What this category of data looks like
Data sources for this category include backup systems, database logs, object storage, and disaster recovery platforms. Daily documents cover backup policies, recovery drill reports, disaster recovery switchover plans, and data validation records. Fields include backup task ID, recovery point objective, drill execution time, and data consistency validation results, among others. Data update cadences are divided into three categories: daily, monthly, and quarterly. Backup task logs update daily, drill reports are archived monthly, and long-term validation records are retained quarterly.
What constraints do these characteristics impose on the knowledge base retrieval and recall workflow
Multiple data sources require retrieval to cover documents with different formats and storage locations, so cross-source synchronization rules must be configured. Different update frequencies require staged synchronization tasks to avoid resource consumption from full synchronization. Documents include structured task fields and unstructured plan text, so adaptation for both keyword and semantic retrieval is needed. Documents with many tables and images account for a certain share, so chunking logic must be optimized to preserve structured information. Content from the most time-consuming links, such as recovery point confirmation and drill script execution, must be accurately recalled. Therefore, retrieval thresholds must strictly match the relevance requirements of the business scenario.
How to set configurations
This document is written based on FastGPT v4.17.0. Below are configuration schemes adapted to this category:
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Filter low-relevance content to ensure retrieved documents are strongly related to backup recovery and disaster recovery drills |
limit | Top 3–5 entries | Control the number of returned results, adapting to the need for quick browsing of multiple reference documents during operational on-call |
searchMode | hybrid | Combine keyword and semantic retrieval to cover mixed content such as task numbers in backup policies and validation results in drill reports |
usingReRank | Enabled | Reorder initial retrieval results to raise the priority of content related to recovery point confirmation and data consistency validation |
datasetSearchUsingExtensionQuery | Enabled | Automatically optimize ticket questions and expand query terms matching backup policies and disaster recovery plans |
collectionIds | Bind to the dedicated backup recovery knowledge base collection | Limit retrieval scope to avoid mixing in other operational documents and improve accuracy |
Three common configuration mistakes
- Phenomenon: Only 1 knowledge base matching result is returned. Cause: The
limitparameter is set to 1, which does not adapt to the need for multiple reference documents in backup recovery scenarios. - Phenomenon: Table and image content in imported Word documents is poorly recognized. Cause: Chunking rules are not adjusted for structured documents, and the default chunking logic does not preserve structured information from tables and images.
- Phenomenon: Unable to manually select a target knowledge base during conversation. Cause: Multi-knowledge base binding workflow nodes are not configured, and only a single backup recovery knowledge base collection is locked.
How to confirm the configuration is correct
- Initiate a test query containing "recovery drill steps", check if the number of returned results matches the
limitparameter setting. - Upload a backup policy document containing tables, check if the retrieved content includes the structured information of the tables after retrieval.
- After configuring multi-knowledge base binding logic, initiate a test query to confirm that the interface displays a knowledge base selection dropdown menu and supports manual switching.
- View the retrieval log to confirm that the similarity scores of retrieved results fall within the range set by the
similarityparameter.
The values provided are common starting points and should be measured 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.