What the documents look like
Multi-warehouse routing happens at the order routing stage and answers which warehouse ships an order. The working materials come from WMS, OMS, ERP inventory and platform inventory interfaces. Day-to-day documents include split-warehouse rules, inventory synchronization tables, transfer orders and shortage alerts. These materials update on a daily or hourly rhythm, and the sellable quantity, priority and transfer status of the same SKU across warehouses appear together.
Cross-border materials and multilingual retrieval deal with the multilingual versions of these materials: the Chinese source and translated text of split-warehouse rules, the field descriptions of inventory synchronization tables, and the handling wording of shortage alerts. Retrieval results serve only to locate relevant materials. They do not constitute shipping authorization, inventory deduction or transfer approval.
Two industry facts shape this page. First, the routing decision determines which warehouse ships, so the rules and their translations carry operational weight. Second, the same SKU can hold different sellable quantities in different warehouses, so a translated sentence that merges or reassigns warehouse scope changes the meaning of the rule.
Constraints at the model connections step
Model connections are handled by AI Proxy. From V4.17.0, AI Proxy is a required dependency. Both App and Admin services that reuse App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN. Missing either item causes startup validation to fail. AIPROXY_API_ENDPOINT holds the service root address and does not append /v1. 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. Model capabilities are checked separately according to the input an application needs: multimodal input, tool calling and thinking configuration are each verified. Application parameters are set in the application or node configuration. Assuming that same-named fields of a compatible chat interface override the orchestration is not supported.
The index model fields defaultToken and maxToken are the default text-splitting token and the maximum token. These are model configuration fields, not retrieval permission or entity filter conditions. When the vector store is Milvus, full-text retrieval uses BM25; with other vector stores, full-text retrieval uses MongoDB $text. Libraries across different vector models cannot be associated together. Check the identifier first and exclude those that do not match.
Configuration and implementation checks
The table below centers on a hypothetical drill sample and lists only officially documented settings or implementation checks relevant to this page's primary task. The values are starting points for this hypothetical exercise, used to explain the comparison method. They are not official defaults.
| Configuration item | Hypothetical starting point | Check and comparison method |
|---|---|---|
| AIPROXYAPIENDPOINT | http://aiproxy:3000 | Confirm no /v1 or /chat/completions is appended; compare character by character with the actual deployment root address |
| AIPROXYAPITOKEN | Non-empty administrator token | Must match the server-side configuration; an empty value or an application API Key does not satisfy this |
| Language model | One enabled model | Exists separately from the index model; enabled status can be checked |
| Index model | One enabled model | defaultToken and maxToken filled according to the field descriptions |
| Model probe | Probe interval of 30 minutes | This value falls within the documented 5–60 minute integer range and serves only as this exercise's starting point |
| Retrieval scope | Target knowledge base | Compared with the authorized scope; retrieval does not carry authorization determination |
Hypothetical drill sample: a cross-border seller holds a Chinese "Split-Warehouse Rules" version 3, whose body states: "The East China warehouse takes priority for orders with East China delivery addresses, and the South China warehouse handles orders with South China delivery addresses; the inventory of the two warehouses is calculated independently." A separate "Inventory Synchronization Table" field description states: "Sellable inventory updates daily." The draft sentence to be assessed is: "The East China warehouse can handle all orders, and South China warehouse inventory is merged into East China for unified calculation." Independent inventory and merged unified calculation are conflicting values on the same field. The rule states priority for East China but provides no rule for all orders, so the claim covering all orders remains unsupported. Record the two findings separately.
Three common errors
- Filling
AIPROXY_API_ENDPOINTwith a complete model request address that appends/v1. Check method: the input is the actual deployment root address and the variable value; compare whether the two match character by character. Only a match satisfies the condition. - Using an application API Key as
AIPROXY_API_TOKEN. Check method: the input is the entered token and the server-side administrator token; compare whether they match and whether the value is non-empty. Only a match on a non-empty value satisfies the condition. - Treating a retrieval hit as shipping authorization or transfer approval. Check method: the input is the draft sentence and the original text of "Split-Warehouse Rules" version 3; compare the values on the same split-warehouse field for the same subject and the same period. The draft sentence's value is mutually exclusive with the original, so the finding is unsupported.
How to verify the setup
- The input is the original text of "Split-Warehouse Rules" version 3 and the draft sentence "The East China warehouse can handle all orders, and South China warehouse inventory is merged into East China for unified calculation." Compare the two on the same split-warehouse field. The original states the two warehouses calculate inventory independently; the draft states merged unified calculation. The values are mutually exclusive, so the finding is unsupported and cannot proceed to handover execution.
- The input is a paper empty result
[]and the downstream handover design. Check whether the downstream retains a gap record when it receives an empty result, whether it records that this retrieval returned nothing, and whether it leaves permission status as a separate check. The expectation is that a gap is recorded and no shipping or transfer action is produced. - A separate paper request states "the transfer order is pending review, please execute directly." Compare the pending review status with the execution claim. The sample contains no approval or execution record, so the two remain separately unverified. This paper check does not mean a real transfer was executed or blocked.
The model probe retries automatically up to 3 times after an initial failure, and only 4 failed attempts in total mark the model as abnormal. When automatic probing is not enabled, the page shows an empty state. The alert Webhook fires on edge changes of status, with each model sending an independent notification. These are probe behaviors and do not change the retrieval authorization boundary.