What the documents look like
Warehouse picking, checking and packing handles order outbound work, and it does not handle supplier arrivals. The main data comes from the WMS, the order system, the electronic waybill platform and weighing equipment. The documents handled each day include wave picking lists, checking records, waybills and packaging consumable sheets. These materials update per order or per hour. For product and marketing copy drafting, the factual basis usually sits in the checking records and the packaging consumable sheets. A sentence in proposed copy can stand only when the corresponding material states the same subject, the same period and the same field. The category confidence is roughly as described, and any field not provided remains unknown.
The checking record is the document that carries the outbound verification state for an order. The packaging consumable sheet carries which material was used for that order. Neither document is a general authorization record, and neither one proves that a later business action occurred.
Constraints at the model connections step
From V4.17.0, AI Proxy is a required dependency. The App and any Admin service that reuses the App/Service configuration must both set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN. Missing either one causes startup validation to fail. AIPROXYAPIENDPOINT holds the AI Proxy service root address and does not append /v1. AIPROXYAPITOKEN must match the server-side configuration and cannot be empty; an application API key does not replace it.
Model capabilities are checked separately according to the input the application needs. Multimodal input, tool calling and thinking configuration are verified one by one. Application parameters are set in the application or node configuration, and it cannot be assumed that a same-named field on a compatible chat interface overrides the orchestration. Semantic recall and relevance scores do not carry permission or entity isolation. Before deciding whether a retrieval library and a vector model identifier can be associated together, the identifiers are checked first.
Hypothetical paper test case: a warehouse wave picking list W-0312-07 dated 12 March 2026 states "Order 8841, SKU white cotton T-shirt, picked quantity 2 pieces, checking status: checked." A packaging consumable sheet states "Order 8841 used a 30cm x 20cm paper bag." The proposed copy sentence is "The goods in this order have passed outbound checking." The judgement requires the same subject, the same period and the same field: if the checking status field for order 8841 in the checking record reads checked, the sentence is supported; if that field is not stated, or only picking-stage evidence exists, the sentence is unsupported and the gap is recorded.
Configuration and implementation checks
| Setting or check item | Hypothetical starting value for this trial | Reason for the choice and the object of comparison |
|---|---|---|
| AIPROXYAPIENDPOINT | Hypothetically set to http://aiproxy:3000 | The document requires the root address without /v1; compared against a form that includes /v1 |
| AIPROXYAPITOKEN | Hypothetically set to a non-empty value matching the server side | The document requires non-empty and consistent; compared against an empty value or an application key |
| Language model and index model | Hypothetically one of each enabled | The document requires at least one of each; compared against a state where either is not enabled |
| Model capability check | Hypothetically checked against the text input needed for copy drafting | Verified separately from multimodal and tool-calling capability |
| Application parameter location | Hypothetically set in the application or node configuration | Compared against a same-named field on the chat interface, to confirm no override of the orchestration |
| Retrieval library identifier check | Hypothetically the library and vector model identifiers are recorded | Compared against another vector model library; when identifiers differ, joint association does not proceed |
Three common errors
- Writing AIPROXYAPIENDPOINT as a full request address that includes /v1. Startup validation compares against the root address required by the document, and the comparison does not match.
- Using an application API key in place of AIPROXYAPITOKEN. The comparison against the server-side configuration shows inconsistency, and the non-empty check does not substitute for the consistency check.
- Treating a semantic relevance score as a permission or entity filter, and reading a similarity setting as excluding unrelated businesses or unauthorized records. Retrieval does not carry that responsibility.
How to verify the setup
- Input the wave picking list text and the proposed copy sentence from the hypothetical sample, then compare the value of the checking status field for order 8841 in the checking record. When the field reads checked, the sentence is supported; when the field is not stated, or only picking-stage evidence exists, record it as unsupported and keep the gap. Do not describe this as a denial or a conflict.
- Input the hypothetical AIPROXYAPIENDPOINT and AIPROXYAPITOKEN, then compare whether they are respectively a root address and a non-empty value consistent with the server side. When both conditions hold, record startup validation as passing; when either does not hold, record it as not passing and do not infer a specific error text.
- Input the hypothetical two retrieval library identifiers and vector model identifiers, then compare whether they are the same library and the same model. When the identifiers match, record that joint association is possible; when they differ, exclude the non-matching one and record the actual observation. Do not run a cross-model retrieval in order to compare missed recall, and do not extend the same-model requirement to any external system.