Omnichannel stock transfers: model connections for multilingual work

What the documents look like

Brand omnichannel stock transfers move goods between stores and warehouses. The working materials are transfer requests, stock snapshots, store shortage lists and transfer logistics notes. Source systems are ERP inventory, store POS, ecommerce orders and WMS. In a cross-border setting these materials frequently appear in more than one language, so retrieval has to cover multilingual text in the same search scope.

A transfer decision depends on the stock snapshot and the shortage list describing the same point in time. Material updates arrive daily or weekly, and the retrieval result only locates material. It does not grant transfer authority, and it does not verify that a counterparty is approved to receive or release goods.

Constraints at the model connections step

FastGPT connects to model providers through AI Proxy. App and Admin must both set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; a missing value in either one fails the startup check. The system needs at least one language model and one index model before it can be used normally.

A model ID is the value of the model field in the request body and is globally unique. A model mapping lets a FastGPT-side model name resolve to the actual upstream model, for example mapping gpt-4o-test to gpt-4o. A channel test sends a real request so the configured channel can be checked before use.

Retrieval relevance does not carry permission or entity separation. Libraries held in different vector models cannot be associated with each other; the identifier has to be checked first and non-matching libraries excluded. Semantic recall does not decide whether a material was authorized for the requester, and the multilingual search scope is a retrieval scope, not an access-control boundary.

Configuration and implementation checks

The table below covers settings and implementation checks that this page's documents describe. Values marked as a hypothetical starting point are chosen for a paper exercise and compared against the documented requirement or example; they are not official defaults.

Hypothetical paper sample. Source material, two sentences: "Transfer request T-2026-0312. Requesting party: East China supply chain group. Object: 200 coats moving from the Shanghai warehouse to the Hangzhou store. Status is pending review. Date: 12 March 2026." Draft sentence to be judged: "This transfer request has been approved and the transfer may be executed." The question is whether "has been approved" in the draft matches the source on the same party, the same period and the same field. The source records a status of pending review and carries no approval field. The approval conclusion in the draft is therefore unsupported and the evidence gap stays open.

Three common errors

1. Filling AIPROXY_API_ENDPOINT with a full request address that carries /v1. Input: an address ending in /v1. Comparison: the documented root-address requirement. Observable result: whether the startup check passes.
2. Treating a retrieval relevance score as a permission filter. Input: one multilingual retrieval request. Comparison: the authorization fields carried by the returned materials. Observable result: whether returned materials fall outside the requester's access scope.
3. Writing "pending review" as "approved". Input: the sample source text and the draft sentence. Comparison: same party, same period, same field. Observable result: whether the draft status matches the source status.

How to verify the setup

1. Take the sample source text and the draft sentence. Compare "pending review" against "has been approved" on the same party, the same period and the same field. Where the source carries no approval field, record the conclusion as unsupported and keep the evidence gap open.
2. Take a hypothetical AIPROXY_API_ENDPOINT value ending in /v1. Compare it against the documented root-address requirement and observe whether the startup check fails. Then take a value without /v1 and observe whether it passes.
3. Take one multilingual retrieval request. Compare the authorization fields of the returned materials against the requester's access scope and observe whether any record falls outside that scope. For the empty-result branch, use an explicit paper input of [] to check the downstream design; do not claim that any rule necessarily returns an empty list.

Related reading

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