What the documents look like
Copy drafting for TikTok Shop US store operations draws on a narrow set of working records: product listings, return and refund records, return grading evidence, and settlement statements. The data arrives daily and per order from TikTok Shop Seller Center, the order center, carrier tracking, and advertising reports.
Two features of this category shape how copy is drafted. First, return handling is governed by local return windows, return grading, and overseas logistics constraints, so a listing sentence about returns sits close to operational consequences such as platform automatic refunds, higher return fees, or failed appeals. A sentence that promises a refund condition the materials do not state can move a buyer expectation beyond what the seller can honor. Second, return grading means the condition of opened goods is assessed rather than assumed, so draft copy that treats every opened item as fully refundable conflicts with how the underlying records describe the same object.
In this page the drafting step is the primary task: a proposed listing or marketing sentence is compared against supplied source wording. Model connection and configuration are the supporting conditions that make the drafting step runnable and testable. The configuration described here does not decide whether a sentence is supported; the source wording does.
Constraints at the model connections step
From V4.17.0 onward, AI Proxy is a required dependency. The App service and the Admin service that reuses App/Service configuration both must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN. Missing either one causes startup validation to fail. Older OpenAI environment variables and single-model custom request addresses do not substitute for these two settings.
AIPROXY_API_ENDPOINT takes the AI Proxy service root address, for example http://aiproxy:3000, and must not carry /v1. The token must match the server-side configuration and must not be empty.
The system needs at least one language model and one index model to work normally. A model ID is the value of the model field in the request body and is globally unique. Model mapping lets a model requested by FastGPT resolve to a different upstream model, for example {"gpt-4o-test": "gpt-4o"}, where FastGPT requests gpt-4o-test and AI Proxy sends gpt-4o upstream.
Model status probing applies only to system models that are currently enabled. Disabled models are not probed and produce no history. The probe interval accepts an integer between 5 and 60 minutes. After a first failure the system retries up to 3 more times, so a single round makes at most 4 attempts; one success marks the round valid, and only four failed attempts produce an abnormal result. A model test success with a response latency of 30 seconds or less is normal, success above 30 seconds is high latency, and four consecutive request errors is abnormal. These are the documented conditions for the probe; this page does not assume a scheduled trigger beyond the documented automatic probing.
Configuration and implementation checks
The table below ties documented settings and implementation checks to the hypothetical sample used on this page. The numeric starting values are hypothetical trial starting points chosen for comparison, not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
AIPROXY_API_ENDPOINT | http://aiproxy:3000 | Documented example of a service root address; compare the value for an appended /v1 or /chat/completions |
AIPROXY_API_TOKEN | Non-empty administrator token | Documented requirement that it match the server side and not be empty |
| Language model | Enabled | Copy drafting needs text dialogue capability |
| Index model | Enabled | The system needs at least one index model |
| Model mapping | gpt-4o-test to gpt-4o | Compare the requested model with the upstream model |
| Probe interval | 30 minutes | Documented range is an integer from 5 to 60 minutes; compare the configured value with that range |
Three common errors
- Treating the draft sentence "opened items are also fully refundable" as supported by the supplied material. The material states only that opened items are handled according to the return grading result. It does not state that opening alone produces a full refund, so that sentence is unsupported.
- Entering
AIPROXY_API_ENDPOINTas a full request address that carries/v1or/chat/completions. The documentation requires the service root address, so an appended path does not match the documented form. - Reading the probe rule as "first failure means abnormal." The documentation states that after a first failure the system retries up to 3 more times, and one success marks the round valid. Only four failed attempts produce an abnormal result.
How to verify the setup
The hypothetical paper test case below is labelled as such. It uses no real store, order, or customer, and nothing here was executed or approved.
Hypothetical paper test case. A TikTok Shop US operations staff member prepares to draft one listing sentence. The supplied source material, quoted as text to be checked, is: "This product supports returns within 30 days. Returned items must keep the original packaging intact. Opened items are handled according to the return grading result." The proposed draft sentence to be judged is: "This product supports no-reason returns within 30 days, and opened items are also fully refundable." The comparison scope is limited to this listing draft and the three source sentences above.
- Input the source sentence "This product supports returns within 30 days" and the draft clause "no-reason returns within 30 days." Compare the same object, the product, on the same field, the return condition. The source material does not contain the phrase "no reason," so the draft clause is unsupported on that field. The clause meets the supported condition only when the source wording explicitly states "no reason."
- Input the source sentence "Opened items are handled according to the return grading result" and the draft clause "opened items are also fully refundable." Compare the same object, opened items, on the same field, the refund condition. The source gives handling according to the return grading result; the draft gives a full refund. The two values differ on the same field, so the clause is unsupported or pending verification. It meets the condition only when the source wording explicitly states that opened items receive a full refund.
- Input the hypothetical value
http://aiproxy:3000forAIPROXY_API_ENDPOINTand compare it for the presence of/v1or/chat/completions. Absence of both matches the documented root address requirement. Presence of either does not match the documentation and requires correction before the check is repeated.
A pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Where the source material does not state a refund condition for a given object and stage, that condition remains missing rather than disproven.