What the Data for This Category Looks Like
The data for this category comes from code repositories, CI pipelines, Issue links, and contributor-submitted content. Daily documents include PR descriptions, code review comments, CI results, and merge records. The data structure includes fields such as change scope, associated Issue number, build status, approver, submission timestamp, and others. The update frequency is daily or weekly, and updates are triggered by code commits, pipeline executions, and merge operations.
Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall
Data is scattered across multiple systems including code repositories, CI pipelines, and Issue links. Cross-source indexing must be integrated to cover complete associated information and avoid missing key details during retrieval. The update frequency is daily or weekly, so incremental update logic must be configured to reduce resource usage from full indexing. Daily documents include code change descriptions, review comments, and other content with wide text length ranges, so semantic matching for different lengths must be supported. Metadata such as associated Issue numbers and approvers can be used for precise filtering, but metadata and content must be updated synchronously, otherwise retrieval matching deviations will occur.
How to Configure Settings
This configuration follows the parameter logic of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches code and review text related to PR reviews. This range filters low-relevance general code content and retains results strongly related to change intent |
limit | Top 12 entries | Single recall results adapt to the multi-dimensional information needs of PR reviews, avoiding excessive results that cause context overload |
searchMode | hybrid | Combines semantic and keyword matching to cover technical details of code changes and text information of associated Issues |
usingReRank | Enabled | Re-ranks initial recall results to prioritize review history and CI results that best match the current PR change intent |
datasetSearchUsingExtensionQuery | Enabled | Expands associated Issue and CI-related keywords based on user queries to improve recall rate of cross-source information |
collectionIds | Bind code repositories and CI log collections for the corresponding project | Restrict retrieval scope to project data belonging to the current PR, avoiding interference from irrelevant cross-project information |
Three Common Misconfigurations
- Phenomenon: The order of returned document blocks in retrieval does not match the order of commit records associated with the original PR, or duplicate review comment entries appear. Cause: A global deduplication rule is configured, and the document block association order for PR review scenarios is not retained, resulting in damage to the original matching logic during index merging.
- Phenomenon: Retrieval results only cover code text, and associated CI results and Issue links are not recalled. Cause:
datasetSearchUsingExtensionQueryis not enabled, and associated keywords are not expanded to cover cross-source data. - Phenomenon: A single retrieval returns too many results, requiring extensive filtering of irrelevant information during review. Cause: A reasonable
limitparameter is not set, and the value exceeds the context bearing range for PR reviews.
How to Confirm Successful Configuration
- Initiate a retrieval targeting the current PR change content, and check whether the returned results include the three core document types: PR descriptions, CI results, and associated Issues.
- Adjust the
similarityparameter range, observe changes in the relevance of retrieval results, and confirm that the matching logic meets expectations. - Check the binding configuration of the retrieval collection, and confirm that only the code repository, CI logs, and review record collections belonging to the current PR project are included.
- Test retrieval requests containing duplicate content, and confirm that the deduplication or retention logic of results meets business scenario requirements.
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.