What the documents look like
Marketplace advertising adjusts bids against live traffic and conversion signals. Platform campaign approval is not the governing workflow. The working documents are campaign plans, keyword reports, audience package descriptions and conversion attribution tables. Data arrives daily or hourly from sources such as Juliang Qianchuan, JD Kuaiche, Alimama and store transaction records.
For copy drafting, the practical consequence is that every claim in a draft has to trace back to a dated, scoped document. A campaign plan states a target region and period. A keyword report states a bid range over a defined date window. An audience package description states an eligibility window. A conversion attribution table states an attribution window. These scopes are what a draft may reference, and model connection and configuration is the step that keeps the source material retrievable and checkable.
Constraints at the model connections step
Copy drafting depends on knowledge base retrieval and a language model. AI Proxy is a required dependency. Both App and Admin must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN; if either is missing, startup validation fails. AIPROXYAPIENDPOINT is the AI Proxy service root address and does not carry /v1. AIPROXYAPITOKEN is an administrator token that must match the server configuration and cannot be empty.
The model ID is the value of the model field in the request body and is globally unique. The model name is a display name. The system needs at least one language model and one indexing model to operate. Semantic recall and relevance scores do not perform permission or entity isolation. A retrieval result is not an authorization result. Knowledge bases built on different vector models cannot be associated with each other; the identifier is checked and non-matching bases are excluded.
Hypothetical exercise sample. The source material consists of three sentences: "This campaign targets summer sunscreen in the East China region; the bid range appears in the keyword report daily data for 2026-07-01 through 2026-07-31; the audience package description is limited to users who added to cart but did not pay within the last 30 days; the conversion attribution table uses a 7-day post-click window." The draft sentence to be judged is: "This product is offered to all users nationwide on a long-term basis." Check subject, object, period and field. The material limits region to East China, audience to a 30-day add-to-cart non-payment window, and attribution to 7 days. The draft sentence widens both region and duration, so it is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Comparison method |
|---|---|---|
| AIPROXYAPIENDPOINT | http://aiproxy:3000 | Compare against the root address given by the deployment template; confirm no /v1 is appended |
| AIPROXYAPITOKEN | A non-empty value generated at deployment | Compare against the server configuration; confirm non-empty and consistent |
| Language model | One model that has passed a model test | Run the model test on the model provider page and observe the returned result |
| Indexing model | One model that has passed a model test | Same as above; confirm the indexing model is usable |
| Knowledge base files | The three source material sentences above | After import, check whether the files parsed successfully |
| Retrieval check | Paper query "East China sunscreen bid" | Observe whether returned fragments come from those three materials and record the actual observation |
Reason for the starting values: AIPROXYAPIENDPOINT follows the documented example form so it can be compared against the deployment template. The models are chosen as already passing a test so they can be compared against test output. These are hypothetical exercise starting points, not official defaults.
Three common errors
- Accepting the region and duration in the draft sentence as given. Input is the source material and the draft sentence. Compare subject, object, period and field. The observable result is that the draft sentence exceeds the material scope: the material limits to East China and a 7-day attribution window, while the draft claims nationwide and long-term.
- Treating a relevance score as a permission filter. Input is a score and a record returned by one retrieval. Compare the knowledge base identifier the record belongs to. The observable result is that the score does not change record ownership; a record outside the authorized scope still requires separate checking.
- Treating a pending-review status as granted authorization. Input is the phrase "pending review" in the material. Compare whether a signed or approved record exists. The observable result is that when only pending review exists, the authorization fact remains unsupported.
How to verify the setup
- Input the three source material sentences and the draft sentence. Compare whether the region, audience and attribution window in the draft sentence fall within the material scope. The observable result is that a draft sentence exceeding the scope is judged unsupported.
- Input the paper query "East China sunscreen bid". Compare whether the source identifier of returned fragments belongs to those three materials. The observable result is that fragment sources are checkable; if the return is empty, carry the empty result into downstream design rather than inferring that a rule-based retrieval must be empty.
- Input the phrase "pending review" from the material. Compare whether a signed or approved record exists. The observable result is that with no approval record, the evidence gap is preserved and authorization is not judged complete.