What the documents look like
Producer export tax documentation for the exemption, offset and refund process draws on four document families: the exemption, offset and refund declaration detail sheet, the export contract, the sales subsidiary ledger, and the customs export declaration. These materials update monthly. When a draft cites a product, quantity, trade term or free-on-board basis, the drafter checks the subject, the period and the corresponding batch separately. Two documents sharing a month does not automatically place them in the same transaction. A monthly update frequency does not imply a synchronization schedule, a trigger mechanism or a batch process.
For copy drafting, the practical consequence is that a sentence can only restate a field that a named document carries, and only for the subject and period that document states. A quantity appearing in two documents with matching values remains two separate records until a shared business identifier connects them.
Constraints at the model connections step
Model assistance supports drafting and checking copy. It does not perform refund approval or declaration writes. The App service, and any Admin service that reuses App/Service configuration, must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN. Missing either item causes startup validation to fail, and older OpenAI environment variables or single-model custom request addresses do not substitute for these two settings. The endpoint value is the AI Proxy service root address without an appended /v1. The system requires at least one language model and one index model to operate.
Multimodal input, tool selection and function calling are verified against the actual input the drafting task supplies. Application parameters are set in the application or node configuration; a same-named field on a compatible chat interface does not override orchestration settings. A shared vector model requirement applies to knowledge bases that share a common association, and does not extend to external websites. A relevance score does not replace permission checks or subject verification.
Hypothetical paper exercise
The following materials and draft sentence are a paper exercise only. They describe no real customer, company or legal identity number.
Material one (hypothetical export contract excerpt): "The seller is a domestic production enterprise. The contract subject matter is self-produced household textiles. The trade term is FOB. The contract amount is stated on a free-on-board basis. The contract period is March 2025."
Material two (hypothetical customs export declaration excerpt): "The domestic consignor is that production enterprise. The commodity name is household textiles. The trade term is FOB. The declaration date is March 2025. The declared quantity is 1,200 pieces."
Material three (hypothetical exemption, offset and refund declaration detail sheet excerpt): "The period is March 2025. The export goods name is household textiles. The export quantity is 1,200 pieces. The export sales amount is entered on a free-on-board basis."
Draft sentence to be assessed (hypothetical): "This product was exported in March 2025 in a quantity of 1,200 pieces, under the FOB trade term, with the export sales amount stated on a free-on-board basis."
Configuration and implementation checks
The table below records paper-level connection checks for this page. Trial values are compared against the actual deployment.
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| AIPROXYAPIENDPOINT | Filled as http://aiproxy:3000 | Documented guidance states the AI Proxy service root address with no appended /v1; compare against the documented example format and confirm no extra path segment |
| AIPROXYAPITOKEN | Filled as a non-empty administrator token matching the server side | Documented guidance states the value cannot be empty and an application API key does not substitute; compare for consistency with the AI Proxy server configuration |
| Language model and index model | One of each assumed enabled | Documented guidance states the system requires at least one language model and one index model; compare against the enabled list and confirm both categories are non-empty |
| Model capability check | Assumed checked against plain text input | Documented guidance separates multimodal and tool-calling verification; compare against the actual input type of the drafting task and confirm the required capability is enabled |
| Model probe interval | Assumed set to 30 minutes | Documented guidance supports an integer between 5 and 60 minutes; compare the value against the permitted range |
| Alert Webhook address | Assumed left unconfigured | Documented guidance treats the address as optional; compare against whether status-change notifications are needed, and record an unconfigured state as not enabled |
Model probing covers only system models currently enabled. Disabled models do not participate in probing and produce no history. After a probe request first fails, the system retries automatically up to 3 times, so a single round makes at most 4 attempts; only when all 4 attempts fail is the model judged abnormal. The alert Webhook fires on an edge change of model status, and a model that stays abnormal does not trigger repeated pushes.
Three common errors
- Treating the declared quantity of 1,200 pieces as the contract quantity. The contract excerpt carries no quantity. The declaration and the detail sheet each state 1,200 pieces, but a shared month does not establish a shared batch. Keep each documented field with its own source; cross-document attribution and the contract quantity remain to be verified.
- Treating a relevance score as a permission or entity filter, on the assumption that a similarity setting excludes unrelated business or unauthorized records. The documented material does not assign that function to relevance. Retrieval results do not constitute authorization, and business isolation expectations should not be built on that setting.
- Reading the enabled scope of model probing as covering all models, and setting alert expectations for status changes of disabled models. The documented material states that disabled models do not participate in probing and produce no history. Check enablement status before configuring alerts.
How to verify the setup
- Enter the three excerpts and the draft sentence, and check subject, month and quantity item by item. The declaration states that enterprise, March 2025 and 1,200 pieces. The detail sheet states the same month and the same quantity but carries no subject or business association identifier, so cross-document attribution remains to be verified and equal numeric values alone do not establish a correspondence.
- Enter the phrase "the export sales amount is stated on a free-on-board basis" together with the detail sheet. The source text supports that sheet's own stated pricing basis. Applying it to this product still requires subject and business correspondence evidence; where that is absent, the item remains to be verified.
- Enter the phrase "under the FOB trade term" together with the contract and the declaration. Both carry FOB, so each record can be retained and checked for whether they describe the same transaction. If a draft additionally states "export tax refund eligibility has been obtained," none of the three documents carries an eligibility status, and that claim is unsupported.
These steps are paper-level copy checks and do not execute a declaration or an approval. An empty-reference handoff may be checked against a downstream design using a paper-level empty input; actual query results remain to be observed.