This document applies to FastGPT v4.17.0.
What this category of data looks like
The data for this category primarily comes from GitHub or Gitee Issues, release notes, user environment information, and code repositories. Maintained documents include Issue templates, label rules, response records, and problem reproduction instructions. Data updates occur daily or weekly, with community operations or core contributors syncing the latest Issues and response content on a regular basis. Data fields include problem titles, error messages, reproduction steps, responsible modules, and priority labels. There is no unified fixed format, and some content includes code blocks and multilingual mixed text.
Constraints on knowledge base retrieval and recall
Because data sources are scattered and formats vary, the retrieval link must adapt to multiple text types and code blocks to prevent retrieval failure due to format differences. The daily or weekly data update frequency requires the knowledge base to support incremental synchronization and scheduled updates, ensuring recalled content matches the latest Issue templates and label rules. There is no unified format for data fields, so both semantic and keyword matching must be used to improve classification accuracy and reduce the time spent manually completing reproduction information. If recall results deviate, it will directly lead to incorrect problem classification, extend problem processing cycles, and cause user churn and delayed defect localization.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75-0.85 | Adapts to the problem description characteristics of Issues, balances recall accuracy and coverage, and can be adjusted based on actual testing |
limit | Top 8-12 entries | Controls the context length of retrieval results, avoids excessive irrelevant content interfering with problem classification, and covers common Issue scenarios |
searchMode | keyword+vector hybrid mode | Adapts to the characteristics of Issues that contain both keywords (such as error codes, module names) and semantic descriptions |
usingReRank | Enabled | Filters low-relevance recall results, improves the accuracy of Issue classification, and reduces manual screening costs |
datasetSearchUsingExtensionQuery | Enabled | Supplements context such as user environment and version information to match more accurate Issue reproduction and response documents |
collectionIds | Bind knowledge base collections corresponding to Issue templates and label rules | Limits the recall scope to avoid irrelevant content interfering with the Issue triage and response process |
Three common misconfigurations
- Symptom: Retrieval response time exceeds 10 seconds, and the number of results is less than expected. Cause: The
limitvalue exceeds 15 entries andusingReRankis not enabled, resulting in excessive retrieval resource usage and truncated return results. - Symptom: Chinese file names uploaded to the knowledge base display garbled characters. Cause: The correct character encoding parameter is not specified in the API request header, leading to incorrect file metadata parsing.
- Symptom: Recall results contain a large number of non-Issue-related general documents. Cause:
collectionIdsis not configured to bind category-specific knowledge base collections, resulting in the retrieval scope covering irrelevant content.
How to confirm proper configuration
- Check whether
collectionIdsis bound to the knowledge base collections corresponding to Issue templates and label rules, ensuring that the recall scope is limited to category-specific documents. - Enter common Issue keywords to initiate a test retrieval, and verify whether the recall results under the set
similarityvalue cover core problem classification scenarios. - After enabling
usingReRank, check whether the sorting of retrieval results follows the priority logic of problem relevance. - Simulate high-frequency retrieval requests to confirm that the response time meets the expected rhythm of business processing.
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.