What the documents look like
Franchise fee and deposit settlement deals with funds and breach deductions. It does not handle government filing, which belongs to franchise registration administration. The working records are the franchise fee invoice, the deposit ledger, the deduction notice, and the contract payment schedule. These come from the finance system, the franchise contract, bank statements, and store violation records. The people involved are chain finance staff, the franchise management department, and franchisees. Errors here produce funding disputes, franchise litigation, and deposit refund conflicts.
This page covers one task only: product material and selection questions inside model connection and configuration. It does not build business functions that the documents do not support. For the model connection step, the relevant facts are these. The system needs at least one language model and one index model to work normally. AI Proxy is a required dependency, and both the App service and any Admin service that reuses the App/Service configuration must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN; missing either one causes startup validation to fail. The model ID is the value of the model field in the request body and is globally unique. The model name is for display and can be customized. A model channel corresponds to a provider protocol, and model mapping can map the model requested by FastGPT to the model actually provided upstream.
Constraints at the model connections step
Assumed paper exercise sample. A chain brand franchise management department intends to check the model selection for one deduction notice. The first source text is: "Deduction notice for August 2026, subject is East China Region Store A, reason is liquidated damages for late payment under Article 7.2 of the contract, amount RMB 12,000, status pending review by the franchise management department." The second source text is: "The deposit ledger shows Store A had a deposit balance of RMB 50,000 as of 31 August 2026, of which RMB 30,000 is frozen." The draft sentence under review is: "This deduction has been taken from the deposit and settlement is complete." The pending request is: "Connect this deduction notice to model question answering and confirm whether the deposit deduction has been executed."
The constraints that follow from the documents. Model status probing covers only system models that are currently enabled; disabled models are not probed and produce no history. A relevance score is not a category, entity, or permission filter, and retrieval does not perform authorization. The same-vector-model requirement for a retrieval library does not extend to other external systems. The model channel protocol type is chosen against the actual provider that supplies the model; the documentation states that most providers follow the OpenAI protocol. Model mapping is configured against the channel, as in the documented example where gpt-4o-test maps to gpt-4o.
Configuration and implementation checks
The table below stays with settings the official documents describe and with checks tied to this page's sample. Numeric entries are marked as assumed starting values for this exercise, with the reason for the choice and the value being compared against.
| Configuration item | Assumed starting value for this exercise | Check method and reason |
|---|---|---|
| AIPROXYAPIENDPOINT | AI Proxy service root address, without /v1 appended | The document requires this for App and Admin; compare the format against the documented example http://aiproxy:3000 |
| AIPROXYAPITOKEN | Non-empty administrator token, consistent with the server side | The document requires a non-empty value; compare against the document statement, and do not substitute an application API key |
| Model channel protocol type | The provider protocol matching the model the sample needs | The document states most providers follow the OpenAI protocol; compare against the actual model source |
| Model mapping | Map the FastGPT request model to the upstream actual model | The documented example maps gpt-4o-test to gpt-4o; compare against the channel configuration |
| Probe interval | 30 minutes as an assumed starting value for this exercise | The document supports integers from 5 to 60 minutes; a middle value makes observation easier, compared against the documented range |
| Alert Webhook address | An internal callback URL as an assumed starting value for this exercise | The document marks this optional and supports HTTP/HTTPS; compare connectivity against the documented requirement |
Three common errors
- Treating a pending review status as executed. The sample deduction notice carries the status pending review by the franchise management department, while the draft sentence states the deduction has been taken and settlement is complete. A pending review does not show an absence of operating authority, and it does not mean the deduction was executed. The input is the deduction notice text; the comparison object is the status field; the observable result is that the draft sentence is marked unsupported when the status and the sentence disagree.
- Treating a retrieval result as an authorization basis. The sample request asks whether the deposit deduction has been executed. If the answer relies only on the retrieved ledger balance record, the input is the ledger text, the comparison object is whether the ledger states a deduction action, and the observable result is that an evidence gap is recorded when the ledger does not state the deduction was executed.
- Extending the coverage of model probing. If the sample environment contains a disabled model, the input is the probe configuration, the comparison object is the enabled model list, and the observable result is that the disabled model produces no probe record, so its status cannot be used to infer whether the settlement business step is complete.
How to verify the setup
- Input the sample deduction notice text and the draft sentence. Compare whether the draft sentence agrees with the status field of the notice. The expected observation is that the draft sentence is judged unsupported when it contradicts the pending review status. No live test is claimed.
- Input the sample ledger text and the pending request. Compare whether the ledger states a deduction execution action. The expected observation is that an evidence gap is recorded when the ledger states only a balance and a frozen amount, and no inference is drawn that the deduction was executed.
- Input the model probe configuration and the enabled model list. Compare whether the probe scope covers only enabled models. The expected observation is that a disabled model has no probe record, and its status is not used to judge whether a business action is complete.
Across all three checks, missing evidence stays missing. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination. Retrieval is not authorization, and forms collect values; business verification, approval, database writes, and external synchronization each need their own stated downstream implementation.