What the data for this category looks like
Data for this category primarily comes from art asset libraries, project management systems, engine resource packages, and outsourced delivery packages. Teams produce daily documents including art requirement sheets, asset specification sheets, outsourced feedback sheets, and acceptance records. Fields cover asset polygon counts, texture resolution, delivery formats, rework requirements, acceptance status, and more. Teams adjust data update frequency to daily or weekly based on project progress. Minor differences exist in asset naming and specification fields across different project teams.
What constraints these characteristics impose on reranking and hybrid retrieval
Scattered data sources require hybrid retrieval to integrate structured and unstructured data from heterogeneous systems. Teams must configure multi-source recall pipelines. Differences in asset fields across projects require adapting custom field weights during retrieval to avoid invalid matches from generic searches. Frequently updated data requires regular index refreshes. Without refreshes, retrieval results lag behind current project progress. The variety of outsourced delivery package formats requires pre-retrieval unified format validation. The reranking step must prioritize matching asset specification and acceptance status fields to support rapid specification confirmation and rework communication.
How to set configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
similarity | 0.65–0.75 | Most specification fields for art assets require exact matching. A value too low introduces irrelevant low-match assets. A value too high misses some qualifying rework records |
limit | Top 10–15 entries | Valid documents related to art assets typically fall within 10 entries. Excessive entries increase reranking calculation load. Too few entries miss key outsourced feedback and acceptance records |
searchMode | Hybrid retrieval mode | Teams must recall both structured data from art asset libraries and unstructured documents from project management systems. A single retrieval mode cannot cover all business data |
usingReRank | Enabled | Results recalled through hybrid retrieval have inconsistent field matching priorities. Reranking reorders results using core fields such as asset specifications and acceptance status to improve result relevance |
Three common configuration mistakes
- The number of results returned after reranking is far lower than expected: The interface displays fewer than 5 retrieval results. The cause is that the
limitparameter is set too low, and no upper limit is configured for retained reranking results. Excessive valid content gets filtered during reranking. - Reranking function cannot be enabled normally: Retrieval does not apply reranking logic even after the backend reranking switch is turned on. The cause is that the call key for the reranking model is not configured in the FastGPT model management module. The reranking pipeline breaks as a result.
- Hybrid retrieval results have extremely low match quality: A large number of irrelevant outsourced delivery package files appear in retrieval results. The cause is that the
similaritythreshold is set too high. It fails to cover semantically matching asset specification documents, leading to an overly narrow recall range.
How to confirm configurations are set correctly
- Enter the FastGPT knowledge base configuration page. Confirm that the
usingReRankswitch is enabled and a specified reranking model is bound. - Submit a test query containing "character model polygon count". Check whether retrieval results prioritize displaying polygon count field content from asset specification sheets.
- Adjust the
similarityparameter to 0.7. Compare retrieval results before and after adjustment. Confirm that irrelevant low-match results are effectively filtered. - View retrieval logs. Confirm that hybrid retrieval calls both the art asset library and project management system data sources. No data source call failure errors should appear.
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.