What the documents look like
Product questions for TikTok Shop US operations draw on four document families: product listings, return and refund records, Return Grading evidence, and settlement statements. Compared with shelf-style operations on Douyin Shop, the US marketplace places more weight on local return windows, return grading, and overseas logistics constraints. These records arrive daily and per order, so the same product can carry a different return status on different dates. That matters at the model connection step because the answer scope has to be fixed before any question about answer quality is meaningful. A knowledge base retrieval hit is not an authorization decision. A record that appears in a search result still needs separate confirmation of access rights and business authorization. Datasets built on different vector models cannot be associated together; the model identifier has to be checked first and non-matching datasets excluded. Carrier tracking and advertising reports are stored separately from platform records, so a retrieval range that covers only one source cannot stand in for the others.
Constraints at the model connections step
FastGPT reaches model providers through AI Proxy. From V4.17.0, AI Proxy is a required dependency. Both the App service and any Admin service that reuses the App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN. A missing entry in either one causes startup validation to fail. AIPROXY_API_ENDPOINT holds the AI Proxy service root address and does not carry a /v1 suffix. Older OpenAI environment variables and single-model custom request addresses cannot substitute for these two settings.
The system needs at least one language model and one index model to run normally. A model ID is the value of the model field in the request body and is globally unique; a model name is for display and can be customized. A model channel corresponds to a provider protocol, and one model can be configured across several channels for load balancing. Custom request addresses and keys bypass the model channel, and the documentation marks that path as not recommended.
Model status probing covers only currently enabled system models. After a first failure, the system retries automatically up to three more times, so a single round attempts at most four calls. Only when all four attempts fail is the result judged abnormal. Normal means the test succeeded and response latency was at or below 30s; high latency means success above 30s; abnormal means four consecutive request errors. The alert Webhook fires on edge changes of model status, one notification per model, sent separately.
Configuration and implementation checks
The settings below cover the document scope of this page. Each row states what the official text says and what this page checks.
| Configuration item | Documented content | Check on this page |
|---|---|---|
AIPROXY_API_ENDPOINT | Required, AI Proxy service root address, no /v1 appended | Compare the entered value with the actual service root address |
AIPROXY_API_TOKEN | Required, admin token, must not be empty | Compare with the server-side configuration for consistency |
| Language model | The system needs at least one language model | Confirm an enabled language model exists |
| Index model | The system needs at least one index model | Confirm an enabled index model exists |
| Model ID | Value of the model field in the request body, globally unique | Compare the channel model with the enabled model ID |
| Model status probing | Probes only enabled models; four failed attempts judge abnormal | Confirm the probe scope and the alert Webhook configuration |
Hypothetical paper exercise: a batch of product record checks uses the following material text. "Product A's return records from March 1 to March 31, 2026 total 12 entries, of which 3 are pending review, and 6 Return Grading evidence items have been uploaded." The draft sentence to be assessed reads: "Product A completed all return processing in March." This sample is for paper checking on this page only and does not point to a real store.
Three common errors
- Treating pending review as completed processing. In the sample, 3 return records sit in pending review, while the draft sentence says all return processing is complete. Same subject, same period, same field, mutually exclusive values, so the draft is contradicted.
- Treating a retrieval hit as an authorization pass. A return record appearing in retrieval results does not mean the record was approved for refund or that settlement finished. The business status needs separate confirmation.
- Writing the abnormal judgment of model status probing as abnormal on a single failure. The documentation states that all four attempts must fail before the result is judged abnormal. A single failure does not support an abnormal conclusion.
How to verify the setup
- Enter the sample material text and the draft sentence, then compare whether "pending review" and "completed all return processing" are mutually exclusive on the same field. If both point to the same product, the same period, and the same return status field, the draft sentence is not supported.
- Enter the
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENconfiguration values, then compare whether either is empty and whether a/v1suffix was appended. If both are non-empty and the root address carries no extra path, the startup validation condition is met. - Enter the model probe result, then compare the attempt count with the judgment threshold. If the attempt count is 4 and all attempts failed, the result is judged abnormal; if the attempt count is below 4, the result remains to be confirmed. All three checks are paper checks and were not actually executed.