TikTok Shop US operations: model connections for multilingual work

What the documents look like

TikTok Shop US store operations work with four recurring document families: product listings, return and refund records, Return Grading evidence, and settlement statements. These arrive from Seller Center, the order center, carrier tracking, and advertising reports, and they update daily and per order. US local return windows, return grading, and overseas logistics constraints mean that one return event appears across order records, image evidence, and settlement deductions with different fields in each place.

Multilingual retrieval has to locate the same subject and the same field across these materials. Model connection and configuration determines whether retrieval covers mixed English and Chinese text at all. A return record may carry an order identifier and a review status, an evidence package may carry a grading reference and an image count, and a settlement statement may carry a deduction amount on a stated date. Retrieval returns text; it does not decide whether a return was approved, whether a refund completed, or whether a store is authorized to act.

Constraints at the model connections step

In FastGPT v4.17.2, the App service and any Admin service that reuses App/Service configuration must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN. Missing either one causes startup validation to fail. Legacy OpenAI environment variables and single-model custom request addresses do not substitute for these two settings. The proxy address field in a model channel takes a BaseUrl, not a full request address.

The system needs at least one language model and one index model to work normally. Full-text retrieval follows the vector store in use: with Milvus it uses BM25 over the modeldata_v2 single table, and with other vector stores it uses MongoDB $text. Similarity scores do not act as permission or entity filters, and retrieval does not perform authorization. Datasets built on different vector models cannot be associated together. Model status probing applies only to currently enabled system models; disabled models are not probed and produce no history.

Configuration and implementation checks

Hypothetical paper exercise sample. A handover packet contains three source sentences: "Order A-771 submitted a return request on 2026-03-02, status pending review." "Return Grading evidence package RG-12 was uploaded on 2026-03-05 and contains 4 product images." "Settlement statement B-33 shows a return handling fee deduction of USD 6.50 on 2026-03-10." The draft sentence under review is: "The return for order A-771 was approved and the refund completed." That sentence names order A-771 as subject, approval and refund as objects, and completed as status. The source material states only pending review and records no approval or completed refund.

Three common errors

1. Writing AIPROXYAPIENDPOINT as a full request address that ends in /v1 or /chat/completions. The documented value is the service root address, so startup validation or the call path no longer matches.
2. Treating a retrieval relevance score as a permission or entity filter, on the assumption that a similarity setting excludes other stores or unauthorized records. Retrieval does not perform authorization, and a score is not a category, entity, or permission boundary.
3. Reading a pending-review status as approved or refunded. The source material states pending review and records no approval decision and no completed refund, so the draft sentence is unsupported rather than contradicted.

How to verify the setup

1. Input the source sentence "Order A-771 submitted a return request on 2026-03-02, status pending review" together with the draft sentence "The return for order A-771 was approved and the refund completed." Compare the same order's same status field across both. The source records only pending review and states no approval or completed refund, so the draft is unsupported and the evidence gap stays open. This is an expected outcome on paper, not a completed test.
2. Input the proposed value for AIPROXYAPIENDPOINT. Compare whether it is a root address with no /v1 or /chat/completions suffix. A value carrying /v1 does not match the documented requirement and needs correction before rechecking. A value that is only a root address matches, and the remaining startup conditions stay unverified.
3. Input the source sentence "Return Grading evidence package RG-12 was uploaded on 2026-03-05 and contains 4 product images." Compare the returned evidence package identifier and image count against RG-12 and 4. If retrieval returns a different package identifier, record that as a mismatch and exclude it. Do not run cross-model retrieval to compare for missed recall, because datasets on different vector models cannot be associated together. If retrieval returns nothing, record the empty result and the missing evidence rather than asserting that a rule forces an empty list.

Across all three checks, the same subject, the same period, and the same field must be compared. Only mutually exclusive values in the same field count as a conflict. A missing value, or evidence that covers only an earlier stage, counts as unsupported or pending verification. A pending review does not establish that an operator lacks permission to act. Retrieval and form entry collect and return values; business verification, approval, database writes, and external synchronization require their own downstream implementation.

Related reading

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