Returned-goods inspections: model connections for multilingual work

What the documents look like

Returned-goods inspection handles reverse parcels, a different work direction from in-warehouse picking and packing. The main records come from the WMS returns module, platform after-sales, carrier returns, and the inspection bench. Daily documents include the return inbound note, inspection photos, the grade decision table, and the resale listing note. These materials update per item and per day. Scanning the returned parcel, judging item condition, and deciding resale or scrap are the most time-consuming steps.

Cross-border materials and multilingual retrieval place return inbound notes and grade decision tables in different languages into one retrieval scope, so inspectors can query by item condition. Errors in these records lead to resale of defective goods, refund disputes, and inventory contamination. The retrieval step locates relevant text blocks. It does not determine authorization status and does not replace the inspection conclusion.

Constraints at the model connections step

This page addresses only model connections and configuration for cross-border materials and multilingual retrieval in returned-goods inspection.

FastGPT requires at least one language model and one index model to work normally. AI Proxy is a required dependency from V4.17.0. App and Admin services that reuse App/Service configuration must both set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either item causes startup validation to fail. Old OpenAI environment variables or a single-model custom request address do not replace these two settings. AIPROXY_API_ENDPOINT is the AI Proxy service root address, for example http://aiproxy:3000, and must not have /v1 appended.

In language model fields, maxContext, maxResponse, and quoteMaxToken constrain context, reply, and quoted content respectively. Model capability is checked against the input the application needs. Multimodal input and tool calling are verified separately. When the vector store is Milvus, full-text retrieval uses BM25 over the modeldata_v2 single table, and MILVUS_LANGUAGE_IDENTIFIER selects the language identification engine (lingua by default, or whatlang). Other vector stores use MongoDB $text for full text.

Hypothetical drill sample: a returns warehouse receives one return inbound note reading "Return note RT-2041, item is a portable juicer, outer carton damaged, body has no scratches, accessories complete, arrival date 12 March 2026, quantity 2." Inspection photos for the same batch record the body appearance. The grade decision table marks "pending re-inspection." One drafted sentence reads "This return has completed inspection and is judged ready for resale listing." The subject, object, and status in that drafted sentence must be checked item by item against the materials above. The materials do not state completed inspection or a resale decision, so the drafted sentence remains unverified.

Configuration and implementation checks

The table below centers on the sample above. The values are hypothetical starting points for this paper exercise, used for comparison with documented field descriptions. They are not official defaults.

A relevance score is not a category, entity, or permission filter. Retrieval does not grant authorization, and a similarity setting does not exclude unrelated businesses or unauthorized records. Knowledge-reference limits use the documented unit. Same-model retrieval association, empty-reference output, and context limits keep their documented conditions.

Three common errors

1. Filling AIPROXY_API_ENDPOINT as a complete model request address with /v1 appended. Input: the hypothetically filled address carrying /v1. Comparison: the official requirement not to append /v1. Observable result: startup validation or the call path does not match the documented requirement.
2. Treating "completed inspection and judged ready for resale listing" in the drafted sentence as a fact already stated in the materials. Input: the RT-2041 return inbound note text and the grade decision table marking "pending re-inspection." Comparison: the status description in the drafted sentence. Observable result: the materials do not state completed inspection or a resale decision, and the sentence remains unverified.
3. Treating a retrieval hit as authorization or as an inspection conclusion. Input: one grade decision table text block returned by a multilingual retrieval. Comparison: the original records in the return inbound note and inspection photos. Observable result: the retrieval hit does not constitute completed inspection, listing approval, or refund processing.

How to verify the setup

1. Input the hypothetical RT-2041 return inbound note text and the drafted sentence. Compare whether the subject, object, and status in the drafted sentence correspond item by item to the source text. The drafted sentence is supported only when the source states completed inspection and a resale decision. The current source does not state these, so record the item as unverified.
2. Input the hypothetically filled AIPROXY_API_ENDPOINT and non-empty AIPROXY_API_TOKEN. Compare against the documented required fields and path requirement. When both items exist and the address has no /v1 appended, the condition is met. When either item is missing, record it as not met.
3. Input one multilingual retrieval for "portable juicer outer carton damaged." Compare whether the identifiers in the returned text blocks belong to the same return as the return inbound note, inspection photos, and grade decision table. When the returned identifiers match the materials, the condition is met. When the result is empty or identifiers do not match, record the actual observation without inferring a cause for missed recall.

Related reading

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