Amazon FBM fulfillment: model connections for multilingual work

What the documents look like

FBM sellers fulfill orders themselves and set their own return addresses, so the paperwork does not pass through platform warehouse inspection. The working materials are order pick sheets, shipment confirmation sheets, return address settings, and logistics tracking tables. These items update per order and per day, and their language follows the buyer's marketplace, so customer service email commonly mixes several languages. Cross-border material and multilingual retrieval has to answer questions and cite these materials. A retrieval result only presents document content; it does not authorize an order, a return, or a carrier choice. Update frequency does not change how a model is connected. The connection configuration has to match the languages and field structure of the materials.

Constraints at the model connections step

FastGPT connects to model providers through AI Proxy. Since V4.17.0, AI Proxy is a required dependency. The App and any Admin service that reuses App/Service configuration must both set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN. A missing item causes startup validation to fail. Legacy 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 name and can be customized. A model channel corresponds to a provider protocol, and one model can be configured across different channels for load balancing. A custom request address and key can bypass the model channel, which the documentation marks as not recommended. Using the same vector model across a retrieval library is a precondition for association inside that library, and this requirement does not extend to business registration sites or other external systems. Semantic recall and relevance do not carry permission or entity isolation duties, and retrieval itself is not authorization.

Configuration and implementation checks

Hypothetical paper exercise sample: an order pick sheet reads "Order number FBM-2026-0417, marketplace US, pick item A quantity two, carrier pending, return address field blank"; a shipment confirmation sheet reads "Order number FBM-2026-0417, shipment date 2026-04-18, tracking number not filled." The draft sentence to be judged is "This order has completed shipment and a tracking number can be provided to the buyer." This sample is for paper checking only and does not represent a real order or a real customer.

This hypothetical exercise starting value for channel priority is 50. The reason is that an intermediate value makes switching easier to observe when several channels coexist, and the comparison object is the priority value of other available channels.

Three common errors

1. Writing AIPROXY_API_ENDPOINT as a full request address that includes /v1, or leaving AIPROXY_API_TOKEN empty and substituting an application API key. The check reads the environment variable text and compares whether the address matches the service root format and whether the token is non-empty.
2. Treating the model ID and the model name as interchangeable. The check compares the model field in the channel configuration against the display name and asks whether the two are referenced as one identifier.
3. Treating a retrieval hit as an authorization credential. The check compares the draft sentence against the sample facts that the return address field is blank and the tracking number is not filled, and asks whether the draft claims completed shipment.

How to verify the setup

1. Input the sample pick sheet and shipment confirmation sheet text plus the draft sentence. Compare the claim "has completed shipment and can provide a tracking number" against the field state "tracking number not filled" in the source. The source does not record a tracking number, so the draft claim is unsupported and should be recorded as pending verification. A contradiction can be judged only when the same field carries mutually exclusive values in the source.
2. Before retrieval, check the vector model identifier of the candidate library and exclude those that do not match. Then query with the sample order number and marketplace, and check each cited order and marketplace item by item; record missing fields as pending verification. For empty-result handling, a separate paper input of [] can check the downstream design. Do not presume a query must return nothing, and do not perform cross-model joint association.
3. Input a request asking about the return address, and compare it against the actual state of the return address field in the materials. That field is blank, so the request is in a pending-supplement state. A pending review does not demonstrate a lack of operating permission, and it does not constitute an effective return address. The authorization boundary is this: retrieval and forms collect values, while business verification, approval, database writes, and external synchronization need separately connected system implementations.

Model status probing covers only enabled models. After a first failure, the system retries at most 3 times, and 4 failed attempts in total are judged abnormal; a latency above 30s is high latency. The alert Webhook pushes on edge changes of status, with event types model_status_error and model_status_recovered. System errors and automatic business actions that the documentation does not describe are not asserted here.

Related reading

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