What data for this category looks like
The data for this category comes from four main sources: product help centers, customer organizational structures, training recorded videos, and ticket systems. Update frequency is per ticket or per version. Document formats include training slides, user manuals, training attendance sheets, and problem Q&A lists. User manuals mostly consist of structured operation steps. Transcribed training recorded videos are unstructured long texts. Ticket system data is mostly short texts of specific questions and their corresponding responses. Data fields include role permission descriptions, operation process nodes, and problem classification tags. No unified fixed unit format is applied.
What constraints do these characteristics impose on reranking and hybrid retrieval
The multi-source mixed nature of category data requires the retrieval link to cover both structured manual content and unstructured recorded video and ticket data. A hybrid retrieval mode is therefore needed to adapt to different text formats. Frequently updated ticket data requires retrieval results to have real-time performance, to avoid returning outdated training content. The presence of role permission fields requires retrieval results to match the visible scope of the corresponding role. Additionally, there is a large semantic difference between long-text training slides and short-text ticket responses. The reranking link must adjust relevance weights to prevent short-text ticket content from occupying the display positions of core manual content.
Configuration settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | The category data includes long-text slides and short-text tickets. The reranking link can correct semantic deviations from initial retrieval and improve the sorting priority of core manual content |
limit | Top 10 entries | Adapts to the rapid response needs of operation on-duty, avoids excessive results causing prolonged retrieval time |
searchMode | Hybrid retrieval mode | Adapts to format differences of multi-source data, recalls content from different sources such as help centers and ticket systems |
similarity | 0.75-0.85 | Balances retrieval recall rate and accuracy, adapts to semantic differences of multi-source data, avoids missing core content or returning irrelevant results |
| Reranking model video memory threshold | 60% of total video memory (calibrated via actual testing) | Prevents the reranking model from occupying too much video memory and causing system lag, adapts to FastGPT v4.17.0 resource scheduling rules |
Three common configuration mistakes
- Phenomenon: Video memory remains occupied and not released after the reranking model is called, and the system prompts insufficient memory. Cause: The automatic video memory recycling policy for FastGPT v4.17.0 is not configured, and temporary parameters loaded by the reranking model are not properly released.
- Phenomenon: Ticket data accounts for an excessively high proportion in hybrid retrieval results, and core user manual content is not ranked at the top. Cause: Retrieval weights for multi-source data are not adjusted. The default configuration favors short-text ticket data, and does not adapt to the business priority of the category.
- Phenomenon: Retrieval test prompts an error after configuring the
similarityparameter. Cause: The parameter is set to a value outside the 0-1 range, which does not comply with FastGPT parameter verification rules.
How to verify successful configuration
- Navigate to the retrieval configuration page of the FastGPT knowledge base, confirm that the
usingReRankswitch is enabled, and thatsearchModeis set to hybrid retrieval mode. - Submit a common role permission question for operation on-duty, check the sorting of retrieval results, and confirm that core user manual content is ranked in the top 3 entries.
- Check the number of retrieval return results, confirm that it matches the value set in the
limitparameter, with no extra or missing results. - Check system video memory usage 10 minutes after running, confirm that temporary video memory used by the reranking model has been properly released, with no persistent occupation.
The values provided 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.