What this category of data looks like
The core data of this category includes ad creative copy, landing page screenshots, historical review opinions, product qualification documents, and advertising law compliance clauses. Data is sourced from ad delivery platforms, landing page backends, product qualification systems, and public compliance documents. Update frequency is daily or real-time. The structure of a single data entry includes unique creative identifier, delivery channel, review status, and violation type fields. There is no unified fixed unit, and it adapts to the format of source documents. It is used daily by ad reviewers, agency optimizers, and brand compliance personnel to check the compliance of delivered content item by item.
What constraints these characteristics impose on knowledge base retrieval and recall
Scattered and diverse data formats require the retrieval link to support multi-type document parsing and cross-source association. This avoids missing cross-platform compliance basis and historical review cases. Real-time or daily updated data requires supporting synchronization mechanisms. These mechanisms ensure knowledge base content aligns with the latest compliance clauses and delivery materials. Without them, review basis may lag. Single data entries have no unified fixed unit. Original field information must be retained for accurate matching of violation scenarios. The item-by-item check usage scenario requires recall results to be sorted by violation risk priority. This prevents irrelevant content from interfering with review judgments. The consequences of errors are directly linked to ad delivery risks. Retrieval result traceability information must be complete, and the recall scope must cover all compliance basis and historical review records.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Ad creative review needs to balance compliance judgment accuracy and retrieval efficiency. A threshold that is too low will introduce irrelevant violation judgments, while a threshold that is too high will miss potential compliance risk points |
limit | Top 6–10 entries | A single review needs to cover core compliance basis and historical cases. Too many entries will increase the context processing load, while too few will fail to cover all potential risks |
searchMode | vector+keyword | Ad compliance content includes both semantically related compliance clauses and accurately matched qualification fields. The hybrid mode can balance semantic recall and keyword accurate matching |
usingReRank | Enabled | Recall results must be sorted by violation risk priority to help reviewers quickly locate high-risk content and improve review efficiency |
datasetSearchUsingExtensionQuery | Enabled | User review query statements can be extended, such as expanding "landing page compliance" to "landing page consistency, advertising law landing page clauses" and others, to improve recall accuracy |
collectionIds | Bind the collection ID of the knowledge base exclusive to ad creative review | Isolate knowledge base data from different business scenarios, avoid irrelevant content interfering with retrieval results, and ensure the targeting of review basis |
Three common errors
- Issue: The number of results returned by calling the knowledge base retrieval interface is lower than expected. Cause: The
collectionIdsparameter was not correctly specified, causing the retrieval scope to cover irrelevant knowledge base content not exclusive to ad creative review. - Issue: Retrieval accuracy decreases after multiple rounds of review conversations, and returns to normal after opening a new conversation. Cause:
datasetSearchUsingExtensionQuerywas not enabled and context filtering was not configured, and redundant historical review context interfered with the semantic matching of the current query. - Issue: After replacing the embedding model, the previously imported knowledge base cannot match new retrieval requirements. Cause: The reindexing process for the original knowledge base was not triggered, and vectors generated by the old model were still used for retrieval.
How to confirm the configuration is correct
- View the retrieval configuration panel to confirm that the
collectionIdsparameter is bound to the knowledge base collection exclusive to ad creative review, and no irrelevant collections are included in the retrieval scope. - Initiate a test query to verify that retrieval results are sorted by violation risk priority, and confirm that the
usingReRankfunction is enabled. - After replacing the embedding model, trigger the reindexing process for the original knowledge base, and verify that the new model's vectors are applied to existing data.
- Initiate multiple rounds of test conversations to confirm that retrieval accuracy does not decrease significantly as the number of conversation rounds increases, and confirm that the
datasetSearchUsingExtensionQueryfunction is enabled.
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.