What the documents look like
Marketplace refund and return handling occurs continuously per order and does not necessarily connect to a promise made in a single livestream session. The documents handled day to day include refund applications, return logistics notes, quality inspection photos and refund approvals. When drafting product and marketing copy for this category, the source material comes from product information and after-sales records. Model connections generate text only. They do not determine return responsibility or approve refund amounts. Data updates per order and daily. This page does not derive batch synchronization or transaction-triggered mechanisms from that frequency.
Constraints at the model connections step
Model access depends on AI Proxy. From V4.17.0, AI Proxy is a required dependency. Both the App and an 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. Legacy OpenAI environment variables or a single-model custom request address cannot replace these two settings.
The model ID is the value of the model field in the request body and is globally unique. The model name is for display and can be customized. Drafting copy uses a language model. Recognizing quality inspection photos requires the model to have vision capability. Model probing applies only to currently enabled system models. The probe interval accepts an integer between 5 and 60 minutes. An abnormal result is recorded when four consecutive probe requests all fail.
Hypothetical paper exercise sample: a marketplace after-sales team keeps paper material stating that the refund application for order A-2201 lists the product as a vacuum flask, the return logistics note states the signed delivery date as March 12, and the quality inspection photo caption states the flask body has a scratch. The draft sentence to be assessed is: "This product has been confirmed by quality inspection to have a scratch, and a full refund is supported." The material does not state a quality inspection conclusion or a refund approval result.
Configuration and implementation checks
The settings below concern copy-drafting checks for this page's sample. The values are hypothetical starting points for this exercise, used to compare against values displayed in the model probe panel.
| Configuration item | Hypothetical starting point for this exercise | Comparison method |
|---|---|---|
| AIPROXYAPIENDPOINT | Enter the AI Proxy root address, without appending /v1 | Compare against the deployment document example structure |
| AIPROXYAPITOKEN | Enter a non-empty administrator token | Compare against the AI Proxy server-side configuration |
| Language model enabled state | Enable one language model | Compare against the enabled list on the model probe page |
| vision capability | Decide based on whether quality inspection photos need recognition | Compare against the vision value in the model configuration fields |
| Model probe interval | Take an integer between 5 and 60 | Compare against the allowed range in the probe configuration dialog |
| Alert Webhook address | Optional; enter a callback URL | After clicking test, compare against the push result |
Three common errors
- Writing
AIPROXY_API_ENDPOINTas a complete chat endpoint address with/v1appended, so startup validation fails. Check whether the input is a root address, compare against the deployment document requirement, and observe whether the service passes startup validation. - Treating the draft sentence "a full refund is supported" as a fact already supported by the material. Check the input as the original sample material text, compare whether the material states a quality inspection conclusion and a refund approval, and observe whether that sentence lacks a basis.
- Using a similarity setting to exclude unrelated business or unauthorized records. Check the input as the retrieval configuration item, compare whether it is a permission or entity filter, and observe whether the scope of that setting is wrongly expanded.
How to verify the setup
- Input the three sentences of the hypothetical sample material and the draft sentence. Compare whether the draft sentence goes beyond the scratch fact stated in the material. If "a full refund is supported" appears, judge it as unsupported and keep it pending verification.
- Input
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKEN. Compare whether both are non-empty and whether the endpoint has no/v1appended. If either is missing or the format does not match, startup validation fails and the configuration is incomplete. - Input the paper retrieval result
[]. Compare whether downstream copy drafting handles it as an empty reference. If downstream still outputs a specific promotional conclusion, the empty-result handling did not execute as designed, and the observation should be recorded.