Dealer ordering events: model connections for product questions

What the documents look like

Dealer ordering events serve offline dealer reservations and do not manage platform storefronts. The working materials for this category are the ordering handbook, the ordering contract, the sample list and the credit approval record. Historical sales, dealer inventory and financial credit support the same event cycle. Product information and selection questions must answer whether a given item is orderable in a given ordering season, which handbook clause applies, and whether the sample list contains that model. Materials refresh by season or by year, so a configuration step only needs the imported documents to remain retrievable and answerable once a language model and an index model are enabled.

The practical consequence for model connections is narrow. A dealer question such as whether a model is orderable depends on text that already sits in the knowledge base. The model layer supplies retrieval and generation, not commercial authorization. A retrieved clause can support a statement about the handbook. It cannot convert a pending credit request into an approved credit line, and it cannot convert a sample arrival into an ordering right.

Constraints at the model connections step

From V4.17.0, AI Proxy is a required dependency. Both the App service and an Admin service that reuses App and Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN. Missing either item causes startup validation to fail. Older 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 index model to operate normally.

Inside a model channel, the proxy address field takes a BaseUrl, not a complete model request address. Model mapping translates the model FastGPT requests into the upstream model actually called; the documented example maps gpt-4o-test to gpt-4o. The relation between a retrieval dataset and its vector model belongs to the knowledge base side and does not change with this page's model connection work. A relevance score is not a category, entity or permission filter, and retrieval is not authorization.

Configuration and implementation checks

The following hypothetical paper exercise is labelled as such. It records no real customer, no real approval and no completed business action.

Hypothetical exercise sample: a brand channel department prepares for an autumn 2026 ordering event. It imports Chapter 3 of the Autumn 2026 Ordering Handbook, Sample List A and one credit approval record into a knowledge base for product information and selection questions. Chapter 3 states: "Model A vacuum flask is orderable this season, minimum order 500 units." Sample List A lists the Model A vacuum flask with a status column reading "sample arrived." The credit approval record states that a dealer applied for a credit line of 800,000 yuan, and the approval column is empty. The draft sentence under review is: "Model A vacuum flask is orderable this season, and this dealer has obtained a credit line of 800,000 yuan." The first half has handbook support. The second half rests only on an applied amount with an empty approval column, so it is unsupported.

Three common errors

1. Entering AIPROXY_API_ENDPOINT as a complete request address that includes /v1 or /chat/completions. The input is the environment variable value, the comparison object is the documented root address form, and the observable result is whether startup validation passes.
2. Treating the applied amount on the credit approval record as an obtained credit line in the draft sentence. The input is the record text with an empty approval column, the comparison object is the phrase "has obtained a credit line of 800,000 yuan," and the observable result is that this phrase lacks evidence of completed approval.
3. Using a retrieval result or a form entry as a substitute for authorization execution. The input is a handbook clause returned by the knowledge base, the comparison object is whether the draft sentence claims orderability on that basis, and the observable result is that retrieval supplies textual grounds only and does not constitute ordering authorization.

How to verify the setup

1. Input the handbook text "Model A vacuum flask is orderable this season, minimum order 500 units" together with the first half of the draft sentence. Compare whether subject, object and period match. The observable result is that this part is supported by the source text.
2. Input the credit approval record text together with the second half of the draft sentence. Compare the state of the approval column against the phrase "has obtained a credit line of 800,000 yuan." The observable result is that the approval column is empty, so this part is unsupported and the evidence gap must be recorded.
3. Input Sample List A together with the draft sentence. Compare the sample status "sample arrived" against whether the draft sentence claims orderability. The observable result is that sample arrival and orderability are different facts and cannot stand in for each other.

These checks test the named primary task of drafting product information and selection answers from the imported materials. They do not test identity certification, report analysis or any downstream business action. Missing evidence stays missing: a pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination.

Related reading

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