What Data for This Category Looks Like
Typical entities are power transmission and distribution equipment, switchgear, and transformer manufacturers. Daily generated documents include winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge and withstand voltage test reports, and nameplate parameter sheets. Data sources cover design BOMs, silicon steel sheet and insulating material procurement data, vacuum and drying equipment operation data, oil test reports, and high-voltage test station test data. Data is updated per single product or per batch of raw materials. Fields include professional parameters such as insulation resistance, partial discharge value, and water content, with clear unit identifiers. The text length of individual records varies widely, from short nameplate parameters to several pages of test analysis reports.
What Constraints These Characteristics Impose on Vector Models and Indexes
The update frequency for single products or batches requires indexes to support incremental updates, avoiding performance loss from full reconstruction. The combination of professional parameters and long text requires vector models to have semantic understanding of professional terminology. Chunking strategies must balance long text integrity and embedding efficiency. The traceability requirement for single products means indexes must support precise filtering by fields such as product batch and serial number, avoiding invalid recall across products. The high sensitivity of insulation and high-voltage test data requires index permission control and data isolation to meet security requirements for production scenarios.
How to Set the Configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-paragraph text length of insulating oil reports and withstand voltage test reports, avoiding excessive splitting of professional terminology |
| Overlap Length | 50–80 characters | Retains associations of professional terminology across chunks, preventing core terms such as partial discharge and insulation from being split into different chunks |
| Similarity Threshold | 0.72–0.85 | Matches semantic similarity of professional terminology in transformer manufacturing scenarios, avoiding false or missed recalls |
| Single Search Limit | Top 8 entries | Adapts to the volume of associated data for a single product, balancing recall accuracy and query latency |
maxToken | 8192 token | Adapts to the maximum context length of mainstream embedding models, supporting complete embedding of withstand voltage test analysis reports |
defaultToken | 2048 token | Matches the token consumption limit for conventional embedding tasks, preventing single embedding tasks from exceeding resource limits |
Three Common Errors
- The interface shows "No available channels, under the default group". The cause is that the index model has not been added to the available model list of the default group, so the configured vector model cannot be matched during calls.
- Vector embedding tasks time out. The cause is that the single chunk text length exceeds the limit set by
maxToken, and no truncation or splitting is performed according to the configured chunk length, preventing the model from completing embedding. - Recall results include product data from non-target batches. The cause is that filtering rules based on product serial number and batch number have not been configured, so the index recall range is not limited to the single product currently being traced.
How to Confirm Proper Configuration
- Upload a complete withstand voltage test report, check the embedding task logs, and confirm that token consumption does not exceed the
maxTokensetting. - Enter a query statement containing professional terminology, such as "The cause of partial discharge abnormality in a certain transformer", and check whether the number of recall results matches the single search limit configuration.
- Navigate to the model management page, confirm that the index model has been added to the default group, and that the access status shows normal.
- Manually split a long test record, verify that core professional terms are not incorrectly split after the chunked text is spliced together.
The values provided are common starting points and should be measured against the reader's own samples.