What this category of data looks like
Data sources include advertising platforms, landing pages, product qualifications, and advertising law clauses. Daily processed documents include ad creative copy, landing page screenshots, audit opinions, and modification records. Data updates occur daily or in real time. Each creative corresponds to an independent audit task. Fields include creative type, delivery channel, violation type tag, rectification requirements, and other items. There are no unified fixed units; values change dynamically based on the attributes of individual creatives.
What constraints these characteristics impose on the model integration and configuration link
Ad creative audit data comes from multiple sources and supports daily or real-time updates. Model calls must adapt to high-concurrency real-time requests. Batch offline processing workflows cannot be relied upon. Creatives include copy, images, videos, and other formats. Text and visual content review must be supported simultaneously, so visual recognition capabilities must be configured. Each creative audit must associate multiple related documents. Sufficient context information must be retained to avoid truncation of key related content. Audit results directly affect ad delivery permissions. Judgment consistency must be ensured, and randomized output must be avoided.
The following configuration items are formulated based on FastGPT v4.17.0.
How to set the configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Ad creative audit requires association of multiple documents including creatives, landing pages, and product qualifications. Sufficient context must be retained to cover all related information |
maxTokens | 2000–3000 token | Complete violation point annotations and rectification suggestions must be output, to avoid truncation of key audit conclusions |
vision | Enabled | Landing page screenshots, creative images, or video creatives must be processed, to recognize visual violation content |
temperature | 0.1–0.3 | Consistency of audit conclusions must be maintained, to avoid randomly generated violation judgments |
AIPROXY_API_ENDPOINT | Root address of the AI Proxy service, for example http://aiproxy:3000. Do not append /v1 or /chat/completions paths. Do not use single-model addresses such as VLLM/Ollama/Xinference | Required environment variable for the App/Admin server, to adapt to high-concurrency real-time call requirements |
AIPROXY_API_TOKEN | Administrator token consistent with the AI Proxy server configuration, and cannot be empty. Do not use application API Keys, model vendor keys, tokens issued per application, or token counts estimated per request | Required environment variable for the App/Admin server, must match the AI Proxy server configuration |
Three common errors
- A 400 status code is returned when testing a channel. Troubleshooting steps: Check if the AIPROXYAPIENDPOINT configured on the App/Admin server is a valid root address of the AI Proxy service, with no /v1 or other paths appended. Check if the AIPROXYAPITOKEN matches the AI Proxy server configuration and is not empty.
- Audit results are truncated, and violation point annotations are incomplete. The cause is that
maxTokensis set too small, which cannot support complete output of audit conclusions. - Violation content in image-based creatives is not recognized. The cause is that the
visionconfiguration is not enabled, so audit tasks for visual materials cannot be processed.
How to confirm the configuration is complete
- Upload a single ad creative that includes text and images, trigger the audit process, and check if complete violation point annotations and rectification suggestions are generated.
- View model call logs, confirm that the actual values of
maxContextandmaxTokensmatch the configured values. - Replace creatives with different violation types, verify the consistency of audit conclusions, with no randomized judgment results.
- Upload purely image-based creatives, confirm that visual violation content is correctly recognized and annotated.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.