What the documents look like
Brand dealer ordering events run on offline dealer reservations. They do not directly manage platform storefronts. The working material comes from the ordering event system, historical sales, dealer inventory and financial credit approval. Day-to-day documents include the ordering handbook, the ordering contract, the sample list and the credit approval record. These materials update by quarter or by year, and a cross-border setting adds language versions that must stay aligned with the original.
Model connection and configuration at this step exists to let multilingual retrieval read those documents correctly and to keep document status separate from authorization status. A retrieval hit on a handbook is not an approval. A sample list marked as pending is not a confirmed list. The configuration work is to make the language model and the index model available, to keep the AI Proxy connection valid, and to record which vector model each library actually uses.
Constraints at the model connections step
Multilingual retrieval needs a language model and an index model working together. Per the official documents, the system requires at least one language model and one index model before it can operate normally. Models connect through AI Proxy. The App and Admin services must both set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either one causes startup validation to fail. Older OpenAI environment variables or a single-model custom request address do not replace these two settings.
On the retrieval side, when the vector store is Milvus, full-text retrieval runs through BM25 on the modeldata_v2 single table, and MILVUS_LANGUAGE_IDENTIFIER selects the language identification engine, with lingua as the documented default and whatlang as the alternative. With other vector stores, full-text retrieval runs through MongoDB $text. Libraries built on different vector models cannot be associated together. The correct action is to check the model identifier and exclude the ones that do not match. A relevance score does not act as a permission filter or an entity isolation filter, and retrieval itself is not authorization.
Hypothetical paper exercise sample, marked as such and not drawn from any real record: a brand channel department holds a Chinese ordering handbook whose original text reads, "This handbook applies to the Spring 2026 ordering event. Dealers may submit reservation quantities from March 1 to March 15. The credit limit is subject to the finance approval result." A separate English sample list reads, "Sample list for Spring 2026, subject to final confirmation." One draft sentence awaiting review reads, "The dealer has completed credit approval and may submit reservation quantities under the handbook." The draft sentence merges a document statement with an approval status, and it does not carry the "subject to final confirmation" limitation from the sample list.
Configuration and implementation checks
The table below is limited to settings the official documents describe, or to implementation checks tied to this page's documents. The starting values are hypothetical exercise starting points, not official defaults, and each one states why it was chosen and what it is compared against.
| Item | Hypothetical exercise starting point | Reason and comparison method |
|---|---|---|
AIPROXY_API_ENDPOINT | http://aiproxy:3000 | Documented example form for the AI Proxy root address, with no /v1 appended; compare against the deployed AI Proxy address |
AIPROXY_API_TOKEN | A non-empty administrator token entered for the exercise | Documented requirement that the value is non-empty and matches the AI Proxy server; compare against the AI Proxy configuration |
| Language model enabled | One multilingual language model enabled for the exercise | Documented requirement of at least one language model; compare against the language coverage the handbook and sample list need |
| Index model enabled | One index model enabled for the exercise | Documented requirement of at least one index model; compare against the document languages being indexed |
MILVUS_LANGUAGE_IDENTIFIER | lingua for the exercise | Documented default value used for BM25 language identification; compare against the actual document languages |
| Model mapping | A request model mapped to an upstream model for the exercise | Documented behavior that the model ID in FastGPT can differ from the upstream model; compare against the channel configuration |
Three common errors
- Treating a retrieval result as an authorization basis. In the sample, the draft sentence states that the dealer has completed credit approval, while the handbook original only states that the credit limit is subject to the finance approval result. The approval status is not stated in the original. A retrieval hit on the handbook does not support the claim that approval is complete, so the sentence remains unsupported.
- Associating libraries built on different vector models. If the Chinese handbook and the English sample list are stored in libraries using different vector models, they cannot be associated for joint retrieval. The model identifier must be checked first and non-matching libraries excluded. The observation is recorded as-is, without inferring missed recall or redundant results.
- Treating a pending status as confirmed. The English sample list states "subject to final confirmation," and the draft sentence does not carry that limitation. Pending is not confirmed. This is an evidence gap, and no confirmation clause may be supplied to fill it.
How to verify the setup
- Input the Chinese handbook original and the draft sentence from the hypothetical sample. Compare the draft phrase "has completed credit approval" against the original phrase "subject to the finance approval result." Observe whether the draft sentence stays within what the original supports. If the draft claims approval is complete while the original does not state it, the result is unsupported.
- Input the English sample list original and the draft sentence from the hypothetical sample. Compare "subject to final confirmation" against the draft wording. Observe whether the draft preserves the pending-confirmation limitation. If the draft does not preserve it, the result is an evidence gap.
- Input the two sample documents and a retrieval request. Compare the vector model identifiers recorded for each document's library. Observe whether they belong to the same model. If the identifiers differ, exclude joint association, record each retrieval observation separately, and do not run a cross-model comparison.
A retrieval check observes and records failure branches. It does not guarantee a first-hit result, full recall, or an empty result for an unrelated query. An empty-result case can be checked against a downstream design using a clearly stated paper input of an empty list, without claiming that a constructed rule returns an empty list. The optional paths in the overall document flow are not executed on every request: image retrieval requires a corresponding image input and model capability, and reranking requires a model that is enabled under the settings. When a document field is not provided, it remains unknown. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination.