What this category’s data looks like
This category’s data primarily originates from CRM systems, contract orders, identity management systems, and product backends. The update frequency is per order or daily. Daily produced documents include onboarding application forms, tenant configuration sheets, administrator account lists, and delivery confirmation documents. Data fields include tenant unique identifiers, customer entity names, basic configuration statuses, administrator account lists, onboarding timestamps, and other relevant fields. The data structure is built around operation and status records for the full tenant onboarding workflow.
Constraints imposed on knowledge base retrieval and recall
Data sources are scattered and linked to the full tenant onboarding workflow. Retrieval must accurately match tenant-specific fields to prevent cross-tenant data mixing. Update frequency varies per order, so retrieval data sources must support real-time or near-real-time synchronization. Failure to do so will result in discrepancies between configuration statuses and actual conditions. Diverse document types require differentiated field weight configuration for different document categories. For example, administrator account lists should prioritize matching account fields. Data volume grows with the number of customers, so retrieval and recall processes must balance accuracy and performance to avoid response delays caused by excessive data volume.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches structured data related to tenant onboarding. This range filters low-relevance non-tenant documents while retaining valid matching results |
limit | Top 8 entries | The volume of onboarding workflow data for a single tenant is limited. Limiting the number of returned entries reduces redundant calculations and improves response speed |
usingReRank | Enabled | Tenant data fields are relatively fixed but scenario details vary. The re-ranking model optimizes the sorting logic of initial recall results to improve matching accuracy |
datasetSearchUsingExtensionQuery | Enabled | Documents such as onboarding application forms contain many business abbreviations. Query optimization adds contextual association information to reduce ambiguity |
collectionIds | Specify tenant-specific knowledge base collections | Prevent cross-tenant data from being included, ensuring the retrieval scope only covers onboarding-related data for the current customer |
The above configurations are based on FastGPT v4.17.0
Three common configuration mistakes
- Retrieval results do not match the tenant onboarding workflow, returning irrelevant general documents. Cause: The
collectionIdsconfiguration to limit the retrieval scope was not set, resulting in cross-tenant data being recalled. - Semantic retrieval scores fall outside the 0-1 range, with abnormal score values displayed on the interface. Cause: The threshold verification logic for
similaritywas not properly configured, leading to unnormalized recall scores. - Knowledge base retrieval response times out, with a single question answer taking more than 15 seconds. Cause: The number of returned entries for
limitwas not restricted, and batch processing forusingReRankwas not enabled, resulting in excessive redundant data calculation volume.
How to confirm configurations are correctly set
- Run a test retrieval, enter a query related to tenant onboarding, and verify that returned results only include onboarding application forms, configuration sheets, and other documents for the current customer.
- Check retrieval logs to confirm that
similarityscores fall within the preset range, with no results that exceed the range abnormally. - Simulate multi-round queries to verify that response times meet business requirements with no significant delays.
- After enabling
usingReRank, compare initial recall results with the sorted logic after re-ranking to confirm that the sorting rule is active.
The configuration values listed here are common starting points, and should be evaluated against the relevant organization’s own specific 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.