What the documents look like
Tmall flagship store daily operations centre on brand authorisation, flagship store identity, and Tmall points compensation rules. The working materials are product listing pages, brand authorisation letters, delivery commitment records, and after-sales work orders. These materials are updated daily and per order. The cross-border materials and multilingual retrieval task has to handle foreign-language names, authorisation scopes, and commitment clauses inside those materials, so the model needs the relevant language recognition and comprehension ability. A daily and per-order update rhythm means retrieval results change as documents change, and model connection and configuration must allow re-indexing on demand. At the configuration layer, only enabled system models are checked; models that are not enabled do not take part in probing or retrieval.
Two industry facts shape this step. First, brand authorisation letters define the authorised party, the authorised scope, and the validity period, so any retrieval or drafting that touches authorisation must stay inside the written scope. Second, delivery commitment records carry per-order deadlines, and the model connection must not be treated as proof that a commitment was met. The materials are the source; the model connection only makes them searchable and readable.
Constraints at the model connections step
AI Proxy is a required dependency. Both the App and the Admin service must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either one causes startup validation to fail. Legacy OpenAI environment variables or a single-model custom request address cannot replace these two settings. AIPROXY_API_ENDPOINT takes the service root address and must not append /v1. AIPROXY_API_TOKEN is an administrator token and cannot be replaced by an application API key.
The system needs at least one language model and one index model to work normally. Model ID is globally unique; the model name can be customised; a model channel corresponds to a provider protocol; one model can be configured across different channels for load balancing. A custom request address and key bypass the channel and are not recommended. System model probing covers only enabled system models; disabled models do not take part in probing and produce no history records. Retrieval libraries can be associated together only when they use the same vector model; libraries with different vector models cannot be associated together. Check the identifier and exclude non-matching libraries rather than running a cross-model retrieval comparison.
Hypothetical paper exercise sample: a Tmall flagship store operator submits a scanned brand authorisation letter. The source text contains three sentences: the authorising party is A Brand Management Co., Ltd.; the authorised party is B E-Commerce Co., Ltd.; the authorisation period runs from 1 January 2026 to 31 December 2026; the authorised scope is the sale of A brand Class 25 apparel goods on the Tmall platform. A proposed draft sentence reads "Company B has obtained full-category authorisation for brand A." The pending request reads "Please confirm whether this authorisation letter supports marking full-category authorisation on the product listing page." This sample is for paper checking only and does not involve a real customer or a legal identity number.
Configuration and implementation checks
The table below lists official documented settings or implementation checks relevant to this page's feature and primary task. The hypothetical starting values are for this paper check only and are not official defaults.
| Configuration item | Hypothetical starting value for this exercise | Check method |
|---|---|---|
AIPROXY_API_ENDPOINT | Assume http://aiproxy:3000 | Check that it is the service root address with no /v1 appended |
AIPROXY_API_TOKEN | Assume a non-empty administrator token | Check that it matches the server configuration and is non-empty |
| Language model | Assume one language model is enabled | Check the enabled state; disabled models do not take part in probing |
| Index model | Assume one index model is enabled | Check the enabled state; this satisfies the system minimum |
| Model ID | Assume it matches the model in the channel | Check that it is globally unique and distinct from the model name |
| Channel priority | Assume 50 | Check the range 1 to 100; a larger value is selected more readily |
The starting values are chosen so the paper exercise has concrete inputs to compare against. Each is compared with the documented condition for that setting, not with a remembered default.
Three common errors
- Filling
AIPROXY_API_ENDPOINTwith a complete request address and appending/v1. Check: input the configured value, compare whether it consists only of protocol, host, and optional port as a root address; the presence of/v1or/chat/completionsmeans the condition is not met. - Using an application API key in place of
AIPROXY_API_TOKEN. Check: input the token, compare whether it is an administrator token and non-empty; an application API key means the condition is not met. - Including a disabled model in the retrieval or probing scope. Check: input the model enabled state, compare whether the probing scope contains only enabled system models; a disabled model appearing in a probe record means the condition is not met.
How to verify the setup
- For the sample draft sentence "Company B has obtained full-category authorisation for brand A", input the three source sentences of the authorisation letter and compare the authorised scope field against Class 25 apparel goods. If the draft sentence states full-category authorisation while the source states only Class 25 apparel, the draft sentence is not supported.
- For the sample pending request "Please confirm whether this authorisation letter supports marking full-category authorisation on the product listing page", input the authorisation letter source text and the request, and compare whether the full-category authorisation the request needs appears in the source scope. If the source states only Class 25 apparel, the request remains to be verified and is not judged as authorised.
- For the retrieval scope, input the model enabled state and the vector model identifier, and compare whether the libraries taking part in retrieval use the same vector model. If the identifiers differ, exclude that library and record the actual observation without inferring missed recall or redundant results.
A relevance score is not a category, entity, or permission filter, and retrieval is not authorisation. A pending request is not approval, and an unsigned termination is not a completed termination. Where the source document does not state a field, that field remains unknown.