What the documents look like
Export tax rebate filing work happens after goods leave the country, when tax vouchers are matched against export records. The working documents are the export tax exemption and rebate filing form, the export customs declaration, purchase invoices, and foreign exchange receipt vouchers. For product and marketing copy drafting, these documents supply the checkable facts a draft may cite: product name, specification, quantity, transaction terms, and date range. Records update monthly and per batch, so each drafted sentence should trace to a field in a specific voucher rather than to a general impression of the shipment.
This matters for model connections because the drafting task is a text task with a narrow evidence base. The model does not verify tax status, approve a rebate, or synchronize with any tax system. It receives document text and produces candidate copy. Whether a claim such as completed review or full rebate appears in the draft depends on whether that fact exists in the supplied material, and the material here covers customs and invoicing data, not review outcomes.
Constraints at the model connections step
Model access requires both AIPROXYAPIENDPOINT and AIPROXYAPITOKEN to be configured and non-empty. Missing either one causes startup validation to fail. The endpoint is the AI Proxy service root address, for example http://aiproxy:3000, and must not carry /v1. The token is the AI Proxy administrator token, must match the server configuration, and cannot be replaced by an application API key.
Within model configuration, the model ID is the value of the model field in the request body and is globally unique; the model name is display-only. For classification and content extraction, toolChoice is used first, then functionCall; when neither is supported, the documented prompt mode applies. Model probing covers enabled system models only, attempts at most four times in one round, and judges a model abnormal only when all four attempts fail. On the retrieval side, only libraries sharing the same vector model can be associated together, and a relevance score carries no permission or entity isolation duty.
Hypothetical paper exercise sample: a finance staff member at a foreign trade enterprise supplies material text reading "the export customs declaration states the product is a stainless steel insulated cup, specification 500 ml, quantity 2,000 units, declaration date March 10, 2026" and "the purchase invoice states the same specification insulated cup, 2,000 units, invoice date March 5, 2026." The draft sentence to be judged is "this product has passed export tax rebate filing review and qualifies for a full rebate." The material contains no record corresponding to the review claim.
Configuration and implementation checks
| Configuration item | Hypothetical starting value | Reason and comparison method |
|---|---|---|
| AIPROXYAPIENDPOINT | Assume http://aiproxy:3000 | Documented as a root address; compare the entered value against the no-/v1 requirement |
| AIPROXYAPITOKEN | Assume the administrator token | Must match the server and be non-empty; compare startup validation result |
| Language model tool capability | Assume toolChoice enabled | Copy classification and field extraction use it; compare whether extraction runs in tool mode |
| Model probe interval | Assume 30 minutes | Documented range is an integer from 5 to 60 minutes; compare probe records against that interval |
| Vector library identifier | Assume a single library | Compare whether retrieval scope stays inside that library and excludes cross-library association |
Three common errors
- Entering a full chat address with
/v1appended as AIPROXYAPIENDPOINT. The documented requirement is the service root address. The check compares the entered value against the root address format and records a mismatch. - Using an application API key in place of AIPROXYAPITOKEN. The credential purpose differs. The check compares the token source against the server configuration and records whether they agree.
- Treating the draft sentence "has passed export tax rebate filing review" as supported. The check compares that sentence against the supplied material, which states customs and invoice facts only; the review status does not appear, so the claim is recorded as unsupported.
How to verify the setup
- Input the declaration and invoice text from the hypothetical sample, together with the draft sentence "this product has passed export tax rebate filing review and qualifies for a full rebate." Compare the review status in the sentence against the material for the same subject, same period, and same field. Check separately for review-passage evidence and for the scope or amount basis of "full rebate." Neither appears in the material, so record each as unsupported. A review record alone would not supply the full rebate conclusion.
- Input the product name, specification, and quantity fields from the same batch of material. Compare the customs declaration against the purchase invoice. Both show a stainless steel insulated cup, 500 ml, 2,000 units, so the fields agree; if either differs, record it as pending verification.
- Input a retrieval query containing only the product name, and observe whether results stay within the same vector library. If the result is empty, use the paper input
[]to check the downstream copy design rather than inferring that a rule necessarily returns empty. Libraries on different vector models are not associated together; check identifiers and exclude non-matching ones instead of running a cross-model search to compare recall.
The model connection itself does not authorize, approve, or file anything. Any business verification, approval, database write, or external synchronization requires its own downstream implementation, and none of those steps is established by a passing model test.