What the data for this category looks like
Documents in this category fall into four types: backup policies, recovery drill reports, disaster recovery switchover plans, and data validation records. Fields include backup window (unit: seconds), retention period (unit: days), recovery point objective RTO (unit: seconds), drill completion status, and more. Data sources include backup systems, database transaction logs, object storage, and disaster recovery management platforms.
Update frequencies vary by document type: backup logs are updated daily, drill reports are archived monthly, and disaster recovery plans are revised quarterly.
Constraints on knowledge base retrieval and recall from these characteristics
Differences in format across multi-source data require retrieval to adapt to varying log and document structures, to avoid invalid recall caused by field mismatches. Differing update frequencies must align with the knowledge base’s incremental sync strategy: daily-updated backup logs require daily incremental sync, monthly-updated drill reports can be synced incrementally weekly, and quarterly-updated plans can be fully updated monthly.
Specific business fields such as RTO and RPO require precise matching to avoid recalling irrelevant general operation and maintenance documents. The recovery point confirmation step also requires quick location of corresponding drill reports and backup policy content, so the relevance and quantity of recall results must be strictly controlled.
How to configure settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches key terms in backup policies and drill reports, filters irrelevant general operation and maintenance content |
limit | Top 8 results | Adapts to the length of disaster recovery drill documents, keeps context length within system requirements |
searchMode | Hybrid search | Matches both keywords and semantics, covers structured backup window fields and unstructured drill steps |
usingReRank | Enabled | Re-ranks recall results by disaster recovery scenario relevance, corrects bias from pure semantic search |
datasetSearchUsingExtensionQuery | Enabled | Generates extended queries based on original questions, supplements disaster recovery professional terminology to improve recall accuracy |
collectionIds | Bound to disaster recovery-exclusive knowledge base | Only retrieves dedicated documents such as backup policies and drill reports, eliminates interference from other business knowledge bases |
Three common mistakes
- Symptom: Knowledge base chat works normally during debugging, but public links or chat menus throw
common:core.chaterrors. Cause: Knowledge base access permissions are not configured, or thecollectionIdsparameter is not correctly bound to the public application. - Symptom: Custom reference templates and prompt words cannot be edited. Cause: The installed FastGPT version is lower than v4.17.0; advanced configuration features are only available in this version and later.
- Symptom: Hybrid search configuration does not return results related to data validation records. Cause:
datasetSearchUsingExtensionQueryis not enabled, and professional terminology for disaster recovery scenarios is not supplemented, so keywords in validation records cannot be matched.
How to confirm configuration is complete
- Check if the
usingReRankfunction is enabled, and verify that recall results are re-ranked by relevance in test conversations. - Verify that the
collectionIdsparameter is only bound to the disaster recovery-exclusive knowledge base, and that retrieval results do not include irrelevant business documents. - Trigger an incremental sync task, confirm that daily-updated backup logs and monthly-updated drill reports are updated to the knowledge base on schedule.
- Confirm that the current FastGPT version is v4.17.0 or higher, allowing normal editing of custom prompt words and reference templates.
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.