What the data for this category looks like
The core data for this category comes from raw data exported by evaluation tools, project interview records, asset system inventories, and security device logs. Standardized documents routinely generated include classification determination reports, gap analysis reports, rectification plans, and evaluation reports. Each document contains fixed fields: control item number, compliance requirement, gap description, and rectification measures. Data updates follow individual project cycles. Within a single project cycle, updates only occur alongside rectification progress and evaluation results. No high-frequency iteration is needed.
Constraints for Knowledge Base Retrieval and Recall
Data is updated per individual project, and document structures are fixed. The knowledge base does not require frequent full synchronization. Only incremental updates are needed after a project is completed. Documents contain large amounts of standardized control item content. Retrieval must accurately match control item numbers and compliance requirement fields to avoid irrelevant results. Individual documents have long content. It is necessary to limit the length of recalled fragments reasonably. This ensures retrieval results cover core rectification and evaluation links, while preventing redundant information from occupying context space. In internal ticket and operations on-duty scenarios, most retrieval needs focus on specific issues of the current project. Prioritize recalling project-specific document content.
The following configurations are built for FastGPT v4.17.0, and adapted to knowledge base retrieval needs for cybersecurity classified protection evaluation consulting scenarios.
How to Set the Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirements for standardized control items in classified protection documents, balancing recall precision and coverage |
limit | Top 8 entries | Covers main control items and document fragments under a single project, balancing retrieval efficiency and result completeness |
usingReRank | Enabled | Performs secondary sorting on fragments split from long documents, improving the recall priority of content related to control items |
searchMode | vector + hybrid | Combines semantic vectors and keyword matching, covering both structured fields such as control item numbers and unstructured content such as rectification descriptions |
collectionIds | Bound to the current project | Restricts retrieval scope to the project knowledge base corresponding to the current operations ticket, avoiding interference from cross-project documents |
datasetSearchUsingExtensionQuery | Enabled | Adds contextual descriptions related to the project, improving recall accuracy for specific ticket questions |
Three Common Configuration Mistakes
- Abnormal retrieval scores over 4000 may appear. This happens when the normalization logic for the
similarityparameter is not configured correctly, or non-standard similarity calculation methods are used. - Specified knowledge base documents cannot be retrieved. This occurs when the corresponding project's knowledge base collection is not bound via
collectionIds, or documents are not imported correctly into the specified collection. - Complex tables cannot be retrieved properly. This happens when the semantic association of table structures is not retained during document splitting, causing fragment matching failures.
How to Verify Configurations Are Correct
- Manually enter a control item number from a classification determination report, and confirm the retrieval result includes the corresponding fragment.
- Check the
similarityscore of the retrieval result, and confirm the value falls within the 0-1 range. - Pass in the
collectionIdsof the specified project, and confirm retrieval results only include documents from that project. - Submit a query containing complex tables, and confirm the recalled results include the table's core content.
The values provided here are common starting points, and should be evaluated against the user'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.