What This Category of Data Looks Like
Data sources for ad creative review scenarios include ad platform serving creatives, landing page screenshots, product qualification documents, and advertising law compliance clauses. Update frequency is daily or real-time. Daily processed documents include single ad creative content, landing page compliance check records, violation rectification opinions issued by reviewers, and historical ticket modification records. Data fields cover creative copy, landing page compliance tags, violation point descriptions, rectification requirements, and more. There is no unified fixed unit system, and parsing rules must be adapted based on different source documents.
How These Data Characteristics Impact Reranking and Hybrid Retrieval
Multi-source heterogeneous data sources require hybrid retrieval to adapt to document content in different formats. This ensures ad platform creatives, compliance clauses, historical tickets, and other data can be effectively retrieved. Real-time or daily update rhythms require retrieval and reranking processes to have low latency. This avoids impacting the real-time review efficiency of on-call operations staff. The one-by-one review business model requires the reranking link to prioritize content directly related to the current creative’s violation risks. This reduces interference from irrelevant information. The severity of potential error consequences requires retrieval and reranking results to have high accuracy. This avoids missing key compliance basis or introducing misleading content. The operational on-call and ticket Q&A scenario also requires retrieval results to quickly associate historical review records and compliance clauses. This shortens problem troubleshooting time.
Recommended Configuration Settings
| Config Item | Recommended Setting | Rationale |
|---|---|---|
| Reranking Toggle | Enabled | Ad creative review requires accurate matching of violation-related content. Reranking can improve result relevance |
| Similarity Threshold | 0.75–0.80 | Filter low-relevance non-violation content, avoid interfering with review judgments |
| Single-Pass Recall Cap | Top 10 results | Balance retrieval efficiency and recall coverage, adapt to the information needs of single creative review |
| Retrieval Mode | Hybrid Retrieval | Need to integrate multi-source data including ad platform creative data, advertising law compliance clauses, and historical review tickets |
| Reranking Model Video Memory Release | Automatically unload after retrieval completes | Avoid reranking model from occupying video memory continuously after retrieval, relieve operational server pressure, adapt to FastGPT v4.17.0 version video memory management rules |
| Query Optimization Toggle | Enabled | Optimize the retrieval intent of user tickets, improve matching accuracy of multi-source data |
Three Common Misconfigurations
- Phenomenon: The reranking model continues to occupy server video memory after retrieval completes, and does not automatically release it. Cause: The rule to automatically unload the reranking model after retrieval is not configured, resulting in long-term occupation of video memory resources.
- Phenomenon: After enabling query optimization and result reranking, the review ticket response time exceeds 30 seconds. Cause: The single-pass recall cap is set too high, or the reranking model selection is not adapted to lightweight deployment, resulting in processing delay exceeding business requirements.
- Phenomenon: Reranked results are not displayed during online recall testing, even though the reranking toggle is shown as enabled. Cause: Hybrid retrieval mode is not bound to the reranking function, or the reranking association configuration is not enabled synchronously during the knowledge base indexing phase.
How to Verify Correct Configuration
- View the server video memory monitoring panel to confirm whether the reranking model’s video memory occupancy is automatically released after retrieval completes.
- Initiate a review ticket test for a single ad creative, and verify whether the response time meets the business expected threshold.
- During online recall testing, compare the result ranking with and without reranking enabled to confirm that the reranking logic is active.
- Check the knowledge base indexing configuration page to confirm that hybrid retrieval mode and reranking function are enabled synchronously.
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.