What This Category of Data Looks Like
This category of data primarily comes from change management systems, CI pipelines, monitoring data, and business calendars, and is maintained by operations managers, SREs, change control boards, or technical leads. Daily produced documents include change requests, risk assessment reports, deployment window plans, rollback plans, and more. Data is updated weekly or after each completed change. The document structure includes fields such as change subject, risk level, window time period, rollback trigger conditions, and no additional custom units. All fields are standard text and time-related fields for business scenarios.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall Workflows
The data sources for this category are scattered and the update rhythm is inconsistent, which requires the knowledge base retrieval to support incremental synchronization across multiple heterogeneous data sources. Documents focus on core business fields such as risk ratings, deployment windows, and rollback plans, so the recall logic must prioritize matching high-frequency keywords in business scenarios to avoid interference from irrelevant content. The volume of data associated with a single change fluctuates with business scale, so the retrieval upper limit must adapt to recall results of different volumes, while also ensuring the timeliness of results to match the window time period requirements of current changes.
How to Configure Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the relevance of business keywords in change management scenarios, filters low-relevance historical documents |
limit | Top 10 entries | Controls the total length of recall results to adapt to large model context window limits |
usingReRank | Enabled | Optimizes sorting priority for structured risk assessment and deployment window documents |
datasetSearchUsingExtensionQuery | Enabled | Expands search terms based on core fields of change requests to cover associated documents such as rollback plans |
searchMode | Semantic retrieval | Adapts to semi-structured change-related documents, improves semantic matching accuracy |
collectionIds | Specify the collection ID of the change management knowledge base | Precise restriction of retrieval data sources to avoid mixing documents from other business scenarios |
Three Common Configuration Mistakes
- Symptom: Knowledge base retrieval returns a fixed number of results, and the number of documents recalled per request cannot be adjusted. Cause: The
limitparameter is not configured correctly, or the parameter value is set to a fixed upper limit without enabling custom permissions. - Symptom: When calling the knowledge base during large model conversations, the re-ranking model does not take effect, but it works normally when testing the knowledge base search separately. Cause:
usingReRankis not enabled in the knowledge base configuration for large model calls, and re-ranking settings are only completed on the separate search page. - Symptom: After exporting knowledge base data from an older version, modifying it using the new version’s CSV template and importing it, training data reports an error. Cause: The CSV field format for the corresponding version is not matched, or required fields specified by the template were deleted during modification.
How to Confirm the Configuration Is Complete
- View the knowledge base configuration page in FastGPT v4.17.0, confirm that the values of parameters such as
similarityandlimitmeet the requirements of the current change management scenario. - Initiate a retrieval test simulating a change request, verify that the recall results include core business documents such as risk assessment reports and deployment window plans.
- Check the knowledge base configuration items when calling the large model, confirm that switches such as
usingReRankanddatasetSearchUsingExtensionQueryare correctly enabled. - View retrieval logs, confirm that only documents from the change management knowledge base are recalled, with no content from other unrelated data sources.
The values listed are common starting points and should be assessed against the samples relevant to the specific 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.