This page is compatible with FastGPT v4.17.0.
What Data for This Category Consists Of
Data for this category primarily originates from engine Profiler export files, crash platform reported logs, frame rate and resource data collected via real-device testing, and optimization records from version iterations. Daily produced documentation includes performance test reports, device adaptation lists, optimization tracking records, and crash analysis review documents. The data update rhythm typically follows daily incremental updates or full updates per version. Document fields include frame rate values, CPU occupancy rate, memory usage, crash stack identifiers, and similar metrics, with units mostly being fps, percentage, and MB.
Constraints Imposed on Vector Models and Indexes by These Characteristics
Multi-source heterogeneous data sources impose index format compatibility constraints, requiring simultaneous support for structured performance metric data and unstructured logs and document text. The high-frequency update rhythm requires indexes to support incremental synchronization, avoiding excessive server resource consumption from full reconstruction each time. Document content dense with professional technical terms requires vector models to adapt to semantic encoding of industrial-grade technical text, preventing semantic loss of key performance terms. Unit and identifier information for specific fields must be associated with metadata during the indexing phase, ensuring that results can be quickly filtered by metric type and device dimension during retrieval.
How to Set Configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-paragraph content length of performance test reports and crash analysis documents, avoiding disruption of semantic associations between key performance metrics such as frame rate and CPU usage |
| Overlap Length | 100–200 characters | Retains contextual associations across chunks, adapting to coherent descriptions of optimization steps in long documents |
defaultToken | 4096 token | Matches the single-paragraph token limit of most technical documents, avoiding truncation of complete descriptions of professional terms such as Draw Call and frame rate jitter |
maxToken | 8192 token | Supports integrating context from multiple chunks, covering complete version optimization review content |
| Similarity Threshold | 0.75–0.85 | Filters low-relevance non-performance documents, ensuring retrieval results focus on issues such as client stuttering and crashes |
| Single Search Limit | Top 10 results | Controls the number of returned results, adapting to the retrieval habits of client engineers who quickly locate performance issues |
Three Common Misconfigurations
- QA split collection retrieval created via the knowledge base API runs too slowly. The cause is that incremental index synchronization is not configured, and each retrieval triggers full vector calculation, which cannot adapt to the high-frequency update rhythm of performance data.
- After replacing the embedding model, the imported knowledge base cannot return matching historical performance data. The cause is that vector indexes of original documents are not regenerated, and retrieval still uses encoding results from the old model.
- A large number of irrelevant device adaptation list contents are included in retrieval results. The cause is that the similarity threshold is set too low, failing to filter non-target documents that do not meet the matching standard.
How to Confirm Configuration Is Correct
- Index task logs in the FastGPT backend can be reviewed to confirm that incremental synchronization tasks trigger according to the daily or per-version update rhythm, with no abnormal errors for full reconstruction.
- Query terms related to client performance can be manually entered, the matching degree of returned results can be checked, and the similarity threshold can be adjusted to a range that meets business requirements.
- The knowledge base retrieval API can be called, specifying the device or performance metric dimension, to confirm that the number of returned results matches the single search limit configuration.
- The knowledge base where the embedding model has been replaced can be checked to confirm that the vector index has been regenerated, with no residual prompts from the old model encoding.
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.