What this category’s data looks like
Core data for this category comes from level editors, art resource libraries, player test data, and defect tracking systems. Daily produced documents include level design drafts, resource requirement sheets, test feedback forms, and version submission records. Most document fields contain level IDs, scene coordinates, resource paths, test feedback entries, and submission version numbers. Data updates follow weekly or single version iteration cycles. Single update content typically includes path adjustments for the current level, resource completion, or test correction results.
What constraints these characteristics impose on knowledge base retrieval and recall
Data is scattered across multiple tools. Retrieval must support cross-source recall of documents from different systems. Update cycles are inconsistent. Configure an incremental sync mechanism to match weekly or version-level update rhythms. Most documents contain structured fields. Retrieval must support precise matching of identifying content such as level IDs and scene coordinates. This avoids distracting fuzzy recalls. Level content has strong interconnections. Recall results must retain contextual links between documents. For example, resource requirements for a level must be associated with corresponding test feedback, to avoid isolated information. Retrieval result accuracy directly affects level production progress. Strictly control recall precision and relevance to avoid missing key design information, which would cause production blockages or delays.
How to set configurations
This page’s configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | When matching structured fields such as level IDs and scene coordinates, retain a high similarity threshold to avoid recalling irrelevant level content |
limit | Top 8–12 entries | Single version updates typically produce tens of documents. Limiting recall entries avoids overly long context passed to the model, reducing processing latency |
searchMode | Precise matching priority | Most level content retrieval needs involve precisely finding design documents for specific levels. Prioritizing precise keyword matching improves retrieval efficiency |
usingReRank | Enabled | Multi-source recall results must be reranked to optimize relevance. This ensures key information such as core level designs and resource requirements is returned first |
datasetSearchUsingExtensionQuery | Disabled | Level content retrieval needs are clear. No additional query expansion is required, to avoid introducing distracting irrelevant information |
collectionIds | Bind to level creation-specific datasets | Isolate documents from other categories. Ensure recall results only come from the current team’s level-related resources, to avoid mixing in irrelevant content |
Three Common Mistakes
- Phenomenon: After deploying a 70B-level model, replies generated after knowledge base retrieval time out. Cause: The number of recalled documents per call is not limited, leading to overly long context passed to the model that exceeds the model’s processing capacity threshold.
- Phenomenon: Retrieval results do not match document content related to level titles. Cause: No retrieval rule is configured for document title fields. Only main body content is retrieved by default, making precise matching of title keywords impossible.
- Phenomenon: Calling external data source APIs automatically includes Bearer tokens, causing connection failures. Cause: The default setting for automatically supplementing Bearer tokens is not disabled, which does not meet the authentication format requirements of external data sources.
How to Confirm Configurations Are Correct
- Upload a test level design draft, retrieve the corresponding level ID, and verify that the full content of this document is included in the recall results.
- View retrieval logs to confirm that the number of recalled documents per call falls within the preset range, with no abnormal overages.
- Call the API interface to test the external data source connection, and check that Bearer tokens are not automatically included in the request headers.
- Simulate incremental data sync for version updates, confirm that the system only syncs updated documents, with no redundant recall of old content.
The configuration values listed here are common starting points, and should be evaluated against specific samples and use cases.
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.