What the data for this category looks like
This category's data primarily comes from stress testing platforms, APM monitoring tools, database slow logs, and production traffic baselines. Regularly produced documents include stress testing plans, capacity assessment reports, stress testing data records, and optimization suggestion sheets. Common data fields include throughput rate, latency, resource usage rate, capacity margin, and similar metrics, with corresponding units of requests/second, milliseconds, percentage, core count, or GB. Data is updated monthly or during major version releases. The content size of individual documents increases with the scale of the stress testing run, and most content is structured or semi-structured technical indicator data.
What constraints these characteristics impose on re-ranking and hybrid retrieval
Since most data consists of structured technical indicators, hybrid retrieval must match both keyword and vector features. This avoids filtering out high-value content related to core indicators that has low similarity solely based on similarity scores. Data updates infrequently but individual documents have large content sizes, so a balance must be struck between recall volume and retrieval speed. This prevents excessive invalid fragments from entering the re-ranking stage. Fields have strong professional attributes, so the re-ranking stage must prioritize results related to stress testing bottlenecks and capacity assessment. Targeted configuration of re-ranking trigger logic and sorting weights is therefore required.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | Adapts to the structured characteristics of performance testing data, improving sorting accuracy for indicator-related results |
similarity | 0.75–0.85 | Filters low-relevance stress testing log or report fragments, avoiding invalid recall that interferes with core conclusions |
limit | Top 10 entries | Balances retrieval speed and result coverage, adapting to the typical scale of single-batch stress testing data |
searchMode | hybrid | Combines vector recall and keyword recall, matching fixed indicators in stress testing plans and temporary troubleshooting queries |
| Re-ranking model binding | Bind a lightweight re-ranking model from the same ecosystem as the embedding model | Reduces cross-model call latency, adapting to the fast retrieval needs of performance testing data |
Three common mistakes
- Scenario: After offline deployment of FastGPT v4.17.0, clicking the re-ranking model test prompts a request error. The cause is incorrect port mapping and environment variable configuration for the local re-ranking model, preventing the platform from calling the model interface.
- Scenario: Knowledge base question answering has excessively high response latency after enabling re-ranking. The cause is an improperly set
limitparameter, with too many stress testing data fragments recalled, causing re-ranking computation load to exceed server capacity. - Scenario: Retrieval results are not sorted by indicator priority. The cause is that the
usingReRankswitch is not enabled, or secondary sorting by indicator fields is not specified in the hybrid retrieval configuration.
How to confirm the configuration is complete
- Navigate to the retrieval configuration page of the FastGPT knowledge base, confirm that the
usingReRankswitch is enabled, and that an available re-ranking model is bound. - Run the built-in model testing tool, input a test request for the re-ranking model, and verify that the returned result format is normal with no errors.
- Initiate a query containing stress testing bottleneck keywords, check whether retrieval results prioritize content related to throughput rate and latency.
- Compare retrieval response times before and after enabling re-ranking, confirm that latency is within an acceptable range for the business scenario.
The values provided 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.