What the Data for This Category Looks Like
The core data for this category comes from risk control systems, audit records, user behavior logs, and report evidence. Regularly maintained documents include structured ban records, appeal materials, review opinions, and user notifications. Data is updated daily or synchronized in real time. Fields include account unique identifier, violation type code, appeal submission time, review disposal result, associated evidence file identifier, and more. Units vary by field attributes, with no unified fixed standard.
Constraints on Knowledge Base Retrieval and Recall
The real-time or daily update nature of the data requires the retrieval pipeline to support incremental synchronization or scheduled full updates, preventing the recall of outdated ban or appeal information. The mixed structured and unstructured document structure requires retrieval to adapt to both field matching and semantic understanding, covering precise queries for structured records and semantic matching for unstructured appeal materials. The account-level business focus requires retrieval results to be bound to the target account ID, excluding violation data from unrelated accounts. Additionally, errors in this category directly affect user rights, so retrieval and recall results must have high matching confidence to avoid irrelevant content interfering with review judgments.
Configuration Settings
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category requires precise matching of violation evidence and appeal reasons, to avoid low-confidence content interfering with review judgments |
limit | Top 8–12 results | Balances the amount of recalled information and retrieval latency, covering the associated evidence and historical records required for a single appeal review |
searchMode | Semantic retrieval + keyword hybrid | Adapts to keyword matching for structured records and semantic understanding requirements for unstructured appeal materials |
usingReRank | Enabled | Performs re-ranking on initial recall results to raise the display priority of high-matching content |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms for appeal reasons, covering associated violation scenarios not directly mentioned by the user |
collectionIds | List of collection IDs corresponding to the ban and appeal knowledge base | Limits the retrieval scope to only the document data exclusive to this category, avoiding recall of irrelevant content |
The knowledge base retrieval pipeline in FastGPT v4.17.0 supports the above configuration items.
Three Common Misconfigurations
- Semantic retrieval returns scores outside the 0-1 range, such as results with thousands of points. This occurs when the similarity threshold verification logic is not correctly configured, and non-normalized similarity calculation methods are used.
- Associated file information from the knowledge base is not returned after calling the API. This occurs when the knowledge base's file traceability configuration is not enabled, or the correct
collectionIdsparameter is not specified in the request. - The number of retrieved and recalled results does not match the configured
limit. This occurs when multiple retrieval modes are enabled simultaneously, leading to merged results exceeding the limit, or when parameter configurations are not synchronized to the retrieval engine.
How to Verify Proper Configuration
- Submit a test retrieval request containing the target account ID, and check whether the similarity scores of the returned results fall within the 0-1 range.
- Call the API and check whether the returned results include the account's historical ban records, appeal materials and other associated content.
- Manually update a ban record, run a retrieval request, and confirm that the newly added record has been recalled.
- Adjust the
similaritythreshold range, and observe whether the matching accuracy of the retrieval results changes as expected.
The values provided are common starting points and should be measured against samples relevant to the deployment context.
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.