What data for this category looks like
The data for this category primarily comes from art asset libraries, project management systems, engine resource directories, and outsourced delivery packages. Daily documents include art requirement sheets, asset specification sheets, outsourced feedback sheets, and acceptance records. Document fields include resource numbers, specification parameters, delivery times, acceptance statuses, and other details. Some files include unitized parameters such as resolution, polygon count, and file format. Data update frequency is typically daily or weekly. Single delivery packages may receive temporary updates based on outsourced project progress.
Constraints imposed by these characteristics on knowledge base retrieval and recall
Dispersed data sources mean retrieval requires cross-system connections, so cross-source recall rules must be configured. Documents include unitized parameters such as polygon count and resolution, so precise matching with units must be supported. High update frequencies require regular incremental data synchronization to avoid recalling outdated content. Diverse outsourced delivery package formats require compatibility with multiple file type parsing. Internal work orders mostly revolve around specification confirmation and rework communication. Retrieval must prioritize recalling historical records from the same project and same role, and recall scope must be limited based on business scenarios.
How to configure settings
This content is based on the configuration logic of FastGPT v4.17.0. Specific configuration items and recommendations are listed below:
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirement for art asset specifications, and avoids recalling irrelevant resource parameters |
limit | Top 8–12 entries | Balances context length and recall coverage by covering typical asset entries for a single project in a single retrieval |
searchMode | vector + keyword | Balances semantic matching of asset specifications and keyword matching of resource numbers and project names |
usingReRank | Enabled | Performs secondary sorting on recalled asset entries by relevance, correcting bias from single vector recall |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonymous expressions of work order questions, covering art domain terminology such as "polygon count" and "UV unwrapping" |
collectionIds | Configured by project group | Limits retrieval scope to the art asset library for the current project only, avoiding interference from cross-project irrelevant data |
Three common configuration mistakes
- Symptom: Retrieval results include art asset entries from non-current projects, and some fields are empty. Cause: No
collectionIdsconfiguration was applied to limit the retrieval collection, resulting in cross-project data being recalled. - Symptom: Returned asset specifications do not match the units specified in the work order question. For example, a request for "thousand faces" as the polygon count unit returns results using "faces" as the unit. Cause:
datasetSearchUsingExtensionQuerywas not enabled, so unit parameters were not expanded for matching. - Symptom: Outsourced feedback sheets recalled from the knowledge base lack the latest rework records. Cause: Incremental synchronization tasks were not configured based on the daily or weekly update frequency, resulting in delayed data updates.
How to verify correct configuration
- Submit an internal work order that includes a specific resource number and specification parameters, and check whether retrieval results prioritize returning asset documents with the matching number.
- Adjust the value range of the
similarityparameter, and verify that the relevance of retrieval results changes as expected with the parameter adjustment. - Upload a latest outsourced delivery package, trigger incremental synchronization, then retrieve related work orders, and confirm that recalled results include the latest acceptance records.
- Switch the
searchModeconfiguration, compare result differences between vector recall and mixed recall, and confirm that the matching logic aligns with business scenarios.
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.