What the data for this category looks like
All data for this category comes from real-time or event-triggered business assets from frontline operations teams. Data sources include monitoring alerts, log platforms, APM tools, and user feedback channels. Daily maintained documents include fault tickets, event timelines, emergency response plans, and fault notifications. Core fields for individual data entries include incident number, impact scope, trigger time, associated log ID, and troubleshooting steps. There is no fixed length limit, and data is updated in real time as the incident progresses. Data synchronization triggers each time a new incident occurs or a troubleshooting action is completed.
What constraints these characteristics impose on the reranking and hybrid retrieval link
The decentralized and real-time updated nature of data sources requires hybrid retrieval to cover multi-source heterogeneous monitoring, log, and user feedback data simultaneously, to avoid recall bias from single data sources. Documents include core fields such as incident number, impact scope, and trigger time. The reranking link must prioritize matching fields strongly related to the current fault incident; general semantic matching alone cannot meet accuracy requirements. The real-time update attribute requires retrieval to support incremental recall logic, and the reranking model must adapt to high-priority incident data within short cycles. The operations on-duty scenario prioritizes the timeliness of fault recovery, so the total time taken for retrieval and reranking must align with the fault handling response window.
How to set the configurations
The following configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Similarity threshold | 0.72–0.88 | Filter redundant data with low relevance to the current fault incident, retain core matching results |
| Single recall limit | Top 6–10 entries | Balance recall coverage and total retrieval and reranking time, align with the response rhythm of operations on-duty |
| Retrieval mode | hybrid | Cover both keyword recall and semantic recall, adapt to the characteristics of multi-source heterogeneous operations data |
| Reranking switch | Enabled | Use precise semantic matching from the reranking model to optimize the relevance ranking of initial recall results |
| Reranked recall count | Top 3–5 entries | Control the computational load of the reranking link, avoid excessive video memory usage |
| Retrieval timeout threshold | 600 seconds | Align with the fault handling response window, avoid excessive waiting that affects on-duty efficiency |
Three common mistakes
- Symptom: The reranking model continues to occupy video memory after retrieval completes, and system monitoring shows video memory usage is not released after retrieval ends. Cause: The automatic unloading logic for the reranking model is not configured, so the model is not cleared from video memory after retrieval finishes.
- Symptom: After enabling question optimization and result reranking, message return time exceeds 30 seconds. Cause: The single recall limit is set too high, or the reranked recall count exceeds a reasonable range, leading to excessive consumption of computing resources.
- Symptom: A reranking model is configured and knowledge base index settings are enabled, but reranking results do not take effect during online recall testing. Cause: The reranking switch is not enabled during the retrieval link, or the path configuration for the reranking model is incorrect, so the reranking logic is not triggered.
How to confirm the configuration is complete
- Access the knowledge base retrieval test page in FastGPT, input a fault-related query, and verify returned results are sorted by the priority of fault relevance.
- Review the system monitoring panel to confirm whether the reranking model is automatically released from video memory after retrieval completes, with no persistent occupation.
- Inspect the reranking switch and retrieval mode parameters on the knowledge base configuration page, confirm they match the preset configuration items.
- Simulate a fault ticket query scenario, compare result differences before and after enabling reranking, and confirm the reranking logic has taken effect.
The values provided are common starting points and should be measured 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.