What the documents look like
Brand after-sales repair work turns on warranty responsibility and physical repair locations. The daily materials are warranty cards, repair work orders, inspection reports and exchange approvals. In cross-border and multilingual retrieval, one work order can exist in a Chinese version and in a target-market language version at the same time, so retrieval has to be checked against the language of the actual document and against the identifier of the indexing model. The brief states that this category updates per order and per day. That cadence does not by itself create batch synchronization, full synchronization, real-time triggers or transaction-triggered jobs, and this page does not assume any of them.
The materials are produced with the order. A warranty card records a purchase date and a warranty period. A work order records a work order number, a hand-in date, a fault description and a receiving location. An inspection report records a conclusion. An exchange approval records an approval state. These are separate documents with separate fields, and a retrieval result that returns one of them does not return the others. This distinction matters for multilingual work because a translated work order and its source work order share a number but not necessarily a language, and retrieval configuration has to be selected per document rather than per case.
Constraints at the model connections step
AI Proxy is a required dependency. The App service and any Admin service that reuses App/Service configuration must both set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN. Missing either item causes startup validation to fail. AIPROXYAPIENDPOINT takes the service root address and does not carry a /v1 suffix. Legacy OpenAI environment variables and single-model custom request addresses do not replace these two settings.
The system needs at least one language model and one indexing model. The indexing model fields include defaultToken for default text splitting and maxToken for the maximum token value. These belong to the indexing model entry, not to a chat request body, and changing a same-named field in a chat call does not override the indexing configuration.
Retrieval association between a document and a vector model is limited to documents inside the same vector store. This limit does not extend to external systems, and it does not mean a business registry or another outside service must use the same model. Semantic recall and relevance scoring do not perform permission or entity isolation. A similarity score is not a category filter, an entity filter or a permission filter, and a retrieval result is not authorization. Whether a repair falls inside warranty scope, whether an exchange is approved, and whether a refund may proceed are business determinations that need their own verification and approval steps downstream.
Configuration and implementation checks
The table below is tied to the hypothetical sample in the next section. Values marked as a hypothetical starting point are chosen for this paper exercise and compared against the documented field description; they are not official defaults and they were not tested.
| Check item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| AIPROXYAPIENDPOINT | http://aiproxy:3000 | Chosen as a root address without /v1; compare the entered value against the documented root-address form and check for an appended /v1 or /chat/completions |
| AIPROXYAPITOKEN | A non-empty administrator token | Chosen to satisfy the non-empty condition; compare against an empty value to confirm the field is populated and matches the server-side configuration |
| Indexing model | One multilingual indexing model enabled | The system needs at least one language model and one indexing model; compare against the enabled model list |
| Indexing model maxToken | 3000 | Chosen because the sample work order paragraph is short; compare against the indexing model field description rather than against a chat request field |
| Vector store identifier | One single vector store | Association is limited to the same vector store; compare identifiers and exclude documents whose identifier does not match |
| Retrieval input language | A Chinese work order and a target-language translation | Check against the language of the actual document; compare against a single-language assumption and record which document language was actually indexed |
Three common errors
- Entering a full request address with
/v1into AIPROXYAPIENDPOINT. The check takes the entered field value, compares it against the documented root-address form, and observes whether a/v1segment or a/chat/completionspath is present. A root address without those segments is the condition the document describes. - Treating a retrieval relevance score as a permission or entity filter. The check takes the retrieval configuration and the permission configuration, compares whether the two have been merged, and observes whether a design excludes unauthorized records by similarity score. Relevance scoring does not perform that exclusion.
- Writing an unsupported conclusion into a draft sentence. The check takes the draft sentence and the sample source text, compares the sentence against the fields the source text actually records, and observes whether an inspection conclusion or an exchange approval state has been asserted. If the source text does not record those, the sentence is unsupported and the evidence gap stays open.
How to verify the setup
Hypothetical paper test case, clearly labelled and not drawn from any real customer. A brand after-sales center receives a cross-border repair work order. The source text reads: "Work order number WX-2026-0417, hand-in date 17 April 2026, fault description screen flicker, warranty card states whole-unit warranty of 12 months, purchase date 1 June 2025, receiving location an overseas authorized point." The draft sentence under review reads: "This work order is confirmed to be within warranty scope and an exchange approval can be arranged directly." The source text does not record an inspection conclusion and does not record an exchange approval state.
- Input the sample source text and the draft sentence. Compare the draft sentence against the work order number, hand-in date, fault description, warranty period, purchase date and receiving location that the source text records. Observe whether the source text contains an inspection conclusion and an exchange approval state. If it does not, the draft sentence is unsupported and the evidence gap is recorded rather than resolved.
- Input the hypothetical AIPROXYAPIENDPOINT and AIPROXYAPITOKEN values. Compare the endpoint against the documented root-address form and the token against the non-empty condition. Observe whether either item is missing or malformed. A missing item matches the documented startup validation condition; this page does not claim that validation was run.
- Input the hypothetical retrieval request. Compare the returned results against the vector store identifier and the document language. Observe whether a design associates documents across different vector stores or compares across different models. Cross-model association is not available, so the check confirms identifiers first, excludes non-matching documents, and records what was actually observed. An empty result can be checked against a downstream design using an explicit paper input of
[]; this page does not assert that any particular query returns an empty set.
Across all three checks, the system error text, the model probe outcome and any automatic business action remain unknown unless the supplied documents state them. A pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination.