What Data for This Category Looks Like
Data for this category originates from design BOMs, silicon steel sheet and insulating material inbound records, vacuum drying equipment logs, insulating oil test reports, and high-voltage test station records. Updates are made per single product or batch of raw materials. New data is generated once processing for a transformer or batch of insulating material is completed. Daily documents include winding and core assembly records, vacuum drying process records, insulating oil performance reports, partial discharge and withstand voltage test reports, and product nameplate parameter tables. Fields cover insulating moisture content, partial discharge peak value, withstand voltage rating, lead connection parameters, and additional relevant metrics. All documents are linked via batch numbers and product unique identifiers.
Constraints for Vector Models and Indexes
The update rhythm for single products or batches requires indexes to support incremental synchronization, to avoid resource consumption from full index rebuilding. Multiple document types—structured parameter tables, semi-structured test reports, unstructured equipment logs—require vector models to adapt to mixed-format text. Indexes must also create linked fields for batch and product numbers to enable cross-document associated retrieval. Insulation and high-voltage test data has strict precision requirements, and errors can lead to severe consequences. This means vector recall matching accuracy must be ensured. Chunking strategies must retain key parameter integrity to prevent core information loss from splitting. Additionally, the number of linked documents per single product is limited. A reasonable recall upper limit must be set to balance retrieval efficiency and result relevance.
Configuration Settings
| Configuration Item | Recommended Values | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Documents in this category include long test records and parameter tables; this length preserves the integrity of single set of test data and avoids splitting key parameters |
| Overlap Length | 50–80 characters | Linked parameters across chunks (such as batch numbers and test times) must be covered by adjacent chunks to ensure associated information is not lost during retrieval |
| Similarity Threshold | 0.75–0.85 | High-voltage test data has high precision requirements; this range filters low-matching irrelevant historical data while avoiding missing matching results for abnormal working conditions |
| Single Search Upper Limit | Top 10 results | The number of linked documents per single product is limited; excessive recall increases context processing load |
defaultToken | 4096 | The total token count of test reports and assembly records in this category usually does not exceed this value, adapting to the input limits of mainstream vector models for FastGPT v4.17.0 |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
Common Configuration Mistakes
Three frequent configuration errors are as follows:
- Phenomenon: After switching the knowledge base vector model, the interface displays no progress and cannot revert to the original model. Cause: The forced index reset configuration is not enabled, old index resources are not released, and the new model's index task becomes blocked.
- Phenomenon: Calling the Doubao embedding interface returns a 404 error. Cause: The model's API access path and version identifier are not configured correctly, causing the request to target an invalid service endpoint.
- Phenomenon: Non-target batch test data appears in the summary report. Cause: No linked index field for batch and product numbers is established, so recall results are not filtered by production batch.
How to Verify Correct Configuration
Perform the following checks to confirm proper configuration:
- Upload the insulating oil report and test records for a single transformer, run an associated retrieval, and verify that recall results include other linked documents from the same batch.
- Configure the Doubao embedding model and send a test request, confirm that the interface returns normally without 404 errors.
- Modify the similarity threshold and run a retrieval, observe whether the number of recall results changes in line with the expected adjustment logic.
- View the index task logs for FastGPT v4.17.0, confirm that the incremental synchronization task only targets newly uploaded single product data and does not trigger a full index rebuild.
The values provided are common starting points and should be measured against the reader's own samples.