Parcel labels and handoffs: model connections for multilingual work

What the documents look like

Handoff in this category happens at the point where parcels leave the warehouse, and the central fact is carrier scan responsibility. The document set is made up of electronic waybills, handoff manifests, pickup scan records and exception parcel registers, drawn from courier systems, WMS shipment records and carrier billing. Updates arrive per batch and per day.

For cross-border work the same handoff batch may carry waybills in Chinese, waybills in a foreign language, and bilingual handoff manifests at the same time. Multilingual retrieval across that mixture depends on stable model connections and configuration: the system needs at least one language model and one index model to operate normally, and only one speech recognition model takes effect, so one is enough. Speech recognition is not the primary need here; the language and index models are.

Constraints at the model connections step

From V4.17.0 onward, AI Proxy is a required dependency. Both the App and any Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either one causes startup validation to fail. Older OpenAI environment variables and single-model custom request addresses do not substitute for these two settings. AIPROXY_API_ENDPOINT takes the AI Proxy service root address with no /v1 appended.

Model capability is checked against the input the application actually needs, with multimodal input and tool calling verified separately. Application parameters belong in the application or node configuration; the same-named fields on a compatible chat interface are not assumed to override orchestration. Retrieval is a document-finding action and does not carry authorization. A retrieval result is not a basis for release, for treating a handoff as complete, or for a refund.

Hypothetical paper exercise: at one warehouse, a cross-border parcel batch dated 10 March 2026 has a handoff manifest whose source text reads "This batch contains 48 items in total, of which 3 items have foreign-language recipient addresses on the waybill; the pickup scan record shows 45 items scanned and 3 items awaiting scanning." The draft sentence under review reads "All parcels in this batch have completed pickup scanning and can be submitted for shipment." The draft conflicts with the source: the source states 3 items awaiting scanning, while the draft claims all are complete. This sample is for paper comparison only and does not represent any real business operation.

Configuration and implementation checks

The table below stays with this page's sample. Values are hypothetical starting points for this exercise, with the reason for the choice and the comparison object stated.

Three common errors

1. Filling AIPROXY_API_ENDPOINT with a full request address that carries /v1. The input is the configured value; the comparison object is the documented instruction not to append /v1; the observable result is whether startup validation passes. This is a configuration error and involves no business execution.
2. Treating the 3 items awaiting scanning as already scanned. The input is the sample source text stating 3 items awaiting scanning; the comparison object is the draft sentence claiming all parcels completed pickup scanning; the observable result is that the two take mutually exclusive values on the same field, so the draft is unsupported.
3. Using a retrieval result in place of an authorization decision. The input is a manifest fragment returned by multilingual retrieval; the comparison object is whether an explicit release or handoff-completion record exists; the observable result is that a retrieval hit does not constitute authorized execution, and where the authorization record is absent the gap remains.

How to verify the setup

1. The input is the hypothetical sample source text "This batch contains 48 items in total, of which 3 items have foreign-language recipient addresses on the waybill; the pickup scan record shows 45 items scanned and 3 items awaiting scanning", and the comparison object is the draft sentence "All parcels in this batch have completed pickup scanning and can be submitted for shipment." The expected observation is that the draft and the source take mutually exclusive values on the same field, so the draft is judged unsupported. No live test is claimed.
2. The input is the hypothetical configuration of AIPROXY_API_ENDPOINT set to http://aiproxy:3000, and the comparison object is the documented instruction not to append /v1. The expected observation is that the configured value carries no /v1 and meets the startup validation condition; a value with /v1 does not match the documentation. No live test is claimed.
3. The input is the 3 items awaiting scanning in the hypothetical sample, and the comparison object is whether a corresponding pickup scan completion record exists. The expected observation is that if no scan completion record for those 3 items falls within the retrieval scope, the evidence gap is recorded; if retrieval returns an empty result, the downstream design is checked against an explicit paper input of [], without asserting that a hypothetical rule makes retrieval return empty. No live test is claimed.

Related reading

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