What the documents look like
Short-video affiliate sales relies on long-term content exposure and product-link conversion rather than live explanation. The working materials for this category include short-video scripts, attached-product lists, commission plans, and violation feedback, with daily or weekly update cycles. Product questions and selection Q&A need the product references in scripts, the product status in attached-product lists, and the validity scope of commission plans to be checked in one view. These records live in separate sources, and a model connection must treat each source record as independently authoritative for its own field.
For the product questions and selection Q&A task, the model connection determines whether the assistant can retrieve script text, list records, and commission records and then answer selection questions. The connection does not itself confirm that a product is approved, that a commission rate is currently valid, or that a product may be attached to a video.
Constraints at the model connections step
Model access depends on AI Proxy. The App and Admin services must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN; a missing value in either causes startup validation to fail. Old OpenAI environment variables and single-model custom request addresses cannot substitute for these two settings. The system requires at least one language model and one indexing model to operate.
The model ID is the value of the model field in the request body and is globally unique. Model mapping translates the model requested by FastGPT into the actual upstream model, for example mapping gpt-4o-test to gpt-4o. Channel priority ranges from 1 to 100, and a larger value is selected more readily. Model probing applies only to currently enabled system models; after an initial failure the system retries up to 3 more times, and only when all 4 attempts fail is the model judged abnormal.
Retrieval recalls text and does not perform permission or entity isolation. Libraries built on different vector models cannot be associated together; identifiers must be checked and non-matching records excluded. A relevance score is not a category, entity, or permission filter, and retrieval is not authorization. Forms collect values; business verification, approval, database writes, and external synchronization require their own downstream implementation.
Configuration and implementation checks
The following hypothetical paper sample illustrates the checking method. It does not indicate that any test or deployment took place. Hypothetical source text: a short-video script paragraph 3 states "the attached product is Model A thermos, commission rate 12%"; the attached-product list for the same period records "Model A thermos, status is pending review, effective date: 2026-03-01"; the commission plan table records "Model A thermos, commission rate 12%, valid until 2026-03-31". Draft sentence to be judged: "Model A thermos has passed review and can be attached at a 12% commission." This sentence treats pending review as passed review and treats the commission rate as having no expiry.
| Configuration item | Value or check for this page | Explanation |
|---|---|---|
| AIPROXYAPIENDPOINT | Hypothetical starting value http://aiproxy:3000, without /v1 appended | Fill the AI Proxy root address; compare against the documented example |
| AIPROXYAPITOKEN | Hypothetical starting value is a non-empty token consistent with the server | An application API key cannot substitute |
| Language model and indexing model | At least one of each enabled | Minimum condition for normal system use |
| Model ID and model mapping | Check that the model field is globally unique | Mapping example gpt-4o-test to gpt-4o |
| Channel priority | Hypothetical starting value 50, compared against the 1–100 range | A larger value is selected more readily |
| Model probing | Applies only to enabled models, retry limit 3 | Abnormality is judged only after 4 failed attempts |
Three common errors
- Treating a pending-review status as passed review. The list records "status is pending review" and the draft sentence states "has passed review". These share the same subject and the same field with mutually exclusive values, which is a contradiction.
- Treating the commission rate as having no expiry. The commission plan table records "valid until 2026-03-31", and the draft sentence states no expiry. The draft claim is unsupported by the supplied record, and the evidence gap remains.
- Treating retrieval recall as authorization. Retrieval returning text about the Model A thermos does not mean the product passed review or may be attached. Authorization status must be checked separately against the list field.
How to verify the setup
- Input the hypothetical source text and the draft sentence, and compare the "status" field. The list gives "pending review" and the draft sentence gives "has passed review". The two values are mutually exclusive, so the comparison result is a contradiction.
- Input the commission plan table record and the draft sentence, and compare the "validity" field. The plan table states "valid until 2026-03-31" and the draft sentence states no expiry. The draft claim is unsupported, and the missing evidence remains missing.
- Input an empty result
[]together with the downstream design, and check how an empty result is handled. If the downstream design does not define empty-result handling, record that as a gap. Do not claim that a hypothetical rule retrieval necessarily returns[].
These checks test the named primary task: comparing proposed copy sentences against the supplied source wording and its limitations. They do not test identity certification, report analysis, or any other task.