What the data for this category looks like
Daily documents for this category include level design drafts, resource requirement sheets, test feedback forms, and version submission records. Data originates from level editors, art resource libraries, player test data, and defect tracking systems. Level design drafts include path node coordinates and trigger condition fields. Resource requirement sheets mark resource types and delivery deadlines. Test feedback forms record stuck positions and feedback times. Version submission records include submitters, version numbers, and modification content. Field structures vary significantly across different document types. Data updates occur weekly or per version.
What constraints do these characteristics impose on reranking and hybrid retrieval
Disparate data from multiple sources requires hybrid retrieval to support recall logic for different fields. For example, match coordinate information from level design drafts and text descriptions from test feedback forms simultaneously. Significant differences in field structures across document types mean a single retrieval mode cannot cover all valid content. Recall must combine semantic and field features. The weekly or per-version update rhythm requires the retrieval pipeline to adapt to frequently changing datasets. It must avoid prioritizing expired content in recall results. Business errors in this category directly lead to level blocks, reduced content quality, or delayed launches. Retrieval results must prioritize high-priority test feedback and version submission records, and reduce interference from irrelevant information.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
usingReRank | Enabled | Retrieval for level content creation requires precise filtering of irrelevant documents. Reranking improves the sorting priority of highly relevant test feedback and version records |
limit | Top 10 entries | The total number of core documents for this category is limited. Too many results increase the context processing burden and reduce retrieval efficiency |
similarity | 0.75–0.85 | Filter low-match non-core documents, and retain level design drafts and test feedback sheets that are highly relevant to the query |
searchMode | Hybrid recall | Support retrieval requirements for multiple types of fields, including coordinate data exported from level editors and resource ID data from art resource libraries |
Three common mistakes
- Symptom: The reranking model call returns empty content, and the result array field is empty. Cause: The custom request address for the reranking model is not configured correctly, or an incompatible model version with FastGPT is used.
- Symptom: The number of retrieval results is far lower than expected. Cause: The
similaritythreshold is set too high, filtering out some valid content, or thelimitparameter value is too small, failing to recall enough relevant documents. - Symptom: The locally deployed reranking model in Docker cannot be called. Cause: The model port is not correctly mapped to the host machine in the docker-compose configuration, or the host machine firewall blocks port access.
How to confirm the configuration is correct
- Access the knowledge base test page of FastGPT v4.17.0. A query related to level blocks can be entered, and the sorting and number of retrieval results can be checked against business expectations.
- Open the request logs of the reranking model. Confirm that the returned result array is not empty, and includes core documents for this category such as level design drafts and test feedback forms.
- Use a curl command to test the local deployed reranking model request address. Confirm that sorted document results are returned normally.
- Adjust the
similaritythreshold to the 0.7 to 0.9 range. Observe changes in retrieval results, and confirm the threshold setting meets business filtering requirements.
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.