What the documents look like
Third-party overseas warehouses can serve several sales channels, and the seller allocates stock across them. Replenishment plans, inbound appointment records, inventory age tables and final-mile delivery notes are produced by different systems. ERP sales forecasts and warehouse management system stock figures may use different counting rules, and forwarder systems and carrier bills may carry different batch numbers. The multilingual retrieval task has to place several versions and several language fields of the same replenishment material into a searchable scope, then call a language model and an index model through the configured model connections. Material updates arrive weekly or daily, and this page stays with that stated rhythm.
Constraints at the model connections step
The system needs at least one language model and one index model before it works normally. From V4.17.0, AI Proxy is a required dependency, and both the App service and any Admin service that reuses App/Service configuration must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN. Missing either one causes startup validation to fail. Older OpenAI environment variables and single-model custom request addresses cannot replace these two settings. A model ID is the value of the model field in an interface request body and is globally unique. Model mapping can translate the model FastGPT requests into the upstream model actually used. Retrieval association is constrained by the vector store and index model identifier: only the same vector model can be associated together, and different models cannot. A relevance score does not carry permission or entity isolation duties, and retrieval is not authorization.
Hypothetical paper exercise sample: on 10 March 2025 a cross-border seller submits a replenishment plan check file. The source text reads "Overseas warehouse A, SKU-1001, available stock on 10 March is 420 units, with 80 units in transit; this replenishment batch plans to enter the warehouse with 200 units on 20 March, and the inbound appointment number has not been returned." A same-period inventory age table reads "SKU-1001, quantity aged over 90 days is 60 units." The draft sentence to be judged is "SKU-1001 has sufficient available stock on 10 March and no replenishment is needed." This sample is a paper exercise; no system call or approval occurred.
Configuration and implementation checks
The table below lists implementation check items as the starting point for this hypothetical exercise. The values are chosen for this paper exercise to compare against another set of values; they are not official defaults.
| Item | Hypothetical starting point for this exercise | Comparison method and reason |
|---|---|---|
| AIPROXYAPIENDPOINT | Enter the AI Proxy service root address without appending /v1 | Compare word by word with the documented root address requirement and confirm no path suffix |
| AIPROXYAPITOKEN | Enter a non-empty administrator token consistent with the server | Compare with the AI Proxy server configuration and confirm it is non-empty and consistent |
| Index model identifier | Record the model ID of the chosen index model | Compare with the vector model identifier used by the retrieval store; only the same model can be associated together |
| Model mapping | Record the correspondence between the model FastGPT requests and the upstream model | Compare item by item with the mapping in the channel configuration and confirm the requested model resolves to an upstream model |
| Language model capability | Check whether the capability required by this task's input is supported | Compare with the documented model capability fields; verify multimodal and tool calling separately |
| Retrieval scope | Record the materials and language fields included in this retrieval | Compare with the target retrieval scope that carries access rights; when missing, record the gap |
Three common errors
- Writing AIPROXYAPIENDPOINT as a full request address that includes /v1 or /chat/completions. The check takes that configured value as input, compares it with the documented root address requirement, and treats any path suffix as non-conforming.
- Using an application API key in place of AIPROXYAPITOKEN. The check takes the token value as input, compares it with the AI Proxy server administrator token, and treats a mismatch or an empty value as non-conforming.
- Treating a relevance score as a permission or entity filter, so that a low score is read as excluding unrelated business or unauthorized records. The check takes one retrieval request as input, compares the returned records with the authorized scope, and treats use of the score as an authorization basis as non-conforming.
How to verify the setup
- For the hypothetical draft sentence "SKU-1001 has sufficient available stock on 10 March and no replenishment is needed," take the replenishment plan source text and the inventory age table source text as input. Compare the values for the same subject, the same date and the same field: available stock 420 units, in-transit 80 units, quantity aged over 90 days 60 units, against the conclusion "no replenishment is needed." That conclusion is not supported, and the evidence gap must be recorded.
- For the hypothetical inbound appointment number "has not been returned," take that field as input and compare whether the material states an appointment number. When it is not stated, record it as pending verification, and do not judge it as an appointment refused or approved.
- For retrieval association, take the index model identifier and the retrieval store vector model identifier as input and compare whether the two are consistent. Only consistency allows association together; when they differ, exclude that store and record the actual observation, without running a cross-model retrieval to compare missed recall. An empty result can be handled by checking the downstream design against the explicit paper input [], without claiming that a fictional rule retrieval necessarily returns empty.