What Data for This Category Looks Like
Data for this category typically comes from art asset libraries, project management systems, outsourced delivery packages, and engine resources. Daily generated documents include art requirement sheets, asset specification tables, outsourced feedback forms, and acceptance records. These contain fields such as asset ID, version number, delivery time, polygon count, texture resolution, and more. Data update frequency is mostly daily or weekly. Some outsourced-related documents are updated temporarily alongside delivery progress.
Constraints Imposed by These Characteristics on Re-Ranking and Hybrid Retrieval
Art asset production management data is scattered across multiple systems, requiring hybrid retrieval to connect to multiple data sources such as asset libraries and work order systems. Field mapping rules for each data source must be clearly defined during configuration. Document formats vary widely: there are both structured specification parameter tables and unstructured feedback text. Hybrid retrieval must balance semantic matching and precise matching of structured fields. Numeric fields related to assets, such as polygon count and resolution, have high matching accuracy requirements. The re-ranking link must consider both semantic relevance and numeric matching degree. In the operations on-duty and work order Q&A scenario, users need to quickly locate the latest asset versions and valid feedback. Therefore, retrieval results must filter expired data, and the re-ranking link must include update time as one of the sorting criteria.
How to Configure Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the dual numeric and semantic matching requirements for art asset specifications, filters low-relevance outsourced feedback or requirement sheets |
limit | Top 10 entries | Controls the number of results returned per retrieval, adapts to the quick review scenario for operations on-duty personnel, and avoids information overload |
searchMode | hybrid | Enables hybrid retrieval mode, which simultaneously recalls structured asset library fields and unstructured work order text |
usingReRank | Enabled | Re-ranks recalled results, prioritizes entries with high asset specification matching degrees, and optimizes the viewing order for operations personnel |
| Re-ranking model selection | Determined based on actual testing | Any mainstream re-ranking model compatible with FastGPT v4.17.0 can be used, and must correspond to the configured API key |
Three Common Configuration Mistakes
- Symptom: A re-ranking model is configured, but retrieval result order does not change. Cause: The
usingReRankparameter is not correctly enabled in the application configuration, or the re-ranking model's API key and access address are not configured in config.js. - Symptom: Retrieval results include expired asset versions or invalid outsourced feedback. Cause: The asset update time field is not included in retrieval and re-ranking rules, and hybrid retrieval does not filter entries with update times earlier than the threshold.
- Symptom: The number of returned retrieval results does not match the set value. Cause: The
limitparameter is not correctly configured, or hybrid retrieval recalls redundant entries from multiple data sources without secondary filtering.
How to Confirm Successful Configuration
- Navigate to the application configuration page of FastGPT v4.17.0, check if the
usingReRankswitch is enabled, and confirm that the bound re-ranking model list includes the configured target model. - Submit a test query such as "Verify the polygon count specification of character assets", compare the result sorting difference before and after re-ranking is enabled, and verify that the re-ranking function is active.
- View the application's retrieval logs, confirm that
searchModeis set to hybrid, and that the number of returned results matches thelimitparameter setting. - Test cross-data source queries such as "Query asset acceptance records from outsourced feedback", and verify that both asset library entries and outsourced feedback form entries are recalled.
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.