What data this category includes
This category of data originates from operational maintenance data sources including Prometheus, Zabbix, cloud monitoring, application logs, CMDB, and others. Four types of documents are routinely generated: monitoring indicator lists, alert rules, on-duty manuals, and alert change records. Monitoring indicator lists include indicator names, collection cycles (unit: seconds), and threshold ranges. Alert rules include trigger conditions, notification targets, and effective time periods. Alert change records include fields such as change time, change content, and operator. The data update frequency is daily or after each alert configuration change.
What constraints these characteristics impose on knowledge base retrieval and recall
The multi-source nature of alert configuration data requires retrieval to match the field formats of different data sources, to avoid incorrect recall of cross-source documents. The uncertain update frequency requires the knowledge base to be configured with on-demand or scheduled synchronization mechanisms, to ensure recalled data aligns with the latest configurations. The diversity of document types requires setting retrieval weights by document category, prioritizing the recall of alert rules and indicator lists. The clear units and threshold requirements of fields require automatic unit matching during retrieval, to prevent invalid recall caused by unit mismatches.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.7–0.85 | Alert configuration has high field precision requirements, so low-match irrelevant operational documents must be filtered out |
limit | Top 8–12 entries | The number of associated documents for alert configurations is moderate; an excessive number will exceed the context window limit |
searchMode | vector | Alert configurations have standardized text structures, so semantic retrieval delivers better matching accuracy than keyword retrieval |
usingReRank | Enabled | Further filter noisy alert rules in retrieval results to reduce false positive interference |
datasetSearchUsingExtensionQuery | Enabled | Alert configurations contain many specialized operational terms, so extended queries improve recall coverage |
collectionIds | Bind dedicated alert configuration knowledge base collections | Isolate business data to avoid irrelevant documents interfering with accurate alert configuration retrieval |
Three Common Configuration Errors
- Symptom: After calling the knowledge base retrieval API, the returned results do not include the identifiers or paths of associated files. Cause: The
datasetSearchUsingExtensionQueryparameter is not enabled, or file metadata synchronization for the knowledge base collection is not configured. - Symptom: Semantic retrieval scores exceed the normal 0–1 range, with values in the thousands. Cause: The normalization logic of the embedding model is not properly adapted, or a non-standardized similarity calculation method is incorrectly used.
- Symptom: After configuring the searchXNG retrieval engine, alert configuration-related documents cannot be retrieved in some scenarios. Cause: No data source filtering rules are configured for searchXNG, or the latest alert configuration documents are not synchronized to the retrieval engine.
How to Confirm Configuration is Complete
- Call the FastGPT v4.17.0 knowledge base retrieval API, pass query text related to alert configurations, and check if returned results include expected documents such as alert rules and indicator lists.
- View the
similarityfield of retrieval results, confirm scores fall within a reasonable range with no abnormally high or low values. - Verify that after the
usingReRankparameter is enabled, the proportion of noisy documents in retrieval results has decreased significantly. - Check the
collectionIdsparameter configuration, confirm only dedicated alert configuration knowledge base collections are bound, with no irrelevant business collections included.
The values provided here 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.