What the data for this category looks like
The data for this category comes primarily from advertising delivery platforms, landing page resources, product qualification documents, and advertising law compliance clauses. Daily processed documents include advertising creative files, landing page screenshots, single audit opinion records, and creative modification history ledgers. Data updates occur daily or in real time. Core fields include creative unique identifiers, delivery channel types, copy content, associated qualification numbers, violation annotation items, and audit timestamps. Unlike the advertising account opening audit category, this category requires checking specific delivered creative content item by item, focusing on compliance verification for single deliveries.
What constraints these characteristics impose on the context and token link
Single processing for advertising creative audit involves splicing multi-source documents. Association verification between creatives and landing pages requires complete contextual association. Real-time updated qualifications and clauses require that context does not expire. Stacking multiple audit opinions and modification records quickly consumes token quotas. Excessively long contexts exceed large model processing limits. Additionally, multimodal metadata of creatives such as image descriptions increases token usage. Effective context scope must be strictly controlled to avoid irrelevant fields occupying token resources, which causes processing timeouts or content truncation.
How to set configurations
This document applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | First 20 context records | The context for advertising creative audit primarily consists of records associated with a single audit workflow. 20 records cover a complete audit modification chain |
quoteMaxToken | 8000 token | Total token count for a single creative + landing page + audit opinion is approximately 1500-3000 token. 8000 token can accommodate 3-5 complete audit context sets |
| Maximum context window | Calibrated based on actual testing | Adapts to the multi-source data splicing requirements of advertising creative audit, avoiding exceeding the large model's native window limit |
| Context recall threshold | 0.75-0.85 | Filters low-relevance historical audit records, reducing invalid token usage |
| Single recall count | First 5 records | Controls total context length, avoiding exceeding token limits |
Three common errors
- Symptom: After upgrading, errors of the type
Error response from daemon: error from registrappear, or the AI reply contains unexpected JSON formatting. Cause: Context configuration is not adapted to the multi-source data structure of advertising creative audit, causing the large model to fail to correctly parse associated fields. - Symptom:
maxContextis set to 30 records, but only 2 context records appear in conversation details. Cause: Creative IDs are not correctly associated with historical audit records, so the system only recalls a small number of directly matched entries. - Symptom: The large model prompts that context token count exceeds the limit. Cause:
quoteMaxTokenquota is not restricted, and complete multi-part creative and landing page metadata is included in the context, exceeding the token limit.
How to confirm configurations are properly set
- Import a single advertising creative and 3 associated historical audit records, check that the context window correctly loads all associated entries.
- Upload a creative package containing copy and landing page URLs, verify that token consumption statistics fall within the preset
quoteMaxTokenrange. - Simulate a multi-round audit modification conversation, confirm that the AI reply can associate previous audit opinions and creative modification traces.
- Adjust the
maxContextparameter, verify that the number of context recall records matches the preset value.
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.