What Data Looks Like for This Category
Daily documents for this category include four types: training courseware, drill plans, phishing email templates, and drill result reports. Data sources include employee contact lists, corporate email gateways, internal training platforms, and user click statistics. Data update frequency is usually quarterly or annually. Document fields cover drill audience tags, email subjects and content, review results, click conversion rates, and more. Obvious structural differences exist between different documents. Template-type documents focus on standardized formats, while report-type documents focus on statistical dimensions.
What Constraints Do These Characteristics Impose on the "Reranking and Hybrid Retrieval" Link
Data sources for this category are scattered and cover multiple systems. Hybrid retrieval needs to support joint recall across data sources. Document structures vary significantly. Reranking models must adapt to semantic weights of different fields to avoid format differences interfering with relevance ranking. Data updates happen infrequently, so high-frequency incremental synchronization is unnecessary. Regular full index updates are needed to cover the latest drill plans and reports. The business has high accuracy requirements for retrieval results. Irrelevant content must be filtered to prevent employee data leaks or audit risks. The reranking link strictly limits the number of low-relevance recall results.
How to Configure Settings
This document is written based on the configuration logic of FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Enable Reranking | Enable | Significant structural differences exist in documents of this category. Reranking filters low-relevance recall results to improve retrieval accuracy |
| Similarity Threshold | 0.72–0.78 | Balances recall coverage and result relevance. Adapts to semantic matching needs of different documents such as phishing email templates and drill reports. Adjust based on actual measurement calibration |
| Single Recall Limit | Top 8 entries | Single-round retrieval for this category needs to cover multiple types of documents. 8 recall results ensure complete coverage of templates, plans, and reports, avoiding missing key content |
| Retrieval Mode | Hybrid Retrieval | Integrates structured data (such as employee contact lists) and unstructured documents (such as training courseware). Adapts to retrieval needs of multi-source data for this category |
| Reranked Recall Count | Top 3 entries | Focuses on high-relevance results, avoiding excessive low-value content from entering the context. Adapts to precise Q&A needs of security training scenarios |
Three Common Misconfigurations
- Phenomenon: Retrieval results are empty or reference fields are empty after enabling reranking. Cause: Deployment parameters of the reranking model are not configured correctly, or hybrid retrieval data sources are not properly associated with the input interface of the reranking link.
- Phenomenon: The number of results after reranking does not meet expectations, and the single recall limit does not match the reranked recall count. Cause: The values of single recall limit and reranked recall count are not aligned, causing the reranking link to fail to obtain enough candidate results.
- Phenomenon: Retrieval result relevance is too low, failing to match security training or phishing drill business scenarios. Cause: The similarity threshold is set too high or too low, not adapting to the semantic characteristics of documents in this category, causing recall results to deviate from business needs.
How to Confirm Successful Configuration
- The FastGPT knowledge base management page is accessed, the retrieval configuration module is opened, and the enable reranking option is confirmed as checked.
- A test query targeting phishing email templates is initiated, the sorting logic of retrieval results is reviewed, and highly relevant documents are confirmed as ranked at the top.
- Retrieval logs are viewed, normal calling of the reranking link is confirmed, and no error or timeout records are present.
- The similarity threshold is adjusted, relevance changes of retrieval results before and after adjustment are compared, and configuration effectiveness is confirmed.
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.