What this category of data looks like
This category of data primarily comes from Prometheus, Zabbix, cloud monitoring, application logs, and CMDB systems. Daily maintenance documents include monitoring metric lists, alert rule configuration files, on-duty handbooks, and alert change records. Data updates occur daily or after each alert rule change. Document fields include metric name, collection cycle, threshold parameters, notification links, scheduling periods, and more. Some metrics have clear units, such as CPU usage percentage and response time in milliseconds.
What constraints do these characteristics impose on the reranking and hybrid retrieval link
Multi-source heterogeneous data sources require hybrid retrieval to recall content from different systems simultaneously, including Prometheus metrics, Zabbix alert rules, and CMDB configurations. The retrieval mode must support multi-source recall. Data is updated daily or after configuration changes. The reranking link must align with data timeliness to ensure recalled content matches currently active configurations. Documents include professional fields like thresholds and notification links. The reranking model must prioritize matching core fields related to the query, to avoid irrelevant content interfering with ranking results. The accuracy of alert configurations directly affects on-call duty efficiency. The ranking accuracy of retrieval results must align with the decision-making needs of operation and maintenance scenarios.
How to configure the settings
This document is based on FastGPT v4.17.0 configuration guidelines.
| Configuration Item | Recommended Value | Description |
|---|---|---|
| Reranking Toggle | Enabled | Alert configuration retrieval results must accurately match the core needs of operation and maintenance queries; reranking filters low-relevance content |
| Similarity Threshold | 0.75–0.85 | Balances precise matching of alert configurations and recall coverage, to avoid missing critical threshold configurations |
| Single Recall Limit | Top 8–12 entries | Adapts to the quick review needs of on-call duty, to avoid excessive redundant content interfering with decision-making |
| Retrieval Mode | Multi-source Hybrid Recall | Adapts to the data characteristics of multi-system sources for this category, integrating content from different monitoring and configuration platforms |
| Reranking Call Timeout | 30 seconds | Adapts to the retrieval response requirements of alert configuration data, to avoid excessive wait times affecting on-call efficiency |
Three common configuration errors
- The interface displays
Invalid URL (POST /v1/rerank)or returns a 401 unauthorized error. This occurs because the reranking model access key is configured incorrectly, or the deployed reranking service address was not correctly entered in the corresponding FastGPT settings. - Ollama-deployed reranking models cannot connect properly. This occurs because the reranking model's API path does not match the standard path required by FastGPT, leading to connection failure.
- Retrieval results do not cover core alert configuration fields. This occurs because multi-source hybrid recall mode is not enabled, so only monitoring data from a single system is recalled, and alert rules and CMDB configuration content are not integrated.
How to verify a successful configuration
- The FastGPT knowledge base configuration page is reviewed to confirm the reranking toggle is enabled and the retrieval mode is set to multi-source hybrid recall.
- A common operation and maintenance query, such as "server CPU alert threshold configuration", is entered, and the ranking of returned results is checked to confirm it prioritizes core fields such as thresholds and notification recipients.
- Reranking model call logs are reviewed to confirm the API request address and key configuration are correct, with no error returns.
- The single recall limit is set to the specified value, and the number of returned retrieval results is checked to match the set limit.
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.