Marketplace category access: model connections for copy drafting

What the documents look like

Category access is decided before a listing goes live or a store joins the platform. The decision rests on the applicant entity and the authorization chain. The documents handled day to day are the business license, the trademark registration certificate, the brand authorization chain, and any industry license. These materials determine which promotional sentences may appear in copy and which must be removed. Materials are refreshed on an irregular schedule and collected again at each onboarding, so the currently valid material scope must be confirmed before drafting begins.

Copy drafting calls a language model to generate or rewrite text. Model connection and configuration determine whether the model receives the original material text, whether output can be constrained to the material scope, and whether generated results can be traced back. The model does not judge whether qualifications are genuine; it generates candidate sentences from the input text.

Constraints at the model connections step

Model access depends on AI Proxy. Both the App and any Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; a missing value in either causes startup validation to fail. AIPROXY_API_ENDPOINT holds the service root address without an appended /v1. Legacy OpenAI environment variables or a single-model custom request address cannot substitute for these two settings.

The system requires at least one language model and one index model to operate normally. The language model handles copy generation; the index model handles retrieval association across material fragments. The model ID is the value of the model field in the request body and is globally unique; the model name is for display only. Model mapping maps the model FastGPT requests to the upstream model actually used, so a request for gpt-4o-test reaches upstream as gpt-4o.

Hypothetical drill sample: an onboarding specialist holds business license text reading "internal drill code QUAL-A (not a unified social credit code), entity name Hangzhou Yunqi Example Technology Co., Ltd., business scope includes internet information services." Trademark registration certificate text reads "internal drill code TM-A (not a real registration number), registrant Hangzhou Yunqi Example Technology Co., Ltd., approved goods in Class 9, valid until March 1, 2030." Brand authorization chain text reads "Hangzhou Yunqi Example Technology Co., Ltd. authorizes the Hangzhou Yunqi Example Technology Co., Ltd. self-operated store to use the above trademark, authorization period January 1, 2025 to December 31, 2027." The draft sentence to be judged is "All goods sold in this store are direct from the brand owner; fake goods carry a tenfold penalty." This sentence involves three kinds of information: entity, authorization status, and compensation commitment. The materials state only authorization for the self-operated store to use the trademark, and do not state direct sales or compensation terms.

Configuration and implementation checks

The table below is a paper configuration check for this page. Trial values are not official defaults.

Three common errors

1. Writing AIPROXY_API_ENDPOINT as a full model request address with /v1 appended. The input is the address string entered; the comparison object is the root address format described in the documentation; the observable result is whether startup validation passes.
2. Using an application API key as AIPROXY_API_TOKEN. The input is the token value entered; the comparison object is the administrator token configured on the server; the observable result is whether the call returns an authentication error.
3. Reversing the model mapping direction and treating the upstream model name as the model ID FastGPT requests. The input is the mapping JSON; the comparison object is the key-value direction in the documented example; the observable result is whether channel testing reaches the target model.

How to verify the setup

1. Provide the business license text, trademark registration certificate text, brand authorization chain text, and the draft sentence "All goods sold in this store are direct from the brand owner; fake goods carry a tenfold penalty" together as model input. Compare whether the entity, authorization status, and compensation commitment in the draft sentence have corresponding fields in the materials. The materials state only authorization for the self-operated store to use the trademark and do not state direct sales or compensation terms, so the sentence is unsupported and must be deleted or changed to wording the materials support.
2. Separately set a paper draft reading "This authorization is valid long term" and compare it with the authorization chain period of January 1, 2025 to December 31, 2027. The draft expands the stated period and should be brought back within the original text range; the direct-sales and compensation claims in the original draft remain checked independently under item 1.
3. Use a paper empty reference [] to check the downstream handling design; whether an actual query is empty remains to be tested. Retrieval association does not perform permission or entity isolation, and a relevance score does not exclude unauthorized records. Libraries across different vector models cannot be associated together; first check the model identifier and exclude non-matching ones, record actual observations, and do not infer missed recall or redundant results.

Model probing takes effect only after it is enabled and only for enabled models; disabled models do not participate. A failed probe retries at most 3 times, and only 4 consecutive failures are judged abnormal. The alert webhook pushes on state edge changes and does not repeat while the model remains abnormal.

Related reading

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