What the data for this category looks like
This category’s data primarily originates from operational tools including Prometheus, Zabbix, cloud monitoring, application logs, CMDB, and others. Updates occur either daily or immediately after each alert rule change. Daily documents include monitoring metric lists, alert rule configurations, on-call handbooks, and alert change records. Data fields include metric names, threshold parameters, alert trigger conditions, notification recipients, change times, and additional relevant fields. Some metrics have clear units, such as CPU usage percentage and response time in milliseconds.
What constraints do these data characteristics impose on knowledge base retrieval and recall?
These data characteristics create multiple constraints for retrieval and recall workflows. Dispersed multi-source data has significant format differences, including structured metric parameters and semi-structured configuration text. This increases the difficulty of semantic matching and may result in irrelevant content being recalled. The update frequency of daily synchronization or change-triggered updates requires the knowledge base synchronization mechanism to align with change nodes. Otherwise, recalled content will not match the currently active alert configurations. Professional technical configuration text includes a large number of specialized parameters and units. Standard retrieval logic may easily match semantically similar but unrelated entries. Data length varies widely, from a few lines of threshold configurations to complete on-call handbooks. This requires adapting chunking and recall strategies for different lengths to avoid context overload or missing information.
How to configure the settings
These configurations are based on the knowledge base retrieval feature of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the density of professional terminology in alert configurations. A value that is too low introduces irrelevant content, while a value that is too high misses relevant alert rules. |
limit | Top 8 entries | Alert configuration-related documents typically have clear entries. 8 entries cover core alert rules and on-call key points, avoiding context overload. |
searchMode | semantic | For professional configuration text, semantic retrieval better matches implicit threshold logic than keyword retrieval. |
usingReRank | Enabled | Re-ranks recalled results to filter out alert configuration entries with weak semantic relevance. |
datasetSearchUsingExtensionQuery | Enabled | Generates extended queries based on the original question to cover common synonyms in alert configurations, such as "threshold" and "trigger condition". |
collectionIds | Bind the dedicated alert configuration knowledge base | Avoid retrieving documents from other categories, ensuring recalled content only comes from datasets related to monitoring alerts. |
Three common configuration errors
- Retrieval results include content unrelated to the current alert configuration. For example, asking "How to adjust CPU alert thresholds" returns an on-duty schedule. This occurs because the
similaritythreshold is set too low, causing low-relevance documents to be incorrectly recalled. - Mathematical formulas in alert configurations fail to render properly. This occurs because the knowledge base chunking process did not preserve the complete structure of the formulas, or did not enable parsing adaptation for the corresponding format.
- Retrieval results for the same alert-related question are inconsistent across multiple attempts. This occurs because the
limitparameter is not fixed, or the locally deployed model does not use a unified embedding generation logic, leading to differences in recalled entries.
How to confirm the configuration is successful
- Submit a test question targeting alert threshold configuration, and verify that the relevance of the recalled results meets preset requirements.
- Check the knowledge base synchronization records to confirm whether the dataset was updated promptly after an alert configuration change.
- Submit the same alert-related question multiple times, and verify the consistency of the recalled results.
- Test a query for an alert rule containing mathematical formulas, and confirm that the formulas render correctly in the output.
The values provided are common starting points and should be evaluated against samples specific to the target deployment environment.
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.