This document is written based on FastGPT v4.17.0.
What this category of data looks like
Data sources for this category include change management systems, CI pipelines, monitoring data, and business calendars. Daily documents include change request forms, risk assessment reports, release window plans, rollback plans, and similar materials. Data fields include change unique identifiers, window start and end times, risk ratings, approval statuses, rollback trigger conditions, and more. Update frequency follows a weekly schedule or occurs after each change. Each document covers the full process information for a single production environment change.
What constraints do these characteristics impose on the knowledge base retrieval and recall link?
This category of data has scattered sources and inconsistent formats. The retrieval link must support unified mapping and field alignment across multiple data sources. Update frequency fluctuates with change cycles. Incremental updates are required to avoid performance losses from full index rebuilding. Document content focuses on the full change process. During retrieval, fragments related to risk ratings, release windows, and rollback plans must be prioritized for recall. Time attributes of changes must also be associated to match current production environment window restrictions.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Match core semantics of change-related documents, avoid recalling irrelevant general operation and maintenance documents |
limit | Top 8 entries | Cover all associated documents required for a single change, avoid excessive redundant content interfering with approval decisions |
searchMode | hybrid | Combine keyword and semantic retrieval, adapt to scenarios where change documents contain both structured fields and unstructured descriptions |
usingReRank | Enabled | Reorder recall results, prioritize displaying matches for key fields such as risk ratings and window times |
datasetSearchUsingExtensionQuery | Enabled | Generate extended queries based on user questions, cover different types of document fragments such as change requests and rollback plans |
collectionIds | Change management exclusive collection ID | Limit retrieval scope to only the knowledge base collection for this category, avoid mixing documents from other businesses |
Three Common Mistakes
- Phenomenon: A large number of non-change management general operation and maintenance documents appear in retrieval results, making accurate matching of target content impossible. Cause:
collectionIdsis not configured to limit the retrieval scope to the change management exclusive knowledge base collection, leading to cross-collection recall of irrelevant data. - Phenomenon: Garbled characters display on the interface after importing Chinese change documents, while English documents show no abnormalities. Cause: The UTF-8 encoding format is not specified during the import process, causing parsing errors for Chinese characters.
- Phenomenon: The number of returned retrieval results does not match the configured
limitparameter, with either too many or too few results. Cause: The result filtering logic after re-ranking is not correctly configured, or the re-ranking parameter exceeds the recall upper limit.
How to Verify Proper Configuration
- Check the collection list in the knowledge base management interface, confirm that
collectionIdshas been filled with the change management exclusive collection identifier. - Run a test query containing "change window" and "risk assessment", verify that the similarity of returned results falls within the 0.75–0.85 range.
- Import a Chinese change request form, confirm that the interface displays no garbled characters and that the vector recall process triggers normally.
- Check the retrieval logs, confirm that
usingReRankis active, and that results are sorted to prioritize key fields such as risk ratings and window times.
The values provided are common starting points and should be evaluated against samples from the actual deployment.
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.