What Data for This Category Looks Like
Data sources include cloud bills, monitoring systems, business forecast reports, and stress test results. Documents update weekly or monthly. Typical documents include capacity planning reports, expansion application forms, resource utilization reports, and budget calculation sheets. Fields include resource usage, peak thresholds, applied resource specifications, and budget amounts, with corresponding units of computing core count, storage capacity units, and currency units. Document structures primarily use structured fields paired with business descriptions, with no large volumes of unstructured casual content.
Constraints Imposed on Reranking and Hybrid Retrieval
Since this category’s data updates weekly or monthly, the hybrid retrieval recall pool can be limited to recently updated document sets to reduce invalid retrieval overhead. Documents contain clear structured fields such as resource specifications and peak thresholds. The reranking model must prioritize assigning higher weights to results that include these core business fields. Business scenarios focus on resource bottleneck positioning and budget coordination. Hybrid retrieval must recall multiple document types including capacity planning reports, monitoring reports, and expansion applications, so multi-source hybrid recall rules must be configured. Additionally, retrieval results for this category directly link to production resource configuration, so recall and reranking accuracy must be ensured to avoid introducing irrelevant documents that cause decision bias.
Configuration Settings
The following configurations are written for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | Retrieval for capacity planning and resource scaling needs to prioritize matching core fields such as peak thresholds and resource specifications. Reranking can filter irrelevant documents and improve result relevance |
similarity | 0.75–0.85 | This category of documents contains clear business fields. An overly high threshold will filter some relevant historical expansion applications, while an overly low threshold will introduce irrelevant content |
limit | Top 10 entries | The scale of core documents for this category is limited. Excessive results will increase reranking time, which aligns with the weekly or monthly document update cycle |
searchMode | Hybrid recall | Multiple types of documents such as capacity planning reports, resource utilization reports, and expansion applications need to be recalled simultaneously to cover all business scenarios |
| Reranking model video memory threshold | 1000 MB | Avoid video memory overflow that occupies system resources, and adapts to the average text length of documents in this category |
| Retrieval cache duration | 7 days | Matches the weekly or monthly update frequency of this category’s data, reducing repeated retrieval overhead |
Three Common Configuration Mistakes
- Symptom: A request error is prompted when clicking the model test, and the curl test for the reranking model fails. Cause: The API key or deployment address of the reranking model is not configured correctly, and the
usingReRankconfiguration item is not enabled. - Symptom: The reranking phase takes too long after knowledge base question answering. Cause: The set
limitvalue is too large, and the update range of retrieved documents is not limited, resulting in too many non-core documents being recalled to participate in reranking. - Symptom: The reranking model continuously occupies video memory and fails to release it. Cause: The video memory threshold of the reranking model is not set, or the rule for automatically recycling idle video memory is not configured, resulting in continuous occupation of system resources by video memory.
How to Verify a Complete Configuration
- Enter the knowledge base configuration page of FastGPT, and check whether the
usingReRankswitch is enabled. - Run the model testing tool, input a test question related to capacity planning, and verify that the reranking model returns results without request errors.
- Check the system monitoring panel to confirm that the video memory usage during the operation of the reranking model does not exceed the preset threshold.
- Initiate a hybrid retrieval request, and verify that the returned results include multiple types of documents such as capacity planning reports and expansion applications.
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.