What the documents look like
Rebate calculation happens after sales close, and its center of gravity is policy wording and financial treatment. The working materials are rebate policies, sales achievement tables, reconciliation statements and red-letter invoices, drawn from ERP sales, dealer remittances, contract ledgers and finance systems, refreshed monthly or quarterly. In cross-border, multilingual retrieval, one rebate policy may exist in several language versions, and a sales achievement table may carry different currencies, periods and statistical scopes by region. Before any model connection is configured, the reviewable range of the material itself has to be established: which subject, which object, which period and which region the text actually states.
Constraints at the model connections step
AI Proxy is a required dependency. Both the App service and any Admin service that reuses the App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; if either is missing, startup validation fails. Legacy OpenAI environment variables and single-model custom request addresses do not substitute for these two settings. AIPROXY_API_ENDPOINT holds the AI Proxy service root address and does not carry an appended /v1. Model capabilities are checked against the input the application actually needs, with multimodal support, tool choice and thinking configuration verified separately. A retrieval dataset requires the same vector model association, and that requirement stays inside the vector retrieval scope rather than extending to business registration sites or other external systems. A relevance score carries no permission, entity or business isolation duty, and retrieval by itself is not authorization.
Hypothetical paper sample: a brand finance reviewer submits one document during cross-border retrieval. The source text reads: "This policy applies to dealer rebate calculation for the East China region for the period 1 January 2026 to 30 June 2026; the rebate rate is confirmed against the quarterly sales achievement table." The sentence put forward for judgement reads: "This policy covers rebate calculation for the South China region in July 2026." The material states its subject, object, period and region, and the proposed judgement sentence falls outside the stated period and region.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this page | How to check |
|---|---|---|
AIPROXY_API_ENDPOINT | Hypothetically filled with the internal AI Proxy root address, without an appended /v1 | Compare character by character against the deployment instruction not to append /v1 |
AIPROXY_API_TOKEN | Hypothetically filled with a non-empty administrator token matching the server side | Confirm the value is non-empty and matches the server configuration |
| Model capability check | Select a language model and an index model according to the input this task needs | Verify multimodal support, tool choice and thinking configuration separately |
| Retrieval dataset vector model | Assume one retrieval dataset uses one vector model | Compare the dataset and model identifiers and exclude those that do not match |
| Proposed judgement sentence scope check | Assume the source period and region are the comparison baseline | Compare the period and region fields of the proposed sentence against the source text field by field |
These starting points illustrate a checking method and are not official defaults.
Three common errors
- Treating the period and region of the proposed judgement sentence as if the source text covered them. The input is the hypothetical source text and the proposed sentence above; the comparison is between the period and region fields of each; when the proposed sentence carries fields outside the range stated in the source text, that judgement is not supported.
- Treating a relevance score as a permission or entity filter. The input is one retrieval request and its returned score; the comparison is between that score and the permission configuration; when no correspondence exists between score level and access rights, authorization cannot be inferred from the score.
- Extending the same-vector-model requirement for a retrieval dataset to external systems. The input is a retrieval dataset identifier and an external data source identifier; the comparison is between their model identifiers; when the external system states no model information, the requirement does not apply and the item is recorded as not supported.
How to verify the setup
- Input the hypothetical source text and the proposed judgement sentence above, compare the period and region fields, and when the proposed sentence carries fields outside the source range, record the judgement as not supported rather than as a negation or a conflict.
- Input one retrieval request and observe the returned results together with the empty-result branch. When the return is empty, check the downstream design against an explicit paper input of
[], without asserting that a fabricated rule must return empty. - Input a retrieval dataset identifier and a different vector library identifier, compare their model identifiers, and when they do not match, exclude that library and record the actual observation instead of running a cross-model retrieval to compare missed recalls.