What the documents look like
Creator sales campaigns run on external creator performance and commission settlement rather than internal brand scheduling. Brand commerce teams, MCN organizations and creator agents work from creator cooperation agreements, commission setting sheets, sample shipment sheets and livestream performance reports. Product question answering has to determine whether a given item can enter a candidate list for a specific campaign, and which document field governs that check. The underlying material arrives from selected alliance sources, Xingtu, creator history and store transaction details.
These documents update per campaign, with some weekly rollups. The inputs to selection question answering are fields carried in the cooperation agreement, commission setting sheet, sample shipment sheet and livestream report. Model connection and configuration determines whether those fields can be read, retrieved and answered. A field the material does not state stays unknown. A retrieval hit does not constitute authorization.
Constraints at the model connections step
FastGPT connects to model providers through AI Proxy. App and Admin must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN; missing either one fails startup validation. The system needs at least one language model and one index model. Model configuration separates language, index, rerank, speech synthesis and speech recognition models, and their fields differ.
Model ID is the value of the model field in the request body and is globally unique; the model name is display only. Model mapping converts the model FastGPT requests into the upstream model actually called. Channel priority ranges from 1 to 100, and a larger value is selected more readily. Model status probing applies only to enabled models; after an initial failure the system retries at most 3 times, so a single round makes at most 4 attempts, and only when all 4 fail is the model judged abnormal.
Retrieval libraries associate with vector models under documented conditions. Libraries built on different vector models cannot be associated together; verify the identifier first, exclude those that do not match, and record what is observed. A relevance score does not carry permission or entity isolation duties, and retrieval does not constitute authorization.
Hypothetical paper exercise sample: the source material reads "Creator A signed a cooperation agreement with Brand B on 1 March 2026, agreeing that item A would be sold in the 10 March special session at a 15% commission; the sample was shipped on 3 March, and the livestream report shows item A sold 200 units." The draft sentence under review is "Creator A is authorized to sell item A in the 10 March special session, with commission settled at 15%." This sample only illustrates the checking method and does not represent a real customer or real cooperation.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to this page's sample. Where a value is a hypothetical starting point for this exercise, the reason for choosing it and the value it is compared against are stated.
| Configuration item | Hypothetical starting point for this exercise | Check method |
|---|---|---|
| AIPROXYAPIENDPOINT | AI Proxy root address, no /v1 suffix | Compare against the deployment document example and confirm no path suffix |
| AIPROXYAPITOKEN | Non-empty administrator token | Compare against server configuration and confirm it matches and is non-empty |
| Language model and index model | One of each enabled | Confirm enabled status on the model provider page |
| Model mapping | Request model mapped to upstream model | Compare the mapping against the channel model |
| Model status probe interval | Assume 30 minutes | Document allows an integer from 5 to 60 minutes; the midpoint is chosen for easier observation |
| Channel priority | Assume 50 | Document range is 1 to 100; the midpoint is chosen for easier ordering comparison |
Model capabilities are checked against the inputs the application requires, with multimodal, tool calling and reasoning settings verified separately. Application parameters are set in the application or node configuration; same-named fields on a chat-compatible interface cannot be assumed to override orchestration. Thinking budget is a reasoning budget and does not serve as a citation token ceiling.
Three common errors
- Treating a retrieval hit as authorization. In the sample, retrieving a fragment of the cooperation agreement only shows the material exists; it does not show Creator A is authorized. Compare the draft sentence against the source text, and record the authorization scope as unsupported when the material does not state it.
- Confusing model mapping with model ID. Model ID is globally unique, while model mapping is a conversion from the requested model to the upstream model. Compare the mapping shown in configuration against the channel model; when they differ, the actual configuration governs.
- Treating a pending state as already effective. In the sample, the sample having shipped does not mean the cooperation is concluded or the contract has taken effect. Compare the status field in the material, and record an evidence gap when the status field is absent.
How to verify the setup
- Enter the sample source text and the draft sentence, then compare whether the authorization, commission and date in the draft share the same subject, the same period and the same field as the source. Only fields the source states explicitly can be judged as stated; an authorization scope the source does not write is judged unsupported.
- Enter the sample source text and a handover request reading "please arrange settlement for 10 March at 15% commission." Compare the commission rate in the request against the commission field in the source. When the rate matches and the subject matches, the condition is met; when the subject differs or the field is missing, record it as pending verification.
- Enter an empty result
[]together with the downstream question-answering design, and compare whether the downstream gives an explicit notice for the empty result. When the downstream gives a notice for the empty result, the condition is met; when the downstream treats the empty result as authorization, record it as a design defect.
These checks are a paper exercise and no claim is made that they were actually run. The model status probe alarm Webhook pushes on edge changes in status as documented, with event types model_status_error or model_status_recovered; this page does not infer other automatic business actions.