FastGPT v4.17.0
What data for this category looks like
This category of data mainly comes from VPC planning documents exported from cloud platform consoles, security group rule configuration files, load balancer traffic monitoring logs, customer architecture topology diagrams, and change orders. Data updates follow each network change or ticket completion. Single data entries include fields such as network segment number, port range, health check threshold, access policy ID, etc. Units include bit rate, milliseconds, port numbers, etc. Most document structures combine structured configuration items and unstructured topology descriptions. The most time-consuming links in daily operations, such as network segment planning and access rule verification, all rely on this type of data.
What constraints do these characteristics impose on the reranking and hybrid retrieval link
Dispersed data sources lead to large differences in recall result formats. Structured configuration items and unstructured topology content require differentiated weights. High update frequency means retrieval must adapt to the latest data after real-time changes to avoid recalling expired configurations. Single data entries have many closely related fields, so hybrid retrieval must match multi-dimensional keywords such as network segments, ports, and health check rules at the same time. The timeliness requirement for results in operation and maintenance scenarios is high. Reranking must prioritize matching change orders associated with current tickets and real-time monitoring data. Incorrect recalled data may cause access failures, traffic forwarding errors, or service exposure risks.
How to Configure
| Configuration Item | Recommended Practice | Rationale |
|---|---|---|
usingReRank | Enabled, connect Jina Rerank service via custom channel | Adapt to the sorting requirements of multi-field associated data for this category, prioritize matching core configuration items associated with tickets |
similarity | 0.72–0.85 | Balance matching accuracy between structured configuration fields and unstructured topology text, avoid mistakenly recalling irrelevant network segment rules |
limit | Top 10 entries | Control the number of single recall results, adapt to the needs of quickly verifying configurations associated with a single ticket in operation and maintenance scenarios |
searchMode | Hybrid retrieval | Combine keyword matching and semantic retrieval, cover multiple types of data such as network segment planning, port rules, and health checks |
Three Common Configuration Mistakes
- Symptom: Calls to the rerank service return a 500 status code or empty results. Cause: Custom channel parameters for the rerank service are not configured correctly, or the corresponding service key is not bound in the FastGPT backend.
- Symptom: Hybrid retrieval results include expired security group rules or offline load balancer configurations. Cause: No filtering logic based on update time is set, and unsynchronized historical data is recalled.
- Symptom: Semantic retrieval results outperform hybrid retrieval, and parameter adjustments do not improve outcomes. Cause: The
usingReRankconfiguration is not enabled correctly, the hybrid retrieval does not trigger the reranking link, and only relies on basic semantic matching.
How to Confirm Successful Configuration
- Enter the knowledge base configuration page of FastGPT, check the
usingReRankswitch status and the bound channel for the rerank service, confirm that it is enabled and associated with a valid service. - Submit a test ticket containing network segment and port keywords, check the sorting priority of retrieval results, confirm that core configuration items are ranked first.
- Compare retrieval results with and without reranking enabled, confirm that matching accuracy and relevance of results improve significantly after enabling.
- Check the retrieval log, confirm that
searchModeis set to hybrid retrieval, and the number of recalled results meets the threshold set bylimit.
The values provided are common starting points and should be measured against relevant deployment 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.