What this category of data looks like
The data for this category typically originates from advertising platform backends, user profile databases, budget scheduling sheets, and attribution analysis tools. Daily generated documents include campaign setup documents, targeting setting configuration sheets, budget scheduling sheets, and conversion tracking configuration files. Data is updated daily or when the corresponding advertising campaign launches. Document fields include targeted audience package identifiers, budget amounts (unit: yuan), conversion event types, delivery time windows, and more. Common units include yuan, person-times, event identifiers, and similar.
How these data characteristics constrain knowledge base retrieval and recall
The data characteristics of this category impose multiple constraints on knowledge base retrieval and recall. Data sources are scattered and formats are inconsistent. Field mapping across advertising platforms and user profile databases must be automatically aligned during retrieval. Documents include multiple formats such as tables, configuration files, and plain text, so chunking rules must be adapted for different content types. Data update frequencies are flexible, with some changing in real time alongside advertising campaigns. Recall must prioritize associating the most recently updated datasets. The accuracy of delivery-related data directly affects subsequent delivery effectiveness. The confidence requirement for retrieval results is higher than that of general question-and-answer scenarios.
How to configure settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
similarity | 0.75–0.85 | The field matching accuracy requirement for campaign setup data is high. This interval filters low-relevance results and avoids incorrect configurations impacting delivery |
limit | Top 6 entries | Single pieces of delivery-related document content are lengthy. Limiting the number of recalled entries prevents context overlength that causes timeouts |
searchMode | vector_search | Adapts to structured data for campaign setup. Vector search can accurately match semantic associations between fields such as targeting and budget |
usingReRank | Enabled | Performs secondary sorting on recalled results to raise the display priority of highly relevant delivery configuration documents |
datasetSearchUsingExtensionQuery | Enabled | Automatically generates extended query terms for delivery-related questions, covering targeting settings, budget configurations, and similar content expressed in different ways |
collectionIds | Configured by delivery business grouping | Distinguishes knowledge base collections for different advertising accounts and campaigns, preventing irrelevant cross-business data from being recalled |
Three common configuration errors
- Phenomenon: Knowledge base recall results have low accuracy and fail to match precise delivery configuration fields. Cause: The
similaritythreshold is set too low, filtering some highly relevant structured documents, orusingReRankis not enabled, causing low-relevance results to be returned first. - Phenomenon: Knowledge base requests return the
request entity too largeerror, or anoffset is out of boundserror occurs when uploading a 190MB campaign setup document. Cause: System-level request size limits are not adjusted, or chunking logic is not adjusted for large-volume documents, causing single requests to exceed limits or chunking indexes to go out of bounds. - Phenomenon: After enabling knowledge base association, question-and-answer response speed slows significantly. Cause: The number of recalled
limitentries is too high, orcollectionIdsare not configured by business grouping, causing retrieval across multiple irrelevant datasets and increasing retrieval time.
How to verify correct configuration
- Input a typical delivery configuration question, check whether the
similarityscores of recalled results fall within the preset interval, and confirm that irrelevant content is filtered out. - Upload a large-volume campaign setup document, verify that indexing completes normally with no error prompts.
- Compare question-and-answer response times with and without knowledge base association, confirm that time spent falls within an acceptable range.
- View the source datasets of recalled results, confirm that only delivery-related documents under the current business grouping are included, with no cross-business irrelevant data.
The values provided here 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.