What the documents look like
Rebate calculation happens after sales. The controlling materials are policy terms and financial treatment, so copy drafting draws on ERP sales records, dealer remittances, contract ledgers and finance systems. The recurring documents are rebate policies, sales achievement tables, reconciliation statements and red-letter invoices, refreshed monthly or quarterly. A draft sentence that states a rebate amount or an achievement conclusion needs a traceable field in one of those materials. Fields the materials do not carry stay unknown.
Three industry facts shape this page. First, the achievement basis is a policy term, so the policy text and the achievement table must agree on the same subject, the same period and the same threshold field before any conclusion is repeated. Second, returns and invoice offsets are separate steps that reduce or adjust the amount, so a draft that states a final figure reaches beyond a policy clause that only names a rate. Third, reconciliation statements and red-letter invoices carry their own status, and a pending item stays pending.
Constraints at the model connections step
Model access runs through AI Proxy. The App and Admin services must both set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; a missing value in either one fails startup validation. Older OpenAI variables and single-model custom request addresses do not replace these two settings.
Model channels are created on the administrator model provider page. The model ID is the value of the model field in the request body and is globally unique. Model mapping sends a system-side model ID to a different upstream model, for example a mapping whose key is gpt-4o-test and whose value is gpt-4o. Copy drafting requests travel to whatever upstream model the mapping names.
Model probing covers only currently enabled system models. A first failed probe is retried up to three times, so one round makes at most four attempts, and only four failed attempts mark the model abnormal. Probe states are normal, high latency and abnormal. The probe interval accepts whole minutes from 5 to 60.
Retrieval association requires the same vector model. Libraries held in different vector models cannot be associated together. This check reads identifiers and excludes non-matching libraries; it does not run a cross-model search to compare recall. An empty result is checked against a stated paper input, and no claim is made that a fictional rule lookup returns an empty array.
Hypothetical paper exercise sample. A brand finance team supplies a rebate policy whose text reads: "For the first quarter of 2026, where dealer A reaches a sales achievement rate of 95% or above, the quarterly rebate is calculated at 2%." The sales achievement table reads: "Dealer A, January to March 2026, achievement rate 96%." The draft sentence awaiting judgment reads: "Dealer A has satisfied the conditions for the 2% rebate for the first quarter of 2026." The comparison points are the policy subject, the object, the period and the achievement threshold against the same fields in the achievement table.
Configuration and implementation checks
| Configuration item | Implementation check on this page | Relation to the sample |
|---|---|---|
AIPROXY_API_ENDPOINT | Enter the AI Proxy service root address without appending /v1 | Model call entry point used by the drafting request |
AIPROXY_API_TOKEN | Enter a non-empty administrator token matching the server configuration | Call credential; an application API key does not substitute |
| Model channel and model mapping | Configure the model inside the channel; the mapping key is the system-side model ID | Determines the upstream model that receives the drafting request |
| Model capability | Check multimodal and tool-calling fields against the input the application needs | Text-only drafting is checked against the fields actually configured |
| Application parameters | Parameters are set in the application or node configuration | No assumption that a same-named chat interface field overrides orchestration |
| Model probe | Probes cover enabled models only; the interval is a whole number of minutes from 5 to 60 | Observes whether the model used for drafting is reachable |
Three common errors
- Reversing the model mapping key and value. The input is the system-side model ID and the upstream model ID. Compare the direction of the mapping JSON, and observe whether the
modelfield in the request body carries the system-side ID. - Treating a pending review state as granted authorization. The input is the draft sentence and the policy text. Compare the subject, period and threshold fields, and observe whether the draft states an authorization or effective conclusion that the source text does not carry.
- Using a similarity setting to exclude unrelated records. The input is the retrieval request and the returned list. Compare the business subject identifiers in the returned records, and observe whether records for another dealer or another period appear.
How to verify the setup
- Enter the policy text and the achievement table text from the hypothetical sample. Compare the subject, object, period and threshold fields of the draft sentence against them. The condition is met when the draft sentence restates only the 96% achievement rate and the 2% rebate basis recorded in the source text, and adds no signing, approval or payment conclusion.
- Enter the draft sentence from the hypothetical sample. Compare its field correspondence with the policy text. Observe whether the phrase "has satisfied" reaches beyond the recorded scope of "the quarterly rebate is calculated at 2%". Any part beyond that recorded scope is recorded as unsupported.
- Enter the period "first quarter of 2026" from the hypothetical sample. Compare it with the January to March coverage in the achievement table. Observe whether the period matches on the same field. Where the material covers only part of the months, record an evidence gap.
Retrieval checks describe observation and failure branches. No guarantee is made that the first item matches, that all relevant items are recalled, or that an unrelated query returns nothing. An empty result is checked against a stated paper input such as an empty list, and no claim is made that a fictional rule lookup returns an empty list. Where libraries sit in different vector models, the check reads identifiers and excludes non-matching libraries; it does not execute a cross-model search to compare missed recall.
The optional paths in the overall flow diagram do not run on every request. Image retrieval requires a corresponding image input and model capability. Reranking requires a rerank model and its configured enablement. A judgment requires the same subject, the same period and the same field; only mutually exclusive values in that same field count as a contradiction. Values that are not recorded, or that show only a related stage, are recorded as unsupported or pending verification. A pending review does not show that an operating permission is absent. Signing, reaching cooperation and contract effectiveness are three distinct facts. A quotation validity period does not state a supply period. Where the material needed for a check is missing from the accessible retrieval scope or the stated manual input, the gap is recorded as missing.