What This Category of Data Looks Like
Health diagnostic reports originate from industrial equipment patrol collection systems, automatically generated patrol results from operation and maintenance management platforms, and electronic copies of paper records entered by on-site patrol staff. Updates are generated after each single patrol task is completed. The single patrol cycle is adjusted based on on-site operation and maintenance requirements. Each document includes unique device identifiers, patrol timestamps, multiple physical detection parameters such as temperature and vibration amplitude, abnormal point annotations, preliminary abnormality grading, and associated on-site image files. Fields include standardized device numbers, detection values with clear units, and abnormality level tags, among others.
Constraints Imposed on Knowledge Base Retrieval and Recall
The structured parameters, unit-tagged detection data, and associated image attachments in health diagnostic reports create specific requirements for the retrieval system:
- The system must support precise matching of values with units, to avoid retrieval failures caused by inconsistent units.
- The patrol-driven document generation rhythm requires the knowledge base to support incremental synchronization per patrol cycle, to eliminate redundant overhead from full training.
- The abnormality grading tags and unique device identifiers in documents require the retrieval link to support field-level filtering, to narrow the recall scope.
- Embedded local image files require the retrieval configuration to support indexing and recall of local file paths, and cannot rely on external image hosting links.
How to Configure Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Health diagnostic reports contain a large number of precise numerical values with units. A threshold that is too low will introduce irrelevant patrol records, while a threshold that is too high will miss matching results for weak abnormalities. This value is calibrated based on actual testing. |
limit | Top 10 entries | Patrol record volume per device is concentrated. Too many single recall results will increase context burden. 10 entries can cover historical data from nearly 3 complete patrol cycles. |
searchMode | vector_only | Structured parameters of health diagnostic reports rely on vector matching to achieve semantic and numerical association. Keyword mode cannot adapt to detection data with units. |
usingReRank | Enabled | Abnormality descriptions in patrol records have similar semantics but different expressions. Reranking can improve the ranking priority of highly relevant results. |
datasetSearchUsingExtensionQuery | Enabled | Queries from patrol personnel may only mention abnormal phenomena. Extended queries can map natural language descriptions to matching conditions for device parameters and abnormality levels. |
collectionIds | Collection IDs corresponding to device groups | Create knowledge base collections by device group to accurately limit the retrieval scope and avoid invalid recall across devices. |
Three Common Configuration Errors
- Retrieval results include patrol records from unrelated devices Symptom: Returned results contain health diagnostic reports from non-target devices. Cause: Failed to limit the retrieval knowledge base collection via
collectionIds, and did not configure the retrieval scope by device group. - Local images fail to display in imported Markdown health diagnostic reports Symptom: Image positions in the knowledge base preview show blank spaces or error messages. Cause: Failed to configure path indexing rules for local images, and only used import methods that rely on external image hosting links.
- Insufficient retrieval accuracy Symptom: Patrol records for weak abnormalities are not recalled. Cause: The
similaritythreshold is set too high, filtering out weak matching results that meet the similarity requirement but do not reach the threshold.
How to Verify Correct Configuration
- A retrieval initiated with a device number for a single device’s health diagnostic reports will return only historical patrol records for that device, if configured correctly.
- Uploading a Markdown-format health diagnostic report that includes a local image path will display the image normally on the preview page, if configured correctly.
- Adjusting the
similaritythreshold to 0.7, then running a retrieval for a query that includes a weak abnormality description, will return patrol records for the corresponding abnormality, if configured correctly. - Triggering an incremental synchronization task will make newly generated health diagnostic reports available for retrieval within 10 minutes, if configured correctly.
Note: The configuration values listed here are common starting points, and should be evaluated against specific samples and use cases.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.