What the Data for This Category Looks Like
Core documents for the transformer and high-voltage equipment manufacturing category include winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge and withstand voltage test reports, and nameplate parameter tables. Data sources include design BOM, silicon steel sheet and insulating material inventory data, vacuum drying equipment logs, oil test reports, and test station test records. Documents are updated per individual device or batch of materials. Each document contains structured fields and long text passages. For example, nameplates list rated voltage in kV and capacity in kVA. Test reports record industry-specific unit parameters such as partial discharge values in pC and moisture content in ppm.
Constraints Imposed by These Characteristics on Document Parsing and Chunking
Documents for this category must be linked to a unique device traceability number, and fields carry industry-specific units. Chunking operations must not damage the integrity of structured tables. The volume of documents for a single device is moderate, but field density is high. Splitting key parameters during chunking reduces the accuracy of subsequent cross-review. Document update frequency shifts with device production schedules, so incremental parsing and chunking must be supported to avoid repeated processing of already parsed content. Multi-page table content in test reports is lengthy. Sufficient parsing time and reasonable chunk length are required to ensure complete data retention.
How to Configure Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to long passages and structured fields in test reports, avoids splitting key parameters such as rated voltage and partial discharge values |
chunkSplitter | Mixed paragraph + fixed length | Balances natural segmentation logic for assembly records and the integrity of short fields in nameplates and test tables |
customPdfParse | Enable structured table parsing | Retains table structures and unit fields in insulating oil reports and withstand voltage test reports |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | Processes complete documents containing multiple test tables, avoids task interruptions due to parsing timeouts |
chunkSettingMode | Manual configuration | For exclusive documents of a single device, no globally unified chunking parameters are required, and adjustments can be made flexibly |
Three Common Misconfigurations
- Symptom: The chunk ID cannot be copied in the knowledge base page. Cause: The new copy component of FastGPT v4.17.0 is not enabled, or the document block ID display permission is not turned on in the configuration.
- Symptom: Duplicate content in custom-split document chunks is automatically deleted after being stored in the knowledge base, resulting in disorder of pre-built index order. Cause:
chunkSettingModeis set to automatic deduplication and duplicate block filtering is not turned off, or the custom splitting rule does not match the unique device identification field of the document. - Symptom: Parsed test report data cannot be presented in table form. Cause: The structured table parsing switch of
customPdfParseis not enabled, or complete table rows and columns are split during chunking.
How to Verify Proper Configuration
- Upload the nameplate parameter table and insulating oil report of a single transformer, check whether the parsed document retains unit fields such as rated voltage kV and moisture content ppm.
- Copy any chunk ID, confirm that it can be directly pasted into external text without format abnormalities.
- Upload a complete document containing 3 or more test tables, check whether the parsing time meets the configuration requirements and there are no timeout errors.
- Manually split an assembly record document, confirm that no duplicate blocks are automatically removed after being stored in the knowledge base, and the index order matches the splitting rules.
The values given are common starting points and should be measured against the reader's own samples.