What the documents look like
Warehouse picking, checking and packing handles order outbound movement. It does not handle supplier arrivals. The working papers are wave picking lists, checking records, shipping labels and packaging consumable sheets. These are produced per order or at an hourly cadence. Picking errors, missed shipments and damaged parcels lead to after-sales compensation. The task on this page, cross-border reference material and multilingual retrieval, requires fields and multilingual notes from those papers to be passed to a model. Model connection and configuration therefore determine whether those materials can be read correctly.
Data sources named in the brief are the WMS, the order system, the electronic shipping label platform and weighing equipment. The model side only receives text or images that are submitted to it. It holds no write permission over those systems.
Two features of this category shape the model connection step. First, the same order appears across several papers with different field names, so a model must be tied to the right paper and the right field before any comparison is meaningful. Second, the material includes destination-country and label-language values, so language handling belongs to the retrieval and reading task rather than to a separate translation step.
Constraints at the model connections step
FastGPT connects to model providers through AI Proxy. In the App and Admin shared variables, AIPROXY_API_ENDPOINT is a required service root address and must not carry a /v1 suffix. AIPROXY_API_TOKEN is a required non-empty administrator token and must match the server-side configuration. An application API key does not substitute for it. The system needs at least one language model and one index model to work normally.
A model ID is the value of the model field in the request body and is globally unique. A model name is a display name and can be customized. A model channel determines the protocol type, and one model can be configured across different channels to achieve load balancing. A model mapping points a requested model to the model actually supplied, for example a request for gpt-4o-test being sent upstream as gpt-4o.
Vector association inside a retrieval library requires the same vector model. That requirement does not extend to business-registration websites or other external systems. A relevance score is not a category, entity or permission filter, and retrieval is not authorization. A rerank model must exist and be enabled under the configured setting. Image retrieval requires a corresponding image input and model capability. Model probing covers only enabled system models; disabled models do not participate in probing.
Assumed paper exercise sample. A warehouse wave picking list reads: "Wave W-0912, order SO-4471, item A quantity 2, destination country DE, label language DE." A checking record reads: "SO-4471 scanned 2 pieces, outer carton weight 1.8kg, consumable carton type M." A draft sentence awaiting judgment reads: "This order has completed checking and packing under a DE-language label and can be handed to transport." The source text states scanned quantity, weight and carton type. It does not state transport authorization and does not state that label printing has completed.
Configuration and implementation checks
The rows below are settings or implementation checks that the official documents describe. Where a value is marked as an assumed exercise starting point, it is used only for paper comparison and is not an official default.
| Configuration item | Documented setting or check | Assumed exercise starting point and comparison method |
|---|---|---|
AIPROXY_API_ENDPOINT | Required AI Proxy service root address, no /v1 suffix | Start with http://aiproxy:3000 and compare against the documented root-address form |
AIPROXY_API_TOKEN | Required non-empty administrator token, consistent with the server | Start with a non-empty string and compare against the non-empty condition |
| Model channel protocol type | Chosen by provider protocol; OpenAI is available for many providers | Start with OpenAI and compare against the protocol of the configured channel |
| Model ID and model mapping | Model ID is globally unique; mapping points the requested model to the actual model | Start with gpt-4o-test mapped to gpt-4o and compare against the model value in the request body |
| Model capability fields | Language, index and rerank models each check fields such as vision and toolChoice | Start with vision set to true and compare against whether the sample contains image input |
| Model probing scope | Probes only enabled system models; a first failure retries at most 3 times | No numeric value set; compare against the enabled condition |
Three common errors
- Writing
AIPROXY_API_ENDPOINTas a full request address that carries/v1or/chat/completions. The input is the configured value, the comparison is the documented root-address form, and the observable result is whether a path suffix is present. - Using an application API key as
AIPROXY_API_TOKEN. The input is the token source, the comparison is the condition of a non-empty administrator token, and the observable result is whether the value comes from the AI Proxy administrator configuration. - Extending the same-vector-model requirement for a retrieval library to external systems. The input is the vector library identifier, the comparison is the condition that only the same vector model can be associated, and the observable result is whether a same-model requirement is imposed on business-registration websites or other external systems.
How to verify the setup
- For the draft sentence "This order has completed checking and packing under a DE-language label and can be handed to transport," the input is the wave picking list and the checking record. The comparison is whether the source text states transport authorization and label printing completion. The observable result is that when neither is stated, the sentence is unsupported and the evidence gap must be recorded. A pending request is not an approval.
- For "SO-4471 scanned 2 pieces" in the sample, the input is the checking record. The comparison is the scanned quantity against item A quantity 2 on the wave picking list. The observable result is that the same subject and the same field carry the same value, so this field can be judged consistent.
- For the model probing configuration, the input is the list of enabled models. The comparison is the condition that disabled models do not participate in probing. The observable result is whether a disabled model appears in the probing record; if it does, the condition in the document is not met. An empty-result check can use the explicit paper input
[]to carry the downstream design, without claiming that a fictional rule search necessarily returns[].
Retrieval checks describe observations and failure branches. They do not guarantee a first hit, full recall or an empty result for an unrelated query. Where libraries use different vector models they cannot be associated together; the identifier is checked first and non-matching libraries are excluded, rather than running a cross-model search to compare missed recall.