What the documents look like
Rebate calculation for dealers happens after sales are complete, so the operational focus sits on policy wording and financial treatment rather than on stock movement. The working materials come from ERP sales records, dealer remittances, the contract ledger and the finance system. Day-to-day documents include the rebate policy, the volume achievement sheet, the reconciliation statement and red-letter invoices used for offsetting issued invoices.
A rebate policy normally states the eligible parties, the achievement conditions, the calculation basis and the effective period. A volume achievement sheet records a period and a quantity. A reconciliation statement records amounts both sides have confirmed. A red-letter invoice supports an offsetting entry. Together these documents form the evidence base for any sentence that mentions rebates. Fields not present in the material remain unknown and cannot be filled in by inference from another document.
The drafting task on this page is product and marketing copy. The question for each proposed sentence is whether the supplied material supports it, not whether the sentence reads well.
Constraints at the context limits step
When drafting copy, the maximum context of the selected model limits how much content a single request can reference. The reply cap controls how long a single answer may be. These are separate settings and should not be treated as one number. The knowledge base citation limit is measured in tokens, so a single long policy paragraph combined with a sample question needs to be checked against that limit to see whether content was truncated. The same question produces different input when history is empty and when history is present, so comparisons should hold the history condition constant. When a fixed reply format is used, field completeness and successful parsing need to be verified rather than assumed from text length.
In retrieval, semantic search and full-text search can be combined as hybrid search, and the results are merged through RRF. RRF is a fusion step in the pipeline, not a standalone switch that can be turned on. The minimum relevance threshold applies only in semantic search or when result reranking is used; it acts as a filter threshold and does not perform permission or entity isolation. Reranking requires a rerank model and must be enabled by configuration. Image retrieval requires image input and a model with the corresponding capability. The optional paths shown in the overall flow diagram do not all execute on every request.
Hypothetical drill sample. Material text: "This policy applies to contracted dealers in the East China region. The rebate period runs from 1 January 2024 to 30 June 2024. After the quarterly purchase target is achieved, the rebate is calculated at 2 percent of the purchase amount." Proposed copy sentence: "All dealers can receive a 2 percent rebate in the first half of 2024." This sentence widens both the eligible parties and the achievement condition. The material supports only contracted dealers in the East China region, and only where the target was achieved.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Maximum context | The value shown for the selected model | Compare against the combined length of the material text and the proposed sentence to confirm no truncation |
| Reply cap | 800 tokens as a hypothetical starting point for this trial | Compare against the proposed sentence plus its supporting explanation to observe early truncation |
| Knowledge base citation limit | 2000 tokens as a hypothetical starting point for this trial | Compare against the length of the single rebate policy paragraph to check whether cited content is complete |
| Minimum relevance | 0.4 as a hypothetical starting point for this trial | Compare against scores returned by semantic search to observe whether low-scoring content is filtered |
| Reply format | Default | Change to json_schema only when a downstream step needs to read fixed fields |
| Memory turns | Default | Compare an empty-history input and a with-history input for the same question to hold context constant |
These values are hypothetical starting points for this trial and are not official defaults.
Three common errors
- Rewriting "contracted dealers in the East China region" as "all dealers." Input is the material text and the proposed sentence. Compare the eligible-party field and observe whether the scope of the proposed sentence exceeds the source.
- Rewriting "after the quarterly purchase target is achieved, calculated at 2 percent" as "can receive a 2 percent rebate." Input is the material text and the proposed sentence. Compare the achievement-condition field and observe whether the proposed sentence drops the precondition.
- Writing a rebate period other than "1 January 2024 to 30 June 2024." Input is the material text and the proposed sentence. Compare the period field and observe whether the two match.
How to verify the setup
- Enter the material text and the proposed sentence, compare the eligible-party field, and treat the condition as satisfied when the scope of the proposed sentence stays within "contracted dealers in the East China region." Record it as unsupported when the scope exceeds that.
- Enter the material text and the proposed sentence, compare the achievement-condition field, and treat the condition as satisfied when the proposed sentence keeps the "quarterly purchase target is achieved" precondition. Record it as unsupported when the precondition is dropped.
- Enter the material text and the proposed sentence, compare the period field, and treat the condition as satisfied when the period matches "1 January 2024 to 30 June 2024." Record it as unsupported when the two differ.
If retrieval returns no results, a paper input of [] can be used to check the downstream handling design; this does not claim that a fictional rule search will necessarily return []. Libraries built on different vector models cannot be associated together. Verify the identifier first, exclude those that do not match, and record the actual observation.