What the documents look like
Brand-operated livestream rooms carry their own scripts, inventory, and after-sales work. The daily documents are the livestream script, the product explanation card, the shift schedule, and the livestream review. For cross-border sessions these documents exist in more than one language, and multilingual retrieval has to keep the language versions checkable against the source material. Updates run per session and per hour, so a translated card can drift from its original between shifts.
This page handles one step only: model connections and configuration. It covers the model channel, the model capabilities, and the retrieval conditions that a multilingual retrieval request depends on. It does not treat retrieval as an authorization decision.
Constraints at the model connections step
From V4.17.0, AI Proxy is a required dependency. Both the App and an Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; if either is missing, startup validation fails. AIPROXY_API_ENDPOINT holds the AI Proxy service root address and does not carry /v1. Older OpenAI environment variables or a single-model custom request address cannot replace these two settings.
The system needs at least one language model and one index model to work normally. Only one speech recognition model takes effect. The model ID is the value of the model field in the request body and is globally unique; the model name is a display label that can be customized. Channel priority ranges from 1 to 100, and a larger value is more likely to be selected first. A model can be configured across several channels for load balancing. Model mapping sends the model that FastGPT requests to the actual upstream model, for example mapping gpt-4o-test to gpt-4o.
Retrieval is a recall process. It does not perform authorization checks, and a relevance score is not an entity or permission filter. Libraries built on different vector models cannot be associated together. The correct handling is to check identifiers first, exclude the ones that do not match, and record what was actually observed.
Hypothetical paper exercise. A brand livestream team prepares multilingual explanation cards for a cross-border session. The source material reads: "The product for this session is Model A insulated cup. The Chinese explanation card states a capacity of 500 ml. The English explanation card states a capacity of 500 ml. Both versions state 'subject to the actual product'." The draft sentence to be judged is: "The English explanation card has confirmed a capacity of 500 ml and can be used directly for this session." The authorization boundary for this request: the explanation card is in a pending-review state, there is no approval record, and there is no multilingual consistency check record.
Configuration and implementation checks
| Configuration item | Setting or check action on this page | Relation to the exercise |
|---|---|---|
| AIPROXYAPIENDPOINT | Enter the AI Proxy service root address, without /v1 | Multilingual retrieval requests leave through this dependency |
| AIPROXYAPITOKEN | Enter a non-empty admin token matching the server side | If missing, startup validation fails and retrieval is unavailable |
| Language model and index model | Enable at least one of each | Multilingual text dialogue and indexing each need one |
| Model capability vision | Check against the input the application needs; do not assume default support | If an explanation card contains images, check this first |
| Model mapping | Map the model FastGPT requests to the actual upstream model | The model field of a multilingual request follows this mapping |
| Channel priority | For this exercise, start at 50 and compare with other values in the same channel | A trial starting value only, not an official default |
Three common errors
- Treating pending review as approved. In the exercise, the explanation card is in a pending-review state. The draft sentence writes "has confirmed." The input is the pending-review status record, the comparison object is an approval record, and the observable result is that the two do not match, so the sentence is unsupported.
- Using a retrieval hit in place of a consistency check. The input is the Chinese and English explanation cards, the comparison object is the capacity field and the "subject to the actual product" wording in both versions, and the observable result is that the two fields match while neither version carries a multilingual consistency check conclusion. Direct use cannot be decided from this.
- Treating a relevance score as a permission filter. The input is a score returned by one multilingual retrieval, the comparison object is the document authorization status field, and the observable result is that the score carries no authorization information, so unauthorized records cannot be excluded on that basis.
How to verify the setup
- Input the source material and the draft sentence, compare "has confirmed" in the draft against the pending-review state in the source, and observe whether the two disagree. When they disagree, the sentence is unsupported and the review gap remains open.
- Input the fields of the Chinese and English explanation cards, compare the capacity figures and the disclaimer wording, and observe whether the fields match while no consistency check record exists. Record this as pending verification and do not decide that direct use is allowed.
- Input one retrieval result, compare the document identifiers in the result against the authorization field, and observe whether non-matching identifiers are excluded. When the authorization field is missing, record the gap. For the empty-result branch, a clearly labelled paper input of
[]can be used to check the downstream design; do not claim that a fictional rule will always return[].