What the Data for This Category Looks Like
Core data for this category comes from content data, advertising revenue sharing systems, financial systems, and creator account backends. Daily operation documents include incentive rules, settlement details, appeal records, and payment lists. Data update cycles are mostly monthly or triggered by operational activities. The document structure primarily uses structured fields, including creator ID, revenue amount, settlement cycle, appeal status, and other fields. Units include yuan, number of people, cycle identifiers, and there is no unified fixed format.
Constraints Imposed on Knowledge Base Retrieval and Recall
Decentralized data sources require the knowledge base to integrate heterogeneous data from multiple systems. Clear field mapping rules for different data sources must be defined to prevent mismatched revenue metrics. Unfixed update frequencies require support for switching between incremental and full updates, to ensure retrieved settlement rules and details are always the latest versions. Documents contain sensitive financial and creator personal information, so permission scope must be strictly controlled during retrieval to avoid data leaks. A large number of structured fields with significant differences require targeted word segmentation and recall rules to ensure accurate matching for revenue metric verification and abnormal fraud investigation. The abnormal investigation and verification process relies on precise document recall, so the redundancy of recall results must be limited to avoid interfering with core verification workflows.
Configuration Settings
The following configurations are based on FastGPT v4.17.0
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches precise fields related to finance and creators, prevents low-relevance settlement rules from being included in results |
limit | Top 8 entries | Covers details, rules, and appeal records required for a single settlement verification, avoids excessive results consuming context space |
searchMode | Semantic Retrieval | Adapts to matching between structured fields and natural language queries, fits revenue metric verification scenarios |
usingReRank | Enabled | Re-sorts recall results by relevance, improves the priority of critical documents during abnormal fraud investigation |
datasetSearchUsingExtensionQuery | Disabled | Prevents extended query processing logic from damaging the original document format retention requirements, ensures accurate recall of original incentive rule text |
collectionIds | Knowledge base collection IDs bound to the current operational activity | Restricts retrieval scope to only incentive settlement-specific documents for the current activity, excludes irrelevant historical data |
Three Common Misconfigurations
- Retrieval results contain extra spaces or uppercase letter tampering. The cause is that
datasetSearchUsingExtensionQueryis not disabled, and extended query processing logic breaks the original document format retention requirements. - Retrieval results omit the latest settlement rule documents. The cause is that incremental update trigger logic is not configured, so the knowledge base does not synchronize monthly updated incentive rule data.
- Recall results include irrelevant creator onboarding documents. The cause is that
collectionIdsis not configured correctly, and the retrieval scope is not limited to the incentive settlement-specific knowledge base collection.
How to Verify Successful Configuration
- Submit a query containing "settlement rules" or "revenue metrics", verify that the document source of the recall results matches the collection bound to
collectionIds. - Enter a query containing sensitive financial fields, check that returned results fall within the preset
similaritythreshold range, with no low-relevance content. - Trigger a knowledge base update operation, confirm that newly uploaded settlement detail documents can be retrieved normally after the update.
- After enabling
usingReRank, verify that the returned result sorting prioritizes documents matching core investigation requirements.
The values provided are common starting points and should be evaluated 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.