What the documents look like
MCN operations manage long-term account compliance for multiple creators rather than serving a single brand campaign. The materials behind this page are creator signing contracts, training records, violation handling forms, and account schedules, updated daily and weekly. When a draft product or marketing sentence is prepared, the question is whether the existing materials support that sentence. The subject, object, status, and validity scope recorded in the materials determine what may be written.
Model connection and configuration sit at the point where material text and a proposed sentence are sent to a language model for comparison. Whether the configuration is correct determines whether the comparison result can be relied on. FastGPT uses AI Proxy to reach model providers, and AI Proxy also supplies load balancing, model logs, and usage dashboards for checking model calls.
Constraints at the model connections step
From V4.17.0, AI Proxy is a required dependency. The App service and any Admin service that reuses App or 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.
Model configuration is performed on the administrator model provider page. A channel requires a BaseUrl, not a full model request address. The system requires at least one language model and one index model to operate normally. Model status probing applies only to currently enabled system models; disabled models do not participate in probing and produce no history. A relevance score from retrieval does not carry permission or entity isolation duties, and retrieval itself is not authorization.
A model ID is the value of the model field in a request body and is globally unique. A model name is a display label that can be customized. A model channel represents a provider protocol, and one model may be configured across channels for load balancing. A custom request address and key can bypass the channel, and the documentation describes this as generally unnecessary and inconvenient to manage.
Configuration and implementation checks
Hypothetical drill sample: an MCN operations staff member needs to check one proposed sentence. The material text supplied for checking consists of three sentences. Creator A and the agency signed a live streaming cooperation agreement on 1 March 2026, valid until 28 February 2027. An attachment to the agreement states that Creator A may promote the beauty and personal care categories. A violation handling form dated 10 April 2026 records that Creator A received one platform warning for exaggerating efficacy during a live stream, with status handled. The proposed sentence to be judged is: "Creator A is the most trustworthy beauty creator on this site, and all product efficacy has been strictly verified."
| Configuration item | Setting or check on this page | Relation to the sample |
|---|---|---|
AIPROXY_API_ENDPOINT | Enter the AI Proxy service root address without appending /v1 | The copy comparison request is sent to this address |
AIPROXY_API_TOKEN | Enter a non-empty administrator token matching the server configuration | If absent, startup validation fails and comparison cannot proceed |
| Language model enabled | Confirm at least one language model is enabled under system models | Comparing the proposed sentence with material text depends on a language model |
| Index model enabled | Confirm at least one index model is enabled | Material retrieval depends on an index model |
| Model status probing | Applies only to enabled models; a 48-hour status chart can be viewed | Determines whether the model used for comparison is currently available |
Channel BaseUrl | Enter a BaseUrl, not a full model request address | A wrong channel configuration causes the comparison request to fail |
Three common errors
- Treating retrieval relevance as a permission filter. Input: the material library contains contracts for Creator A and other creators. Comparison: the relevance score returned by retrieval against access permission. Observable result: a high or low score does not indicate whether a record is visible to this operator; permission must be judged by a downstream implementation.
- Writing a pending or handled status as approved. Input: the violation handling form status field reads handled. Comparison: whether the proposed sentence claims that all product efficacy has been strictly verified. Observable result: the materials do not record efficacy verification, so that wording is unsupported.
- Treating a passing model test as a passing copy compliance check. Input: a channel test shows the model is available. Comparison: model availability against whether the proposed sentence is supported by materials. Observable result: model availability only shows the request can be sent and does not show the copy content holds.
How to verify the setup
- Input: the three material sentences above and the proposed sentence "Creator A is the most trustworthy beauty creator on this site, and all product efficacy has been strictly verified." Comparison: the phrases "most trustworthy" and "all product efficacy has been strictly verified" against the material text. Observable result: the materials record only the cooperation term, the permitted promotion categories, and one violation warning. They record no trustworthiness evaluation and no efficacy verification record, so the proposed sentence is unsupported and must be returned for revision.
- Input: the material sentence "Creator A may promote the beauty and personal care categories" and a hypothetical proposed sentence "Creator A may promote beauty, personal care, and food." Comparison: the category scope in the proposed sentence against the scope stated in the agreement attachment. Observable result: food is outside the stated scope, so the sentence is inconsistent with the materials and food must be removed or additional authorization material supplied.
- Input: a hypothetical retrieval request returns an empty result array
[]. Comparison: the empty result against the downstream node design. Observable result: on an empty result the downstream should follow a no-material branch and must not infer that materials do not exist or that the proposed sentence holds. The evidence gap should be recorded and referred for manual checking.