What the data for this category looks like
Data sources for this category include Prometheus, Zabbix, cloud monitoring services, application log storage, and CMDB systems. Daily maintained documents include monitoring metric lists, alert rule configuration files, on-duty handbooks, and alert change records. Data is updated either daily or in real time when configuration changes are made. Data fields include metric names, threshold ranges, alert levels, trigger conditions, notification link configurations, and more. Some fields have associated numerical attributes such as response duration, error rate, and collection frequency.
Constraints Imposed on Re-ranking and Hybrid Retrieval by These Data Characteristics
Three constraints are imposed on the re-ranking and hybrid retrieval link by the data characteristics of this category. First, data sources are scattered and include both structured metric fields and unstructured document content. Hybrid retrieval must adapt to both data formats to avoid recall bias from single retrieval modes. Second, data update frequencies vary widely. Alert configurations that change in real time require the retrieval pipeline to quickly sync the latest content to avoid recalling outdated rules. Third, the multi-dimensional attributes of fields require the re-ranking model to prioritize matching the operation and maintenance scenarios of user queries, such as threshold adjustments and notification link optimization, to reduce interference from irrelevant alert entries.
How to Set the Configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the semantic similarity of alert configuration text, balances recall precision and coverage |
limit | Top 10 entries | Controls the number of initial recall results, avoids retrieval result overload that impacts response speed |
searchMode | hybrid | Enables hybrid retrieval to handle both structured metric fields and unstructured alert documents |
usingReRank | Enabled | Re-ranks initial recall results to prioritize matching the core alert configuration scenarios of user queries |
| Local re-ranking model inference timeout | 30 seconds | Adapts to the inference latency of locally deployed models, avoids overall retrieval timeout |
Three Common Misconfigurations
- Enabling
usingReRankand optimizing queries leads to retrieval response times exceeding 30 seconds: This occurs because a reasonable timeout threshold for the local re-ranking model is not configured, or the number of initial recall entries is too high, leading to excessive re-ranking computational load. - A request error is displayed after configuring the local re-ranking model: This occurs because the model deployment address is not filled in correctly, the model version is incompatible with FastGPT v4.17.0, or required parameters for model invocation are not configured.
- Enabling the re-ranking switch for knowledge base indexes but seeing no effective results in online recall tests: This occurs because the
usingReRankparameter is not enabled simultaneously in the global retrieval configuration, or the re-ranking model is not adapted to the professional semantic scenarios of alert configuration documents.
How to Confirm the Configuration is Correct
- View the FastGPT retrieval configuration page, confirm that the
usingReRankswitch is enabled andsearchModeis set to hybrid mode. - Initiate an online recall test with keywords related to alert configurations, check whether the returned results are sorted to prioritize matching the core scenarios of the user's query.
- Adjust the
similarityparameter value, observe changes in the number and relevance of recall results to verify that the parameter takes effect. - Test the invocation pipeline of the local re-ranking model, check whether there are prompts for model response timeouts or permission errors in the system logs.
The values provided 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.