What the documents look like
Franchise supervision materials span four document types: store visit forms, franchise contracts, rectification notices and display standards. Cross-border versions of contracts and rectification notices frequently carry two language texts inside one file. Regional supervisors compare head office standards against actual franchisee execution, drawing material from the supervision app, POS sales, the franchise system and store photographs. Language, signing party and validity scope affect whether a retrieved passage can support later verification.
This page addresses one step only: model connection and configuration. It covers whether model channels, model identifiers and retrieval capability are configured correctly once multilingual materials enter the knowledge base. A retrieval result does not establish a franchisee's execution status and does not grant authorization.
Constraints at the model connections step
Cross-border multilingual material requires an index model and a language model to be separately available. The system needs at least one language model and one index model to operate normally; only one speech recognition model takes effect. The model ID is the value of the model field in the request body and is globally unique. The model name is a display label and can be customized. Model mapping translates the model FastGPT requests into the upstream model actually called, for example {"gpt-4o-test": "gpt-4o"}, so the request side and upstream side can carry different names. The proxy address field takes a BaseUrl, not a complete request address, and the need to append /v1 must be confirmed.
AI Proxy is a required dependency from V4.17.0. AIPROXY_API_ENDPOINT holds the service root address without /v1; AIPROXY_API_TOKEN must be non-empty and consistent with the server. Legacy OpenAI environment variables and single-model custom request addresses cannot replace these two settings. Model capabilities must be checked against the input the application requires, with multimodal support, tool calling and thinking configuration verified separately. Application parameters belong in the application or node configuration.
Hypothetical drill sample: a franchise supervision team prepares a cross-border rectification notice. The source text reads: "This notice is delivered to Franchise Store No. 12, East China Region. The issuing party is the Head Office Operations Center. The issue date is 10 March 2026. The rectification period is 15 days after delivery." The second sentence reads: "The attached display standard is the November 2025 version and applies to all ambient shelving." The draft sentence awaiting judgment is: "The store has completed rectification." That sentence omits completion time, acceptance party and acceptance conclusion, so it is unsupported.
Configuration and implementation checks
The table below lists connection and verification items relevant to this page. Values are starting points for this hypothetical drill only and must be compared against the target environment's actual values.
| Configuration item | Hypothetical starting point | Comparison method |
|---|---|---|
AIPROXY_API_ENDPOINT | http://aiproxy:3000 | Compare with the deployment root address; confirm /v1 is not appended |
AIPROXY_API_TOKEN | Empty placeholder, to be filled | Compare with server configuration; confirm non-empty and consistent |
| Language model | One enabled | Confirm enabled on the model provider page |
| Index model | One enabled | Confirm enabled on the model provider page |
| Model mapping | {"gpt-4o-test": "gpt-4o"} | Compare with the upstream model name inside the channel |
| Model probe interval | 30 minutes | Compare with the documented 5 to 60 minute integer range |
Model probing covers only currently enabled system models. After a first failure the system retries at most 3 times, giving at most 4 attempts per round; only when all 4 attempts fail is the model judged abnormal. The alert Webhook fires on edge changes of status and does not repeat while a model stays abnormal.
Three common errors
- Treating a retrieval hit as evidence that a franchise store has completed rectification. Retrieval returns material fragments only; completion requires an acceptance party and an acceptance conclusion. Check: input the rectification notice text above, compare whether the returned fragment contains acceptance fields, and observe whether a completion conclusion appears.
- Writing the upstream real model name into the mapping key. The two sides of the mapping carry different meanings, and the request side holds the FastGPT model ID. Check: compare the mapping key against the channel model list and observe whether the two are conflated.
- Treating similarity or recall results as a permission filter. Retrieval does not perform authorization judgment. Check: run the same query from accounts with different access scopes and observe whether material outside the permitted scope is returned.
How to verify the setup
- Connection check: input the actual values of
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKEN, compare whether the root address condition and the non-empty consistency condition are met, and observe whether startup validation passes. When it does not pass, record a configuration gap. - Mapping check: input the mapping
{"gpt-4o-test": "gpt-4o"}and the channel model list, compare whether the key is the FastGPT-side model ID, and observe whether the upstream request model name isgpt-4o. - Draft check: input the rectification notice text above together with the draft sentence "The store has completed rectification." Compare whether the source text states an acceptance party and an acceptance conclusion, and observe whether the draft sentence is supported. The source text does not state them, so the draft is unsupported; retain the evidence gap without supplying authorization or effective clauses.