What the data for this category looks like
The core data for this category comes from cloud billing, monitoring systems, business forecasts, and stress test reports, with an update frequency of weekly or monthly. Daily output documents include capacity planning reports, expansion application forms, resource utilization reports, and budget calculation sheets. Data fields cover resource identifiers, usage periods, application quotas, actual consumption records, predicted growth values, etc. Some are structured tabular reports, some are unstructured analysis documents, and there is no fixed unified single scenario format.
What constraints do these characteristics impose on the reranking and hybrid retrieval link
The multi-source and inconsistent format data characteristics require hybrid retrieval to adapt to both structured and unstructured data sources, avoiding recalling only a single type of document. The weekly or monthly update rhythm means that retrieval should prioritize recalling the latest recently generated reports to avoid outdated data interfering with planning decisions. The scattered fields covering the full resource lifecycle require the reranking step to sort results by resource dimension and time dimension, ensuring that content related to core planning is displayed first. The business requirements of peak prediction and bottleneck positioning require retrieval results to have sufficient accuracy, avoiding low-relevance content occupying context space and affecting decision efficiency.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Reranking toggle | Enabled | This category has multi-source and scattered fields, and reranking can improve the relevance of results to capacity planning needs |
| Single retrieval limit | Top 10 entries | Capacity planning retrieval needs to cover multiple types of resource data, and excessive results will increase the context processing burden |
| Similarity threshold | 0.7–0.8 | Low-relevance non-resource documents need to be filtered out, retaining content strongly related to capacity planning |
| Retrieval mode | Hybrid retrieval | This category includes structured reports and unstructured analysis documents, and hybrid retrieval can recall both types of data |
| Reranking recall count | Top 5 entries | Need to focus on planning data for core resources, avoiding redundant results interfering with decision-making |
Three common mistakes
- Phenomenon: Custom reranking model returns empty content after being called, and no matching resource planning data is found in the retrieval results. Cause: The request address of the custom reranking model is not configured correctly, or the calling parameter format does not meet the model requirements.
- Phenomenon: Expired historical documents account for too high a proportion in retrieval results. Cause: Reasonable time filtering rules are not set based on the data update frequency, or the similarity threshold is set too low, failing to filter out non-current planning documents.
- Phenomenon: Hybrid retrieval only recalls structured resource reports, failing to match unstructured stress test reports. Cause: The retrieval mode is not switched to hybrid retrieval, and only a single vector retrieval mode is used.
How to confirm the configuration is complete
- Enter the knowledge base configuration page of FastGPT v4.17.0, confirm that the reranking toggle is enabled.
- Initiate a capacity planning test query containing resource type and time period keywords, check whether the retrieval results cover both structured reports and unstructured analysis documents.
- View the reranking model call logs, confirm that the request address and parameter format are correct, with no error returns.
- Adjust the similarity threshold, verify whether the relevance of the retrieval results changes as expected with the threshold adjustment.
The values provided here 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.