What the data for this category looks like
This category’s data is primarily sourced from code repositories, documentation site build logs, user feedback, and version release records. Daily maintained documents include installation guides, API documentation, migration guides, and FAQs. Most document structures follow step-by-step instructions, parameter definitions, and scenario examples. Data is synchronized for updates weekly or when a new version is released. Fields include document title, content snippet, update time, and associated version number, with no mandatory additional units required.
What constraints do these characteristics impose on the re-ranking and hybrid retrieval link
Documents maintained for documentation sites are mostly structured technical instructions, with significant differences in snippet lengths. During hybrid retrieval, adaptations must be made to accommodate recall results of varying lengths, preventing snippet truncation or redundant loading. Data updates occur at a relatively high frequency, so recall results must reflect the latest version content in a timely manner. The version field must be used to filter expired documents, blocking the return of deprecated installation steps or API descriptions. Sources include original code snippets from code repositories and high-frequency questions submitted via user feedback. Weight differentiation between technical details and user questions is required, ensuring retrieval results cover the core needs of operations on duty and ticket handling. Response speed requirements are strict in operations scenarios, so the re-ranking link must quickly calibrate result order, preventing irrelevant content from interfering with problem identification.
How to configure the settings
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Filter out document fragments unrelated to operations tickets, meeting core relevance requirements for technical documentation |
limit | Top 10–15 entries | Control the number of results returned per retrieval, adapting to the fast response needs of operations on duty |
searchMode | hybrid | Combine semantic and keyword retrieval, covering technical content with different structures such as installation guides and API documentation |
usingReRank | Enabled | Re-sort recall results based on ticket keywords, prioritizing content matching operations issues |
| Version Adaptation | FastGPT v4.17.0 | Adapt to the re-ranking and hybrid retrieval function logic of the current version |
Three common misconfigurations
- Phenomenon: An error is returned when testing the re-ranking model, while other configuration processes operate normally. Cause: The access key and endpoint of the private re-ranking model have not been configured in FastGPT v4.17.0, or the model format is incompatible with the re-ranking logic of the current version.
- Phenomenon: When the re-ranking function is disabled, the relevance field of retrieval results has no practical reference value, and irrelevant content cannot be filtered. Cause: Basic hybrid retrieval only relies on initial semantic matching scores, and does not use the re-ranking link to calibrate results combined with ticket keywords. The relevance score is only a preliminary matching result and has not been optimized for the target scenario.
- Phenomenon: The number of hybrid retrieval results exceeds the set upper limit, and expired documents are included. Cause: The document version update time field is not used as a filter condition, causing old version installation guides or API descriptions to be incorrectly recalled.
How to confirm successful configuration
- Enter the knowledge base configuration page of FastGPT, check whether the
usingReRankswitch is enabled, and whethersearchModeis set to hybrid. - Submit an operations-related keyword, such as "dependency installation failed", and verify whether the sorting of retrieval results prioritizes document content corresponding to the scenario.
- Adjust the
similarityparameter to 0.7, verify whether non-target document fragments with low relevance are filtered out, and confirm that the number of retrieval results matches the range set bylimit. - Check the re-ranking model call records in the system logs, confirm that the model call is successful and the returned sorting results match expectations.
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.