What the documents look like
Site selection and store opening work happens before a store begins trading, and the core deliverables are lease and construction handover. Copy drafting for goods and marketing therefore describes a store that has not opened. The working materials include trade-area footfall, map points of interest, historical store-type data from POS records, and landlord documents. The day-to-day paperwork is a site evaluation form, a lease letter of intent, a renovation budget, and an opening schedule. These materials are updated at irregular intervals and once per store, so drafting can rely only on the version currently on file. There is no basis for assuming batch synchronization or real-time refresh. Any floor area, lease term, opening date, or business category that appears in proposed copy needs a corresponding field in those documents. Content the documents do not state remains unsupported and must stay marked as pending verification.
Constraints at the model connections step
Model connections separate into three layers: model channels, model configuration, and application-side parameters. A channel carries the protocol type, model mapping, proxy address, and API key. Model configuration carries the model ID, context and response limits, image-input support, and tool-calling support. Application parameters are set inside the application or node configuration, and there is no basis for assuming that identically named fields on a chat-compatible interface override the orchestration. The system requires at least one language model and one index model to work normally. AI Proxy is a required dependency: both the App and an Admin service that reuses the server-side configuration must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN, and a missing item causes startup validation to fail.
Hypothetical paper exercise: a chain headquarters expansion team intends to draft opening warm-up copy for a community store that has not been signed. The material text available for checking reads as follows. The site evaluation form states that the shop unit has a gross floor area of about 120 square metres and that the trade area is community type. The lease letter of intent states that the intended lease term is five years and that rent terms remain under negotiation. The renovation budget states that the works are expected to take 45 days and that the start date is not yet set. The proposed sentence is: "This store has confirmed a five-year lease and is expected to open in three months." The question to decide is whether the available materials support that proposed sentence.
Configuration and implementation checks
The table below covers the hypothetical exercise and lists settings or implementation checks that the official documents describe for this feature and primary task. Where a value is marked as a hypothetical starting point, it is a comparison baseline for this paper exercise, chosen to show the reason for the choice and the object of comparison. It is not an official default.
| Configuration item | Value or object for this page | Notes and comparison method |
|---|---|---|
| Model types | Language model and index model | At least one of each is required; check that both are enabled |
| Model channel | Protocol type and model mapping | Check that the mapped model ID matches the model field in the request body |
| Proxy address | BaseUrl | Check that only the root address is entered and that /v1 is added only when required |
| Application parameters | Set in the application or node | Check that parameters take effect on the application side; do not assume a same-named chat interface field overrides orchestration |
| Context and response | Hypothetical starting points for maxContext and maxResponse | Chosen from the length of the exercise materials; compare against the actual character count of those materials |
| Image input | Checked against whether the exercise materials contain images | Plain-text materials need no image capability; image materials require checking the model vision field |
Three common errors
- Treating a letter of intent as an effective contract. In the exercise, the lease letter of intent states only that the intended lease term is five years and that rent terms remain under negotiation, while the proposed sentence states that a five-year lease has been confirmed. A letter of intent records intent; signature and effectiveness are separate facts. The check compares the status wording in the materials against the definitive wording of the proposed sentence.
- Writing an undetermined date as a fixed date. The renovation budget states that the start date is not yet set, while the proposed sentence states an expected opening in three months. The materials do not state an opening date, so the wording is unsupported and the evidence gap must be recorded rather than filled with an invented schedule clause.
- Treating a passed model test as business authorization. Channel tests and model tests show only that a model can be called; they do not constitute business review of copy content. The check separates model availability from factual support for the copy, since neither substitutes for the other.
How to verify the setup
- Input the three source sentences from the exercise materials and the proposed sentence, then compare "has confirmed a five-year lease" against the letter of intent wording "the intended lease term is five years and rent terms remain under negotiation." A contradiction requires mutually exclusive values in the same field for the same subject and period. This case is unsupported rather than contradictory, so the expected result is a pending-verification mark with the source quotation retained.
- Input the renovation budget wording "the start date is not yet set" and the proposed sentence "expected to open in three months," then compare whether a date field is stated in the materials. Where the materials do not state it, the expected result is that the sentence fails factual checking, and the check record should name the missing field as the opening date.
- Input the model mapping in the channel configuration and the model field in the request body, then compare whether the two match. If the mapping is gpt-4o-test and the request body is gpt-4o-test, the upstream actual model is gpt-4o, and the expected result is that the mapping relationship is traceable. If the test returns an error, record only the error information and the number of attempts, and do not infer a specific failure cause.