 2026 rebate has completed calculation and can be offset by invoice issuance." The subject, object, period and status are all stated in the sample, and the sample is used to check whether the retrieval result and the draft sentence agree.
Configuration and implementation checks
| Configuration item | Value or check method on this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Rebate policy base and reconciliation materials base; check that the vector model identifiers match before associating | The node permits associating multiple knowledge bases that use the same vector model |
| Search mode | Compare semantic retrieval and hybrid retrieval for policy-basis questions; compare full-text retrieval for identifiers and invoice numbers | Different modes suit different material forms |
| Citation limit | Set a hypothetical trial starting point of 2000 tokens and compare with 500 tokens for content completeness | This value is a hypothetical starting point for this exercise, not an official default |
| Minimum relevance | Label relevant and irrelevant samples first, then compare recall changes after adjusting the threshold | Takes effect only under semantic retrieval or result re-ranking |
| Question optimization | Compare the original question and the optimized question in multi-turn follow-ups, and check whether the subject was altered incorrectly | Used for coreference resolution and question expansion |
| Result re-ranking | When a re-ranking model is available and enabled by configuration, compare ordering before and after re-ranking | Re-ranking is an optional path and does not run on every request |
Three common errors
- Treating minimum relevance as a permission or entity filter. The input is a rebate question spanning dealers; the comparison object is the cited content returned by retrieval; the observable result is whether other dealers' materials still appear after the low-relevance threshold is adjusted. If they appear, the threshold does not isolate entities, and a separate permission implementation is required.
- Treating a pending-confirmation status as completed. The input is the sample source text "pending dealer confirmation, and the red-letter invoice has not been issued"; the comparison object is the draft sentence "has completed calculation and can be offset by invoice issuance"; the observable result is that the status fields in the two texts do not agree, and the draft sentence is not supported.
- Associating knowledge bases built on different vector models. The input is two bases with different model identifiers; the comparison object is the model identifier in the node association configuration; the observable result is that bases with non-matching identifiers must not be associated together, and only the identifier check result is recorded.
How to verify the setup
- Input the sample source text and the draft sentence, compare the period, subject and status fields in the retrieved cited content against the draft sentence, and observe whether the status fields are mutually exclusive between "pending confirmation" and "completed". When they are mutually exclusive, the draft sentence is not supported.
- Input an explicit paper empty result array [] and compare how the downstream design handles empty citations. The observable result is that the output path completes without producing fabricated content. No claim is made that a retrieval for a fictional rule necessarily returns an empty array.
- Input an identifier-type question and a natural-language question, and compare the recalled content under full-text retrieval and semantic retrieval respectively. The observable result is that the identifier hits the corresponding source text and the natural-language question recalls semantically close passages. A first-hit guarantee or a full-recall guarantee is not provided.