What data for this category looks like
Core data for this category comes from configuration files exported from cloud platform consoles, network topology diagrams, real-time traffic monitoring reports, and architecture diagrams submitted by customers. Daily maintenance documents include VPC planning plans, security group rule lists, load balancing configuration sheets, and change approval forms. Data update frequency syncs with business change tickets, and updates are typically made per ticket or each configuration adjustment. Data fields include subnet masks, port numbers, health check intervals, backend server weights, and similar items. Units mostly follow subnet mask formats, port number conventions, seconds, percentages, and other standard units.
Constraints on reranking and hybrid retrieval
Cloud network and load balancing configuration data mostly consists of structured rule-based documents, with concentrated fields and strong correlations. Hybrid retrieval must set categorized recall rules for different types of configuration documents to avoid mixing unrelated security group rules and load balancing configurations. Data update frequency is high and syncs with change tickets. The retrieval system must filter expired configuration documents in real time. Failure to do so will return invalid network rules, leading to customer access exceptions. Errors in this category directly affect business connectivity. The reranking link must prioritize matching high-priority configuration fields, and control total retrieval and reranking latency to avoid exceeding customer waiting thresholds. Local deployment of reranking models consumes large amounts of GPU video memory. Lightweight API-called models should be prioritized to reduce resource consumption.
How to set configurations
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Hybrid retrieval recall count | Top 8-12 entries | Cloud network configuration documents have a limited number of entries. Excessive recall will increase total response latency |
| Reranking model call timeout | 1200 milliseconds | Matches the real-time response requirements of cloud network customer technical support, to avoid exceeding business waiting thresholds |
| Reranking TopK value | Top 3-5 entries | Key rules for cloud network configurations are concentrated in the top 3-5 entries. Excessive values will increase GPU video memory usage |
| Retrieval data source filter | Only include change tickets from the past 7 days and configuration documents for the current project | Cloud network configuration changes occur frequently. Expired data will lead to invalid rule matching |
| Hybrid retrieval similarity threshold | 0.75-0.85 | Keyword matching accuracy requirements for network configurations are high. Too low a threshold will introduce unrelated configuration rules |
| Reranking model call method | Lightweight API model | Avoids excessive GPU video memory usage caused by local deployment, and adapts to the elastic resource model of cloud services |
Three Common Configuration Errors
- Phenomenon: GPU video memory remains occupied at 6-7 GB after reranking calls, and multiple GPU nodes show this occupation. Cause: A lightweight reranking model is not selected, or the reranking TopK value for single requests is not limited, leading the model to load too many parameters.
- Phenomenon: Total hybrid retrieval response time exceeds 10 seconds, exceeding business waiting thresholds. Cause: Hybrid retrieval recall count and reranking TopK value are not limited, or expired cloud network configuration documents are not filtered, leading to an overly large retrieval scope.
- Phenomenon: An offline reranking model access prompt shows rerank error, returning an error message containing the
detailfield. Cause: The model path and environment variables are not configured according to the FastGPT v4.17.0 offline model deployment documentation, or model files lack dependent packages.
How to Confirm Proper Configuration
- Review reranking model call logs to confirm that each call’s response time meets the preset timeout configuration.
- Test retrieval of different types of cloud network configuration documents to confirm that all hybrid retrieval recalled documents are valid configurations for the current project.
- Review GPU video memory usage to confirm that single request calls do not exceed the business allowable video memory threshold.
- Trigger a test request for the offline reranking model to confirm that no rerank error type errors are returned.
The values provided here are common starting points, and should be measured against samples specific to the 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.