What Data for This Category Looks Like
Daily documents for this category typically include four types: performance test reports, device compatibility lists, optimization records, and crash analysis reports. Data is sourced from engine Profiler, crash platforms, real-device test results, and frame rate collection data. Updates follow a daily or per-version release cycle. No unified mandatory format exists for these documents. Common fields include device identifiers, frame rate values, resource occupancy ratios, and crash log snippets. Token usage per single document varies significantly based on content length.
Constraints on Context and Token Configuration
FastGPT v4.17.0 context and token configurations for this category must align with its data properties. The scattered, flexible format of data increases the difficulty of filtering valid content for single-round queries. Precise control of context token quotas is required. The high-frequency update feature requires context caches to refresh per version or daily cycle. This prevents reference to outdated performance data. The multi-field document structure requires filtering valid content to avoid excessive token consumption by irrelevant fields. Additionally, game client performance data is often tied to device and version dimensions. Context must aggregate data by corresponding dimensions to avoid token waste from cross-dimensional data mixing.
How to Set Configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
quoteMaxToken | 1200–1800 token | Matches the average token usage of a single valid block of game client performance report content. Prevents truncation of critical performance data due to overly long single blocks |
maxContext | 8000–12000 token | Meets the aggregation needs of multiple performance documents. Covers associated content such as device compatibility and optimization records required for single-round queries |
| Chunk overlap length | 100–200 characters | Preserves temporal continuity of performance data. Prevents cross-block frame rate change analysis from being truncated |
| Single retrieval recall limit | Top 6 entries | Matches the number of daily updated performance documents. Avoids excessive historical data consuming context tokens |
| Knowledge base refresh cycle | Daily or per version | Aligns with the update frequency of this category. Ensures context references use the latest version of performance data |
Three Common Misconfigurations
- Setting
quoteMaxTokento 1500 token still results in reference results that exceed the quota. The cause is a knowledge base chunk size set to 5000 token. Full single-block content is recalled directly, without truncation perquoteMaxTokensettings. - Context history is not cleared after workflow runs. Cumulative tokens exceed the
maxContextlimit. The cause is failure to bind the correct context clearing operation after judgment conditions. Historical conversation records continue to consume token quotas. - Token usage for each query cannot be viewed. The cause is failure to enable token statistics in system settings, or failure to enable the corresponding log recording option.
How to Verify Proper Configuration
- Upload a typical performance test report. Initiate a query requesting performance data. Check if the token usage of the returned content matches the
quoteMaxTokensetting. - Initiate multiple consecutive queries. Verify that once the
maxContextthreshold is reached, subsequent queries automatically truncate early context content. - Check the knowledge base refresh cycle settings. Confirm that the latest performance document data syncs automatically daily or per version.
- View the query details page or system logs. Confirm that token usage data for each query is properly generated and recorded.
The values provided are common starting points. They should be evaluated against local test 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.