What the data for this category looks like
The data for this category primarily comes from frame rate and resource usage data output by the engine Profiler, device adaptation lists and frame rate fluctuation records generated by real-device testing, stack analysis reports from crash platforms, as well as optimization records and performance test reports during version iterations. Data update frequency follows daily or per version iteration cycles. Fields include frame rate (unit: fps), package size (unit: MB), crash stack path, device model, performance difference before and after optimization, and more. Document structures mostly combine structured tables and log text.
What constraints these characteristics impose on the "source citation and traceability" workflow
Data sources are scattered across engine tools, testing platforms and version documents. Precise matching of formats across different data sources is required. Therefore, the traceability workflow must support associating and binding multiple types of data sources. Data update frequency follows version iterations. It is necessary to ensure that cited content is the latest data for the current version, to avoid citing expired optimization solutions. Fields include professional performance indicators. Traceability must retain original units and test environment information to ensure the credibility of citations. Client performance data has strong timeliness. Multi-turn conversation context must be associated with the current test device and version to avoid invalid citations across scenarios.
How to set configurations
This document is written based on the configuration logic of FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale for This Setting |
|---|---|---|
quoteMaxToken | 800–1200 characters | Matches the length of individual performance test report blocks, avoids truncating critical frame rate and package data |
detail | Enabled | Retains original test environment, device model and unit information to ensure complete traceable content |
responseData | Structured text | Adapts to the tabular structure of performance documents, enabling quick location of citation sources |
| Citation data source filter | Only include optimization records and real-device test data for the current version | Adapts to daily or per-version update frequency, avoids citing expired solutions |
| Single recall count | Top 2–4 entries | Matches the output scale of performance data for a single version, avoids interference from redundant information |
| Context binding rule | Bind current test device and version number | Adapts to scene consistency in multi-turn conversations, ensures cited content corresponds to the current optimization workflow |
Three common configuration mistakes
- Phenomenon: Unable to select variables cited from the knowledge base in the code running node. Cause: Structured output of the
responseDataparameter is not enabled, so the system cannot recognize the field structure of performance data. - Phenomenon: Only citation content is returned, with no reply text. Cause: The
quoteMaxTokensetting is too small, truncating the main reply content and only retaining citation traceability information. - Phenomenon: Unable to associate context device and version information in multi-turn conversations. Cause: No context binding rule is configured, and the current test environment is not bound to the conversation turn.
How to confirm the configuration is complete
- Initiate a query related to performance optimization, check whether the reply includes source links or document paths of original test data.
- Check whether the cited content includes original field information such as device model, frame rate unit (fps), and package size (MB).
- Initiate two linked queries related to device and version, confirm that the cited content in the second reply matches the test environment of the first query.
- Adjust the
quoteMaxTokenparameter, verify that the length match between reply content and citation fragments is correct, with no truncation.
The values provided here 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.