What the data for this category looks like
Core users of this category include DBAs, data platform operations teams, and backend teams. Daily processed documents include database change orders, slow SQL reports, backup records, and permission application forms. Data sources primarily come from database monitoring systems, audit logs, SQL execution plan files, and internal ticketing systems. Update frequency is daily, or triggered each time a database change occurs. Most document structures include fields such as instance identifier, timestamp, and operation type. Slow SQL reports include quantitative metrics like execution duration (milliseconds) and number of scanned rows (rows). Change orders note the count of affected instances (units).
What constraints these characteristics impose on knowledge base retrieval and recall
Data sources cover structured monitoring logs, semi-structured tickets, and execution plans. This requires retrieval to support both keyword matching and semantic association, to avoid missing key execution details. Update frequency is triggered by changes, so the recall link must support incremental synchronization without fully rebuilding the knowledge base. Documents are strongly bound to specific database instances. Retrieval must filter scope by instance ID to prevent irrelevant cross-instance content from interfering. Slow query troubleshooting and change review are core time-consuming scenarios. Recall results must accurately match business keywords to reduce context overhead from invalid information.
How to Configure Settings
These configurations are based on FastGPT v4.17.0 knowledge base retrieval logic:
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.7–0.85 | Matches the precision of database operation documents, filters redundant results with large semantic deviations |
limit | Top 10–15 results | Adapts to the number of documents in single scenarios such as slow SQL reports and change orders, avoids excessive results increasing context processing overhead |
searchMode | Semantic + keyword hybrid | Matches both SQL keywords and business semantics, meets the dual retrieval needs of slow query troubleshooting and change review |
usingReRank | Enabled | Performs secondary sorting on initial recall results to filter duplicate or low-relevance backup records and permission application forms |
datasetSearchUsingExtensionQuery | Enabled | Expands the retrieval scope to associated tickets and monitoring logs, such as permission application records linked to change orders |
collectionIds | Bind the knowledge base ID of the current operation instance | Limits the retrieval scope to only documents for the database instances currently being managed, preventing cross-instance interference |
Three Common Misconfigurations
- Empty retrieval results: The phenomenon is an empty array returned when calling the knowledge base file list interface. The cause is incorrect configuration of
collectionIds, which fails to limit the retrieval knowledge base scope. - Unreachable original document links: The phenomenon is a 404 error returned when clicking the original link of retrieval results. The cause is failure to update nginx proxy rules in a timely manner, causing the download link path to not match the actual storage path.
- Excessive redundant retrieval results: The phenomenon is a large number of irrelevant backup records returned during slow query troubleshooting. The cause is the
similaritythreshold set below 0.7, which matches documents with similar semantics but irrelevant to the business.
How to Verify Proper Configuration
- Enter the core keywords of slow SQL, execute retrieval, and check if returned results include relevant documents such as slow SQL reports and SQL execution plans.
- Call the knowledge base file list interface with the specified
collectionIds, confirm that returned results only include knowledge base files for the bound instance, with no content from other instances. - After enabling
usingReRank, compare the number of initial recall results and rearranged results, confirm that redundant content is effectively filtered. - Submit a small database change, wait for the knowledge base incremental update to complete, retrieve keywords for the newly generated change order, and confirm the new document can be normally recalled.
The values listed above 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.