What the Data for This Category Looks Like
Data sources include design BOM, silicon steel sheet and insulating material procurement data, vacuum and drying equipment operation records, oil test reports, and partial discharge and withstand voltage test records from the test station. Update frequency is once per product or per material batch. Daily documents include winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge and withstand voltage test reports, and nameplate parameter tables. Fields include insulation resistance, partial discharge signal strength, test voltage, and others. Common units are megaohms, picocoulombs, and kilovolts.
What Constraints These Characteristics Impose on Model Integration and Configuration
Single product documents have considerable length and contain large volumes of structured test data. Model context configuration must adapt to long input segments to avoid truncating critical parameters. Data updates per product or batch. The system must support retrieving associated documents by batch or product ID, so model integration must adapt to multi-source structured data calling logic. Audits of insulation parameters and test results require rigor. Model output must maintain consistency, so the temperature parameter must be kept in a low range. Data sources are scattered across different devices and systems. The configuration link must support multi-interface model integration methods.
How to Configure Parameters
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the length of complete test reports and assembly records for a single transformer, avoiding truncation of critical insulation parameters |
maxTokens | 4000–6000 token | Meets the output length requirements for audit conclusions and anomaly analysis, while controlling single-call costs |
quoteMaxToken | 2000–3000 token | Ensures complete quotation of retrieved test report segments, covering insulating oil indicators and withstand voltage test data |
temperature | 0.1–0.3 | Reduces model output randomness, ensuring the rigor and consistency of audit conclusions |
AIPROXY_API_ENDPOINT | Fill in the exclusive address provided by the third-party model service provider | Adapts to the access path of third-party models, completing interface pointing configuration |
AIPROXY_API_TOKEN | Fill in the key generated by the third-party model service provider | Completes identity verification for model calls, ensuring interface access permissions |
Three Common Misconfigurations
- Phenomenon: The knowledge base can retrieve data such as insulating oil reports, but no output is returned after passing the data to the large model. Cause: The
maxContextconfiguration value is too small, truncating the complete retrieved test data segment, causing the model to fail to obtain valid input content. - Phenomenon: After upgrading to FastGPT v4.17.0, offline model rearrangement for configuration fails. Cause: The field structure of structured documents for this category is not adapted, and the input format of the rearrangement model does not match the dedicated fields for insulation parameters and test values.
- Phenomenon: Custom model configuration fails to be called even when using the One API service. Cause:
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENare not filled correctly, or the interface call permission for the corresponding model is not activated.
How to Verify Proper Configuration
- Upload the complete test report and assembly records for a single transformer, initiate an audit test, and check whether the model output includes complete insulation parameter analysis and conclusions.
- Check the model call log to confirm that the actual usage values of
maxContextandmaxTokensfall within the configured value ranges. - Submit a retrieval request for abnormal partial discharge data, and verify whether the model outputs targeted anomaly cause analysis and rectification suggestions.
- Call the test interface, confirm that the return status code is 200, and there are no identity verification-related error messages.
The values provided are common starting points and should be measured against the reader's own samples.