Export tax rebate filings: model connections for product questions

What the documents look like

Export tax rebate filing happens after goods leave the country. The work matches tax vouchers against what was declared, and the main data comes from the electronic tax bureau, the export rebate filing system, customs electronic messages, and the finance ledger. The documents handled day to day are the export tax exemption and rebate filing form, the export customs declaration, purchase invoices, and foreign exchange receipt vouchers.

Product data and selection questions sit on top of these documents. A question about a product, a supplier, or a document field has to retrieve the right record and answer from it. Model connection and configuration decide whether the answer can cite the correct field, and whether the reply stays inside what the documents actually state.

Two features of this category shape that work. First, the filing is a post-export voucher matching exercise, so a product question often turns on whether a declaration line, an invoice line, and a receipt line agree. Second, the material arrives monthly and per batch, which means the retrieval target changes as each batch is filed.

Constraints at the model connections step

The system needs at least one language model and one index model before it works normally. AI Proxy is a required dependency. Both the App service and an Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN. Missing either one causes startup validation to fail. AIPROXY_API_ENDPOINT takes the service root address and does not carry a /v1 suffix. Older OpenAI environment variables and single-model custom request addresses do not replace these two settings.

Model channels, model mapping, and custom request addresses are configured on the administrator model provider page. A model channel carries a protocol type, the models available on that channel, and an API key. Model mapping translates the model FastGPT requests into the model sent upstream. Model status probing covers only system models that are currently enabled, and the probe interval accepts an integer between 5 and 60 minutes.

Retrieval is not authorization. A similarity score is not a category, entity, or permission filter, and it does not exclude unrelated businesses or unauthorized records. Dataset and vector model association must be checked by identifier first. Datasets built on different vector models cannot be associated together; the check is to confirm the identifier, exclude the ones that do not match, and record what was actually observed. An empty retrieval result can be checked against a downstream design by supplying an explicit paper input of []. This page does not claim that a rule-based lookup returns [].

Assume a paper exercise sample. A document clerk at a foreign trade enterprise is organizing one batch of export goods documents in March 2026. The source material reads: "Export customs declaration number 20260301001, export date 2026-03-01, goods name steel fasteners, quantity 12,000 pieces; corresponding purchase invoice number 20260305002, invoice date 2026-03-05, amount consistent with the customs declaration." The draft sentence under review reads: "This batch of goods has completed the rebate filing and received the rebate payment." The source material contains only the match between the customs declaration and the purchase invoice. It does not state a filing status and does not record receipt of any rebate payment, so the draft sentence is not supported.

Configuration and implementation checks

The table below lists settings or implementation checks that the official documents describe and that relate to this page's feature and primary task. The assumed trial starting values are for this paper comparison only and are not official defaults.

Three common errors

1. Filling AIPROXY_API_ENDPOINT with a full request address that carries /v1. The input is an assumed http://aiproxy:3000/v1. The comparison object is the service root address the document requires. The observable result is that the value carries a /v1 suffix and does not match the requirement.
2. Treating the draft sentence "This batch of goods has completed the rebate filing and received the rebate payment" as a confirmed fact. The input is the source material above together with the draft sentence. The comparison object is whether the source material states a filing status and records receipt of a rebate payment. The observable result is that the source material contains only the customs declaration and purchase invoice match, so the draft sentence is not supported.
3. Treating a similar record returned by retrieval as an authorized action. The input is a set of similar documents returned by one retrieval. The comparison object is whether an authorization or approval record exists. The observable result is that only a similarity score is present and no authorization basis exists, so the record cannot be treated as authorized.

How to verify the setup

1. Input the assumed AIPROXY_API_ENDPOINT value. Compare whether it is a service root address without a /v1 suffix. The condition is met when the observable result is a value with no /v1 suffix that matches the deployment template.
2. Input the source material and the draft sentence given above. Compare whether the draft sentence is stated in the source material. The condition is met when the observable result is that the source material does not state a filing status or record receipt of a rebate payment, and the draft sentence is marked as not supported.
3. Input one retrieval result set and the corresponding authorization records. Compare whether an authorization basis exists for the same subject, the same period, and the same field. The condition is met when the observable result is that only a similarity score is present and no authorization record exists; the item must not proceed to execution and is recorded as an evidence gap.

Related reading

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