What the documents look like
Commission settlement in live commerce and creator distribution happens after a sale closes. The working materials for this category are commission setting records, settlement statements, refund commission deduction details, and invoices. The review focus falls on valid promotion orders, refund deductions, and invoicing and payment. Creator contracts and finance systems supply a second set of figures, so a draft copy sentence has to name the same subject, the same period, and the same field as the material it cites.
Four document types carry the values that copy drafting will quote. Commission setting records hold the rate and the deduction conditions for a product. Settlement statements hold the payable commission amount for a creator over a period. Refund commission deduction details hold the deduction status for refunded orders. Invoices hold the billing and payment stage. A sentence about a rate, a settlement cycle, or a deduction condition has to match the subject, period, and field on these materials before it can be treated as supported.
For this page the task is product and marketing copy drafting, and the feature under discussion is model connections and configuration. The drafting step consumes text that has already been checked against the four document types. Model access and configuration do not verify business content, and retrieval or generation results do not carry permission or entity isolation.
Constraints at the model connections step
Model access runs through AI Proxy. The App and Admin services must both set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN. If either is missing, startup validation fails. Older OpenAI environment variables and single-model custom request addresses do not replace these two settings.
Model channels, model ID, model alias, maximum context, maximum response, maximum quoted content, and maximum temperature are defined in model configuration. Application and node parameters are set on the orchestration side. A compatible chat interface field with the same name does not override orchestration. Index models, rerank models, speech synthesis, and speech recognition each have their own type. The system needs at least one language model and one index model to work.
Model probing covers only currently enabled system models. Disabled models are not probed and produce no history. The probe interval accepts an integer between 5 and 60 minutes. A model is judged abnormal only when all four attempts in one round fail. A green probe status describes model availability and does not describe whether a draft sentence is supported by the settlement materials.
Hypothetical drill sample: during one commission settlement review, the source material reads "the March 2026 settlement statement shows creator A's valid promotion order commission as 8,000 yuan", "the refund commission deduction details for the same period record a deduction of 1,200 yuan on the refunded order", and "the commission setting record for that product shows a commission rate of 15%". The draft sentence to be judged is "creator A received 8,000 yuan in commission at a 15% rate in March 2026, and refunds are not deducted". The sentence involves subject creator A, object that product, period March 2026, and the fields commission rate and deduction status. The authorization boundary is limited to those three materials.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this drill | Reason and comparison method |
|---|---|---|
AIPROXY_API_ENDPOINT | Assume the AI Proxy root address is entered | The document requires no /v1 suffix; compare against the service address given by the deployment template |
AIPROXY_API_TOKEN | Assume a non-empty admin token is entered | The document requires it to match the server side; compare against the AI Proxy configuration value |
| Model channel and model ID | Assume the channel name and model ID are named for the copy task | The document requires the model ID to be globally unique; compare against the model configuration list |
| Maximum quoted content | Assume the value is filled per the selected language model field | The document lists it as a language model field; compare against the selected model's capability value |
| Probe interval | Assume 30 minutes is set | The document allows an integer from 5 to 60 minutes; compare against the allowed range |
| Alert webhook address | Assume it is left empty for now | The document lists it as optional; after filling, compare against the test button result |
Three common errors
- Treating the settlement statement commission amount as the rate basis for a draft sentence. Input is the statement amount and the draft sentence rate. Compare the statement field against the commission setting record field. The condition is met only when subject, period, and field value agree; otherwise the sentence is recorded as unsupported.
- Treating a deduction record in the refund commission deduction details as grounds for the draft sentence "refunds are not deducted". Input is the deduction details and the draft sentence. Compare the deduction status field. When the two values are mutually exclusive, record a contradiction; when the field is not stated, record it as pending verification.
- Treating a green model probe status as grounds for the correctness of the copy content. Input is the probe status and the draft sentence. Compare the probe object against the copy object. The probe describes model availability only and says nothing about whether the cited materials support the sentence.
How to verify the setup
- Input the three source materials above and the draft sentence. Compare the commission rate field and the deduction status field. The expected result is that the draft sentence disagrees with the materials on the refund deduction part and should be rewritten or marked pending verification.
- Input the hypothetical values for
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKEN. Compare whether either is missing. The expected result is that startup validation passes when both are present and the token is non-empty; when one is missing, record the gap. - Input the hypothetical 30-minute probe interval. Compare whether it falls inside the 5-to-60-minute integer range. The expected result is that it can be saved inside the range; outside the range, record it as not passing.
These checks describe what to compare and what remains unknown. No test, deployment, approval, or business action is claimed to have occurred.