What data for this category looks like
The data for this category mainly comes from cloud database consoles, slow logs, monitoring metrics, and internal ticket systems. Daily processed documents include instance change orders, backup and recovery records, performance diagnosis reports, and parameter adjustment records. Fields include instance ID, change time, backup capacity, slow SQL execution time, ticket number, and more. Units involve GB, milliseconds, and count. Data update frequency is daily or after each configuration change. Individual document lengths range from hundreds to thousands of characters.
What constraints do these characteristics impose on the reranking and hybrid retrieval link
Data for this category comes from multiple systems. Hybrid retrieval must recall heterogeneous data such as console configurations, slow logs, and tickets at the same time, and requires adapting to field mapping across different data structures. Individual document lengths vary widely. Long documents may exceed the input character limit of reranking models, so truncation must be performed in advance. Data update frequency is high, so retrieval must match real-time changed instance configurations and ticket statuses. Professional fields such as slow SQL execution time and backup capacity need reranking to filter highly relevant results. Multi-tenant data isolation also requires that retrieval results strictly match the permission scope of the corresponding tenant.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Similarity threshold | 0.65–0.75 | Filter low-relevance heterogeneous data, matching the strict relevance requirements of database operation and maintenance scenarios |
| Single recall limit | Top 10 entries | Control the number of initial recall results to avoid overloading the reranking model input |
| Retrieval mode | Hybrid retrieval | Adapt to the heterogeneous data recall requirements across consoles, slow logs, and tickets |
| Enable reranking | Yes | Reorder recalled heterogeneous results by relevance, improving matching accuracy for professional fields |
| Number of entries to retain after reranking | Top 5 entries | Calibrated based on actual testing; only retain the most relevant results after reranking, adapting to the need for quick problem location in operation and maintenance scenarios |
Three common misconfigurations
- The symptom is that the reranking status field in the retrieval result shows false, and the returned results are not sorted by reranking. The cause is that the corresponding reranking model is not bound in the retrieval workflow, or the reranking model's interface configuration is not correctly associated with the FastGPT retrieval node.
- The symptom is that when the retrieval content exceeds 2000 characters, the FastGPT backend log shows that the reranking model call is not triggered. The cause is that the long document automatic truncation configuration is not enabled, or the input length limit of the reranking model does not adapt to the long document data of this category.
- The symptom is that the Qwen3-Reranker-8B option cannot be found in the reranking model selection dropdown list. The cause is that FastGPT v4.17.0 does not have built-in adaptation configuration for this model currently, and the model deployment path and call parameters need to be added manually.
How to confirm the configuration is complete
- Enter the FastGPT retrieval test page, enter a query term related to database slow SQL, and check whether the reranking status identifier in the retrieval result is true.
- Upload a performance diagnosis report with a length exceeding 1500 characters, initiate a retrieval, and confirm that the reranking model is called normally and generates sorted results.
- Check whether the reranking switch in the retrieval configuration is enabled, and that the reranking model's deployment address and call parameters are configured correctly.
- Compare the initial recall result list and the reranked result list, confirming that results with more relevant professional fields are ranked first.
The values provided here 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.