Marketplace advertising: model connections for multilingual work

What the documents look like

Cross-border advertising work for platform traffic is organized around four document families: delivery plans, keyword reports, audience package descriptions, and conversion attribution tables. Delivery adjusts pricing against traffic and conversion in near real time, so the working data changes daily or hourly, and the records come from platform ad consoles and shop transaction data. In a multilingual setting, one delivery plan can exist in both a Chinese version and an English version, keyword reports carry entries in several languages, audience package descriptions require the audience definition to be checked across languages, and conversion attribution tables must be read against transaction records in different currencies and time zones. These materials enter a knowledge base as text chunks, so the model connection and configuration step determines whether cross-language material can be indexed and recalled correctly.

The relevant industry facts for this page are the document set, the near-real-time pricing basis, and the multi-version nature of the same plan. Update frequency does not change the configuration work described here, so it is not treated as a driver of any setting.

Constraints at the model connections step

From V4.17.0, AI Proxy is a required dependency. The App service and any Admin service that reuses App/Service configuration must both set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either one causes startup validation to fail. Older OpenAI environment variables and single-model custom request addresses do not replace these two settings. Model configuration happens on the administrator model provider page, and the proxy address in a channel is a BaseUrl, not a full model request address. The system needs at least one language model and one index model to work, and only one speech recognition model takes effect.

For multilingual retrieval, the index model handles text chunks in more than one language. The defaultToken and maxToken fields of an index model bound the text splitting and vectorization range. Language model capabilities such as image input and tool use are checked against what the application actually receives. A retrieval library that uses the same index model is a precondition for association; libraries built on different vector models cannot be associated together, so the identifiers are checked first and non-matching libraries are excluded.

Hypothetical paper exercise. A delivery specialist submits a handover file whose source text reads: "From 1 September 2026 to 30 September 2026, plan A runs on the platform ad console with a daily budget of CNY 800 and a target ROI of 2.5. In the keyword report, the Chinese entry for 'summer sunscreen' is confirmed, and the English entry 'summer sunscreen' is pending translation review. In the audience package description, 'high-spending women' is defined as an average order value above CNY 500 in the last 30 days." The proposed sentence under review reads: "Plan A has a daily budget of CNY 800 and a target ROI of 2.5 for September 2026 and can be used directly for the English-language delivery." The pending request reads: "Please connect the above material to the multilingual retrieval library for checking the English-language delivery plan."

Configuration and implementation checks

The settings below relate to the model connection and retrieval check for the sample above. The values are starting points for this hypothetical exercise and are compared against the documented behavior of the corresponding variables.

Three common errors

1. Entering AIPROXY_API_ENDPOINT as a full request address with a /v1 suffix. The input is an address carrying /v1; the comparison is the documented root address format; the observable result is that the format does not match.
2. Using an application API key in place of AIPROXY_API_TOKEN. The input is an application API key; the comparison is the documented administrator token requirement; the observable result is that the credential type does not match.
3. Treating "can be used directly for the English-language delivery" in the proposed sentence as confirmed. The input is the proposed sentence; the comparison is the sample source text, which states only that the English entry is pending translation review and does not state that English-language delivery is authorized; the observable result is that the claim is unsupported.

How to verify the setup

1. Input the sample source text "From 1 September 2026 to 30 September 2026, plan A runs on the platform ad console with a daily budget of CNY 800 and a target ROI of 2.5" and compare it with the proposed sentence "Plan A has a daily budget of CNY 800 and a target ROI of 2.5 for September 2026." The same subject, the same period, and the same field values agree, so the condition is met.
2. Input the sample source text "the English entry 'summer sunscreen' is pending translation review" and compare it with the proposed sentence "can be used directly for the English-language delivery." The source states only a pending review status and does not state that translation is complete or that delivery is authorized, so the claim is unsupported and the evidence gap is recorded.
3. Input the pending request "Please connect the above material to the multilingual retrieval library for checking the English-language delivery plan" and compare it with the authorization status in the sample source text. The source does not state that the request has been approved, so the retrieval connection remains pending review and does not constitute authorized execution. If a retrieval returns an empty result, the downstream design is checked against an explicit paper input of []; no claim is made that a fictional-rule retrieval necessarily returns empty.

Related reading

• self-host/config/env
• self-host/config/model/intro
• FastGPT model connections