What Data for This Category Looks Like
Daily documents for this 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 data, vacuum and drying equipment, oil test results, and test station test data. Data updates occur once per device or per batch of materials. Each document contains professional fields such as insulation water content (ppm), partial discharge peak value (pC), test voltage (kV), and must be bound to the unique traceability ID of the corresponding device. This supports association of process and test data across the full lifecycle of a single device.
Constraints Imposed by Data Characteristics on Vector Models and Indexes
The professional data characteristics of this category impose three constraints on the vector index process: First, a single device must be bound to a unique traceability ID. The index must support grouped retrieval by device ID to prevent cross-device data confusion. Second, documents include fields with professional units such as insulation water content and partial discharge peak value. Chunking must retain full field integrity, and must not split context across cross-professional parameters. Third, data updates occur per device or per batch. Flexible switching between incremental indexing and full reindexing must be supported to adapt to the frequency of process iteration and test updates.
Configuration Settings
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the single-page content length of insulating oil reports and partial discharge test reports, retains professional field integrity |
| Overlap Length | 50–80 characters | Prevents context of cross-professional parameters such as insulation distance and water content from being split |
defaultToken | 4096 | Adapts to the basic content length of a single process document, complies with the basic configuration standards of FastGPT v4.17.0 |
maxToken | 8192 | Supports complete indexing of full batch process and test data for a single device |
| Similarity Threshold | 0.75–0.85 | Distinguishes normal and abnormal partial discharge data, matches the accuracy requirements of high-voltage electrical equipment |
| Single Search Limit | Top 3 entries | Focuses on core test data, avoids redundant information interfering with anomaly analysis |
Three Common Mistakes
- Phenomenon: Calling the knowledge base interface to generate an index returns a 400 status code, with a prompt that the embedding model does not match. Cause: The model was not adjusted for the professional fields of high-voltage electrical equipment, and a general embedding model was used mistakenly. This model cannot adapt to the semantic characteristics of professional data such as insulation and testing.
- Phenomenon: After replacing the embedding model, retrieval results for the already imported knowledge base become abnormal. Cause: The vector index for the already imported knowledge base was not recreated. Only vectors for newly imported data were updated, leading to inconsistent vector spaces between old and new data.
- Phenomenon: After batch importing process documents, the indexing task times out. Cause: The chunk length is set too large. The content of a single chunk exceeds the
maxTokenlimit, preventing the indexing task from completing normally.
How to Verify Proper Configuration
- View the vector index chunking log to confirm that each chunk’s length falls within the 800–1200 character range.
- Call the knowledge base search interface, input keywords related to partial discharge anomalies, and verify that returned results include professional fields such as insulation water content and test voltage.
- Import a batch of new process documents, trigger incremental indexing, and confirm that retrieval results for both old and new data return normally.
- Review the indexing task duration log to confirm that a single indexing task does not exceed the preset reasonable threshold.
The values provided are common starting points and should be measured against relevant samples.