Dealer rebate calculation: model connections for product questions

What the documents look like

Dealer rebate calculation happens after the sale. The working materials are the rebate policy, the sales attainment table, the reconciliation statement and the red-letter invoice. The underlying data comes from ERP sales records, dealer payments, the contract ledger and the finance system, and the update rhythm is mainly monthly and quarterly. For product questions and selection answers, these documents have to be reachable through a model connection and configuration that can call policy wording, attainment figures and invoice offset information correctly.

The difficult part is that wording differs between documents. The same product can carry a different rebate rate, a different treatment of returned goods, and a different invoice offset rule under different policies. A question about a product therefore has to be answered against the policy text that actually applies, not against a general impression of how rebates usually work.

Constraints at the model connections step

Model connections are configured through AI Proxy. From V4.17.0, AI Proxy is a required dependency: both the App service and the Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; if either is missing, startup validation fails. The endpoint holds the service root address and does not carry /v1. An older OpenAI variable or a single-model custom request address does not replace these two settings.

Model capability is checked against the input the application actually needs. Multimodal input, tool choice and function call, and thinking configuration are verified separately, because a model that handles one does not automatically handle another. Retrieval libraries and index models are checked by identifier: a library built on one vector model cannot be associated together with a library built on a different vector model. Semantic recall and relevance scores do not carry permission or entity isolation duties, and a retrieval result is not an authorization.

A clearly labelled hypothetical paper test case follows. In September 2026 a brand finance reviewer checks third-quarter dealer rebates. The material text is: "2026 Q3 rebate policy: Class A dealers with a sales attainment rate of 90% or above receive a 3% rebate; returned goods amounts are deducted from the attainment base." The draft sentence to be judged is: "Class A dealers reached 92% attainment in Q3, returned goods amounts have been deducted, and the rebate is 3%." The question is whether the draft sentence agrees with the policy text on subject, period and field.

Configuration and implementation checks

The settings below relate to this sample and this task only. The values are hypothetical starting points for this paper comparison, chosen so that the sample material can be checked against them. They are not official defaults.

Three common errors

1. Entering a full request address with /v1 in AIPROXY_API_ENDPOINT. The input is the configured value; the comparison object is the documented root-address form; the observable result is whether the address ends in /v1.
2. Treating a retrieval result as the authorization basis for a rebate calculation. The input is the returned document fragment; the comparison object is the subject, period and field of the sample material text; the observable result is whether the fragment agrees with the policy wording.
3. Treating the rebate policy as evidence that 92% attainment has been reached and returned goods have already been deducted. The material states a policy threshold of 90% for a 3% rebate and a deduction rule, but supplies no attainment table or deduction record. Compare the policy with the draft and check the source for each of those two claims. Missing data remains pending verification; a policy does not establish attainment or an executed deduction.

How to verify the setup

1. Verification one: input the sample material text and the draft sentence, then compare the three elements of the draft — the 92% attainment rate, the deduction of returned goods, and the 3% rebate — against the subject, period and field of the original text. The original states that a rate of 90% or above receives a 3% rebate and that returned goods amounts are deducted from the attainment base. The 92% figure and the deduction claim in the draft need corresponding data support. Whatever the material does not state is recorded as awaiting confirmation.
2. Verification two: input the hypothetical App/Admin configuration values from the table. Check that AIPROXYAPIENDPOINT is the AI Proxy service root without /v1, and that AIPROXYAPITOKEN is non-empty and consistent with the corresponding server configuration. If either required item is missing, record the unmet startup condition. No actual environment configuration is supplied, so this paper check does not establish a successful startup or model connection.
3. Verification three: compare the hypothetical model probe interval of 30 minutes in the table with the documented integer range of 5 to 60 minutes. The value 30 is within that field range. Whether automatic probing is enabled and what an actual probe returns require separate checks; the paper range comparison does not report an executed probe.

Across all three checks, a judgement requires the same subject, the same period and the same field. Only mutually exclusive values in the same field count as a contradiction. A field that is not stated, or evidence that covers only a related stage, is recorded as unsupported or awaiting confirmation. An unsigned termination is not a completed termination, a pending request is not an approval, and an offered option is not an execution. Where the material needed for a check is missing from the accessible retrieval scope or from a stated manual input, the gap is recorded as a gap.

Related reading

• self-host/config/env
• self-host/config/model/intro
• FastGPT model connections