Overseas stock replenishment: model connections for copy drafting

What the documents look like

Third-party overseas warehouses can serve multiple sales channels, and the seller allocates stock across them. The working papers for this category are replenishment plans, warehouse inbound appointments, inventory age tables, and last-mile delivery notes. They arrive from the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems, and carrier bills. Copy drafting for this category quotes quantities, statuses, and date ranges that appear in those papers, and the papers themselves update weekly or daily.

Two features of this material shape the drafting task. First, a replenishment plan records a planned inbound date and a planned quantity; it does not record whether the goods actually arrived. Second, an inventory age table records how long stock has been held, which is a different fact from what is scheduled to arrive. A draft sentence that converts a planned date and quantity into a completed arrival claim moves beyond what the source paper states. The model connection step matters here because the draft is produced through a configured language model, and the model can only work from the material supplied to it.

Constraints at the model connections step

Model access depends on AI Proxy. From V4.17.0, AI Proxy is a required dependency. Both the App service and an Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; a missing value for either causes startup validation to fail. Older OpenAI environment variables or a single-model custom request address do not replace these two settings. The endpoint is the AI Proxy service root address, for example http://aiproxy:3000, and it must not carry a /v1 suffix.

Models are grouped by type: language model, index model, rerank model, speech synthesis, and speech recognition. The system needs at least one language model and one index model to operate normally. A model ID is the value of the model field in the request body and is globally unique; the model name is for display and can be customized. A model can be configured across more than one channel to spread load. Channel tests and the page-level model test send an actual request.

Model status probing covers only system models that are currently enabled. A probe retries after a first failure, up to three retries, so a single round makes at most four attempts; one success is enough to count as valid, and only four failed attempts produce an abnormal result. Health states are normal (green, response latency at or below 30s), high latency (yellow, response latency above 30s), and abnormal (red, four consecutive request errors). The probe interval accepts whole minutes from 5 to 60. Channel priority has a range of 1 to 100, and a larger value is selected more readily.

A hypothetical paper exercise sample follows. The source paper text reads: "This replenishment plan covers the German overseas warehouse, the planned inbound date is 10 March 2026, and the quantity is 600 units. This warehouse currently has 3 SKUs with an inventory age over 180 days." The draft sentence to be judged reads: "This replenishment batch will be fully received into the warehouse on 10 March, 600 units." The paper gives a planned inbound date and a quantity. It does not state that receiving was completed. The draft sentence is not supported by the supplied text.

Configuration and implementation checks

These values are starting points for this paper exercise. They are not official defaults.

Three common errors

1. Writing a planned inbound date and quantity from the source paper as a completed arrival fact. Check: take the source paper text as input, compare the draft sentence for completion wording such as "will be fully received," and observe whether the two agree. In the exercise sample, the paper states a planned date and a quantity, and the draft asserts completed receiving, so the draft is not supported.
2. Replacing AIPROXY_API_ENDPOINT and AIPROXY_API_TOKEN with older OpenAI environment variables. Check: take the deployment environment variable list as input, compare whether both required settings are present and non-empty, and observe whether startup validation passes. Whether a specific deployment passes is not observed here.
3. Describing the probe abnormal condition as a single failed request. Check: take the probe configuration description as input, compare the retry count and the final judgment condition, and observe whether the description matches four failed attempts. A description stating one failure is enough does not match the documented condition.

How to verify the setup

• Verification one. Input the hypothetical source paper text and the draft sentence "This replenishment batch will be fully received into the warehouse on 10 March, 600 units." Compare whether the draft goes beyond the planned date and quantity that the paper states. If the draft contains completion wording, it is not supported by the supplied text, and the evidence gap remains open. The paper does not record actual receiving status, and no observation of receiving is available.
• Verification two. Input the deployment environment variable list. Compare whether AIPROXY_API_ENDPOINT is a root address without a /v1 suffix and whether AIPROXY_API_TOKEN is present and non-empty. Both settings being present and matching the official example form satisfies the documented startup validation condition. Whether a real deployment actually starts is not tested here.
• Verification three. Input the values entered on the model probe configuration page. Compare whether the probe interval is a whole number from 5 to 60 and whether the channel priority is within 1 to 100. A value inside the official range satisfies the configuration condition. The actual probe result is not observed, and no probe outcome is claimed.

Retrieval and relevance scores do not carry permission or entity isolation. Libraries built on different vector models cannot be associated together; check the identifier first and exclude those that do not match, then record what is actually observed. A relevance score is not a category, entity, or permission filter, and a similarity setting does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization. An empty-reference output and context limits keep their documented conditions. Forms collect values; business verification, approval, database writes, and external synchronization require their own downstream implementation.

Related reading

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