What the data for this category looks like
The data for advertising and marketing research at financial institutions comes primarily from ad campaign backend reports, ad placement data exported from media monitoring tools, competitor ad material libraries, industry trend analysis documents, and user behavior tracking logs. Data update cycles cover real-time campaign data, daily summary reports, and weekly deep analysis reports. Individual documents include fields such as ad channel classification, delivery time slots, impression counts, conversion costs, and material creative types. Some material documents also include format parameters like resolution and duration. Most data fields use numerical values, classification tags, and text descriptions.
What constraints these characteristics impose on model access and configuration
Advertising and marketing research data includes structured reports, unstructured material text, and image content. It also includes high-frequency real-time updated data. This requires model access to support both text and image embedding and inference. Structured data with multiple fields requires models to adapt to semantic encoding of different fields, to avoid field confusion. Real-time campaign data call demands that model API response delays match business update cycles. It also requires compatibility with batch calls to handle daily summary report data. Additionally, the diversity of material documents requires reserving access ports for multimodal models in the configuration, to support analysis of creative materials in different formats.
How to set the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
embedding_model | Select qwen-text-embedding-v2 for text data, select qwen-vl-embed for image data | Adapts to the two types of data sources for advertising and marketing: text reports and image materials, and matches the official capabilities of mainstream domestic large models |
model_api_base | Enter a locally deployed node or transit address, such as http://localhost:8000/v1 | Adapts to model call requirements deployed via OneAPI or local Xinference, and unifies API entry management |
max_batch_tokens | 8000–12000 tokens | Matches the text length of individual advertising and marketing reports, to avoid call failures caused by exceeding model context limits |
api_request_timeout | 60 seconds | Adapts to the call delay requirements of real-time campaign data, to avoid interrupting batch report processing due to timeouts |
image_embedding_mode | base64 | Matches the local upload scenario for advertising and marketing materials, and adapts to the input format requirements of backend deployed models |
recall_top_k | Top 10–15 results | Adapts to the multi-dimensional recall needs of advertising and marketing research, covering associated data across different delivery channels and material types |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis. It is recommended to conduct testing on internal samples prior to finalizing configuration settings.
Three common mistakes
- Unable to select the Tongyi Qianwen series models in the interface, with a prompt that the model is not configured. The cause is failing to enter the API address and key of the corresponding model in model management, or failing to enable the model's visibility permissions.
- A
400 Bad Requesterror is returned when calling the embedding model, and the backend log shows input format mismatch. The cause is failing to convert image materials tobase64format before passing them, or passing animage_urlparameter that does not meet the input format requirements of the deployed model. - Unable to trigger model calls after importing the OneAPI address and key, with an
invalid api keyerror returned. The cause is the OneAPI transit key not being bound to the corresponding model, or the key entered in the configuration not matching the key generated by OneAPI.
How to confirm the configuration is complete
- Navigate to the model management page, confirm the status of added embedding models and inference models is displayed as "Normal".
- Upload an advertising and marketing material document and a campaign report, trigger the embedding test, and verify the embedding progress bar completes with no error prompts.
- Initiate a knowledge base recall test, input a research query term, and check that returned recall results include field data from this category.
- Review API call logs, confirm the return status code for each model request is
200 OK, with no timeout or format error records.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.