What the documents look like
Cross-border creator commission work draws on four source families: Selected Alliance backend records, order after-sales data, finance system data, and creator contracts. Settlement happens after a sale closes, so review centers on valid promotion and after-sales commission deduction rather than on pre-sale terms. Finance settlement staff, Selected Alliance operations staff, and creator business staff all read the same materials, which means one commission event can carry contract clauses, statements, and refund deduction details in different languages. Materials update weekly or monthly, so retrieval configuration is checked per material batch and per language rather than once for the whole program.
A single commission event may therefore have a Chinese settlement statement and an English creator contract that describe the same period. The statement states order counts and deduction counts; the contract states the settlement rhythm and the deduction principle. Neither document, by itself, states whether payment has occurred.
Constraints at the model connections step
This page addresses model connections and configuration only. It does not address commission calculation. Per the official documents, the App and Admin services share configuration, and from V4.17.0 AI Proxy is a required dependency: AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN must both be set, and a missing item causes startup validation to fail. Model configuration is performed on the administrator model provider page. Model ID is globally unique and is the value of the model field in the request body. Index models index text chunks for related-text retrieval; rerank models reorder retrieval results to improve ranking. Retrieval is the act of fetching materials. It does not constitute authorization, and it does not replace finance verification, approval, database writes, or external synchronization.
Hypothetical drill sample: a cross-border creator cooperation project has commission settlement materials consisting of one Chinese settlement statement and one English creator contract. The statement reads: "Valid promotion orders for this period: 120. Refund commission deductions: 8. Settlement period: March 2026." The contract reads: "Commission shall be settled monthly, subject to after-sales deduction." A draft sentence reads: "This period's commission has been settled and paid in accordance with the contract." This draft sentence is pending judgment. The available materials support the settlement period and the deduction count. They do not state payment status.
Configuration and implementation checks
The table below is organized around this sample. The values are hypothetical starting points for this drill, with the reason for choosing them and the comparison target stated.
| Configuration item | Hypothetical starting point for this drill | Comparison method and reason |
|---|---|---|
AIPROXY_API_ENDPOINT | Hypothetically set to http://aiproxy:3000 | The official document requires the service root address without appending /v1; compare against the documented example and confirm no extra path |
AIPROXY_API_TOKEN | Hypothetically a non-empty string | The official document requires it to match the server side and be non-empty; compare against the AI Proxy administrative configuration |
| Index model | Hypothetically one index model enabled | The official document requires the system to have at least one language model and one index model; compare against the model list |
| Model ID | Hypothetically gpt-4o-test | The official document states model ID is the value of the model field in the request body; compare against the channel model mapping |
| Rerank model | Hypothetically enabled as needed | The official document states rerank improves retrieval ranking; compare against retrieval settings, and when not enabled, reranking does not run |
| Retrieval scope | Hypothetically limited to this settlement material batch | Retrieval does not constitute authorization; compare against the target scope that carries access rights, and record a gap when absent |
Three common errors
- Treating a retrieval hit as completed settlement or payment. In the sample, the draft sentence reads "This period's commission has been settled and paid in accordance with the contract," while the materials state only the settlement period and the deduction count. The input is the draft sentence plus the two source texts; the comparison is whether a payment field exists; the observable result is that payment status is unsupported.
- Mixing up model ID and model name. In the sample, the channel mapping states
gpt-4o-testcorresponds togpt-4o. If the display name is used as the request value, the input is the model configuration; the comparison is themodelfield against the channel model; the observable result is a mapping mismatch. - Treating a pending review request as approved. In the sample, if a settlement handover request appears with pending review status, the input is the request status; the comparison is whether a completed approval record exists; the observable result is that pending review does not constitute approval.
How to verify the setup
- Input the two source texts and the draft sentence from the sample. Compare whether the payment wording in the draft sentence corresponds to a payment field in the source texts. The condition is met only when the source texts explicitly state completed payment; otherwise record the result as unsupported.
- Input the model configuration. Compare whether
AIPROXY_API_ENDPOINTcontains a/v1suffix and whetherAIPROXY_API_TOKENis non-empty. The condition is met only when both items conform to the official requirements. - Input a retrieval request. Compare whether the returned results fall within the target material batch that carries access rights. The condition is met only when results fall inside that scope. For an empty result, use the explicit paper input
[]to check the downstream design, without claiming that a fabricated rule necessarily returns empty.
This page has been checked against public technical facts for identifying related content in the documents. It does not require copying those facts as a fixed five-line table; the body is written independently.