This configuration guide applies to FastGPT v4.17.0.
What this category of data looks like
Data for this category originates from terminal management platforms, asset systems, user complaint channels, and software repositories. Daily processed documents include terminal trouble tickets, installation lists, software installation records, and remote assistance records. Trouble tickets include fields for device unique identifiers, fault symptom descriptions, and historical troubleshooting steps. Installation lists list hardware models and pre-installed software versions. Data updates occur either daily synchronously or in real time based on user requests. Individual document lengths vary widely, ranging from several hundred characters of fault descriptions to several thousand characters of bulk installation records.
What constraints do these characteristics impose on reranking and hybrid retrieval
Multi-source heterogeneous data sources create significant differences in field structures across different documents. Hybrid retrieval must accommodate recall results in varying formats. Individual document lengths vary widely, from short fault symptom descriptions to long bulk installation records. The reranking process must limit single input lengths to prevent resource overload. Real-time updated user complaint data requires retrieval modes that support real-time recall, balancing retrieval speed and result accuracy. Core fields such as device unique identifiers and fault symptoms must take priority during matching. Reranking adjusts the priority order of retrieval results.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Reranking toggle | Enabled | This category has large differences in matching degree between fault descriptions and solutions. Reranking filters low-relevance recall results |
| Similarity threshold | 0.62–0.78 | Adapts to semantic matching accuracy of terminal fault descriptions, avoids recalling irrelevant installation lists or asset records |
| Single recall limit | Top 8–12 entries | Controls the number of single recall results, adapts to the single-ticket rapid troubleshooting scenario for desktop terminal support |
| Retrieval mode | Multi-source hybrid mode | Supports heterogeneous data from terminal management platforms, asset systems, and user complaint channels |
| Reranked result count | Top 3–5 entries | Limits the number of reranked output entries, reduces resource usage from subsequent context |
| Real-time retrieval toggle | Enable as needed | Adapts to rapid retrieval requirements for real-time user complaints, balances retrieval speed and accuracy |
Three Common Mistakes
- Phenomenon: After the reranking function is enabled, system video memory usage continues to rise until it overflows, and an insufficient video memory error appears on the interface. Cause: The input document length and output count during the reranking stage are not limited. Excessively long installation lists or bulk records continuously occupy video memory resources.
- Phenomenon: After submitting a terminal fault retrieval, the returned results are not sorted by reranking, and relevance does not match the recall order. Cause: The reranking toggle is not enabled in the knowledge base configuration, or the reranking model is not correctly bound to the corresponding retrieval link.
- Phenomenon: After configuring reranking parameters, the API call returns a prompt that the corresponding configuration cannot be found, and the reranking function cannot be enabled normally. Cause: The reranking model is not bound in the configuration page for the corresponding FastGPT scenario, or the configuration is not synchronized to the global retrieval rules.
How to Confirm the Configuration is Complete
- Access the retrieval configuration page for the corresponding knowledge base, confirm that the reranking toggle is enabled.
- Submit a typical terminal fault description, verify that retrieval results are sorted to prioritize documents related to fault symptoms.
- Review system logs, confirm that no errors related to abnormal video memory usage or missing parameters appear during the reranking stage.
- Adjust the similarity threshold or recall limit, observe whether the number and relevance of retrieval results change synchronously.
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.