What data for this category looks like
Data for this category primarily comes from cloud database consoles, slow logs, monitoring metrics, and ticketing systems. Daily produced documents include instance change orders, backup recovery records, performance diagnostic reports, and parameter adjustment records. A single document typically contains fields such as instance ID, operation time, change content, and execution result. Some monitoring metrics have fixed units like milliseconds or percentage. Data is updated daily or synchronized after each configuration change. The primary storage formats are structured logs and semi-structured ticketing documents.
What constraints these characteristics impose on knowledge base retrieval and recall
Multiple heterogeneous data sources require the retrieval process to adapt to different document formats. It must support both keyword matching for structured logs and semantic recall for semi-structured ticketing documents. Update frequency triggered by configuration changes requires recall logic to prioritize recently updated documents. This avoids using expired instance configuration information. Documents include sensitive fields such as instance ID and tenant identifier. Retrieval must associate tenant permissions for data isolation, preventing cross-tenant data leaks. High-professionalism documents like slow logs and performance reports require higher semantic matching thresholds. This filters low-correlation results. The process must also cover operational domain terminology variants.
How to configure settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | For high-professionalism documents such as slow logs and performance reports, this range balances recall coverage and precision, filtering low-correlation results |
limit | Top 8–12 entries | Operational documents associated with a single instance typically number fewer than 10. This range covers all associated content without exceeding context window limits |
searchMode | hybrid | Adapts to the mixed data characteristics of structured logs and semi-structured ticketing documents. Combines keyword and semantic retrieval to improve recall effectiveness |
usingReRank | Enabled | For field-dense documents such as parameter adjustment records and change orders, re-ranking improves the ranking priority of highly relevant content |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms for operational questions. For example, expands "slow SQL" to "slow query log analysis" to cover terminology variants |
collectionIds | Configured by tenant instance group | Isolates cloud database data for different customers, meeting security isolation requirements for multi-tenant operations |
Three common configuration mistakes
- Phenomenon: After increasing the
limitparameter, the large language model output does not reference any retrieved content. Cause: The total context window length is exceeded, so retrieved document fragments cannot be fully passed to the model, or highly relevant content is crowded out by lower-priority entries after re-ranking. - Phenomenon: Knowledge base queries return results, but the large language model cannot output content. Cause:
usingReRankis not enabled, orsearchModeis configured incorrectly. This causes retrieved document formats to be abnormal and cannot be correctly parsed by the model. - Phenomenon: Answer accuracy is insufficient after local deployment. Cause: The
similaritythreshold is not adjusted for high-professionalism documents such as slow logs and performance reports, ordatasetSearchUsingExtensionQueryis not enabled. This fails to cover operational domain terminology variants.
How to confirm correct configuration
- Log in to the FastGPT v4.17.0 knowledge base management interface. Check if the
searchModeconfiguration for the target dataset is set to hybrid, and confirm it matches the preset configuration. - Enter a test question such as "Query the slow log analysis results for the current instance". Check if retrieved results include slow log documents from the last 3 days, and if the
similaritythreshold matches the expected range. - Adjust
limitto Top 10 entries. Verify that the number of returned referenced content meets expectations, and no context overflow prompts appear. - After enabling
datasetSearchUsingExtensionQuery, enter "Adjust the connection number parameters for the instance". Confirm that retrieved results include the corresponding parameter adjustment record documents.
The values provided are common starting points and should be measured 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.