What the Data for This Category Looks Like
The core data for this category comes from level editor export files, internal art resource libraries, player test feedback forms, and defect tracking systems. Daily produced documents include level design drafts, resource requirement lists, version submission records, and playability debug logs. Fields typically include level number, scene coordinates, resource reference paths, test feedback severity levels, and version numbers. Data updates follow a weekly cadence, or sync with game version iterations. Single-update content mostly consists of path adjustments for a single level, resource completions, or bug fix records.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
The data sources for level content creation include structured editor export data and unstructured test feedback logs. This requires knowledge base retrieval to support mixed recall of multi-format content. Data update frequency is relatively high. The retrieval link must adapt to incremental knowledge synchronization to avoid recalled content lagging behind the current version. Some data contains fields such as precise coordinates and resource paths. Retrieval must balance semantic matching and precise field matching to prevent returning irrelevant cross-level content. The consequences of errors in this category directly affect launch progress. The relevance threshold for recall must match the business’s high requirements for content accuracy.
Configuration Settings
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | High matching precision is required for level content configuration to avoid returning irrelevant cross-level content |
limit | Top 3–5 entries | The total number of relevant knowledge entries for a single level is limited. Excessive results increase content filtering costs |
usingReRank | Enabled | Semantic expressions in test feedback forms and design drafts vary widely. The re-ranking link can optimize the relevance ranking of results |
datasetSearchUsingExtensionQuery | Enabled | Level creation often uses custom terminology. Extended queries can cover more synonymous or near-synonymous question phrasing |
searchMode | Hybrid retrieval | Adapts to both structured level coordinate and resource path fields, and unstructured debug log content |
collectionIds | Specify according to the level knowledge base collection for the current iteration version | Only recall knowledge from the current version to avoid introducing outdated old version configurations |
Three Common Misconfigurations
- Irrelevant non-target level content appears in retrieval results, accompanied by file reference identifiers. Cause: The
similaritythreshold is set too low, orcollectionIdsis not limited to the level knowledge base collection for the current version. - When an uncollected level-related question is entered, the system still returns file references from the knowledge base. Cause:
datasetSearchUsingExtensionQueryis enabled and no reasonable fallback logic is set, causing the extended query to hit irrelevant content. - Imported external level configuration data cannot be retrieved correctly. Cause: The original field mapping rules were not retained during import, preventing structured coordinate and resource path fields from being matched.
How to Verify Correct Configuration
- When a known collected level design-related question is input, the number of returned results matches the
limitsetting, and the content matches the target level. - When an uncollected general game-related question is input, the system does not return any file references from the knowledge base.
- The application's retrieval configuration panel is reviewed, confirming that the switch states of
usingReRankanddatasetSearchUsingExtensionQuerymatch the configuration requirements. - When a test level configuration fragment is uploaded, the corresponding resource path field is retrieved, confirming that structured content can be recalled normally.
The values given 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.