What this category of data looks like
This category of data mainly comes from product help centers, customer organizational structures, training recorded videos, and ticket systems. Daily documents include training courseware, user manuals, training attendance sheets, and question and answer lists. Data is updated per order or version iteration. Most document structures include role permission descriptions, step-by-step operation breakdowns, and common question entries. Some documents include table and image materials. Fields cover training topics, operation nodes, permission scopes, and ticket association identifiers.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Data sources are scattered, and update frequencies fluctuate with business iterations. Retrieval logic must adapt to multi-format documents. Training courseware and user manuals mostly contain long-text operation steps, so chunking rules need to be optimized to avoid semantic fragmentation. The recognition effect of documents with tables and images is easily affected by formatting, so targeted adjustments to retrieval matching rules are needed. The question and answer lists associated with ticket systems must be retrieved in association with ticket priorities. At the same time, permission fields for internal operation and maintenance and customer sides must be distinguished to avoid cross-role information leakage. Retrieval needs for both internal operation duty scenarios and customer training scenarios must usually be balanced to ensure returned content complies with the access permissions of the corresponding role.
How to set the configurations
The following configurations are formulated based on the knowledge base retrieval and recall logic of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Adapt to long-text matching with operation steps, avoid mixing low-relevance results |
limit | Top 6–8 entries | Match the content volume of a single page of training courseware and user manuals, control context length |
searchMode | hybrid | Balance keyword matching and semantic recall, adapt to mixed retrieval needs of multi-format documents |
usingReRank | Enabled | Optimize the recall sorting of documents with tables and images, improve the display priority of highly relevant content |
datasetSearchUsingExtensionQuery | Disabled | Avoid generating AI questions outside the knowledge base, align with the fixed question and answer list requirements of this category |
collectionIds | Configured by training scenario grouping | Distinguish knowledge base permissions between internal operation and maintenance and customer training, match the grouping logic of business data |
Three common mistakes
- Retrieval results for imported documents with tables and images are missing or have low matching accuracy, because
usingReRankis not enabled or matching rules are not adjusted to adapt to format differences. - Recommended associated questions during retrieval do not hit the question and answer list in the knowledge base, because
datasetSearchUsingExtensionQueryis enabled to generate extended questions, which deviates from the fixed question and answer requirements. - The knowledge base for the corresponding scenario cannot be switched in real time during conversation, because the dynamic binding logic of
collectionIdsis not configured, and only a single scenario's knowledge base collection is fixedly bound.
How to confirm the configuration is correct
- Upload training documents with tables and images, initiate a query containing operation steps, check whether the returned results include specific operation content and table information in the document.
- Initiate a query containing role permission descriptions, check whether the returned results only include document content within the corresponding permission scope, with no cross-role information leakage.
- Test switching knowledge base groups for different training scenarios, check whether the retrieval results match the knowledge base corresponding to the currently selected
collectionIds. - View the associated questions recommended by retrieval, check whether all come from the question and answer list entries in the knowledge base, with no AI-generated irrelevant questions.
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.