What This Category of Data Looks Like
The data sources for this category include design BOM, silicon steel sheet and insulating material ledgers, vacuum drying equipment logs, insulating oil test reports, partial discharge and withstand voltage test records, and equipment nameplate parameter tables. Data is updated per single product or each batch of raw materials. Documents contain fields such as assembly process details, test parameters, and material test results. Each document is associated with a unique product batch identifier. Core fields include insulation moisture content, partial discharge value, tap position parameters, and others.
Constraints Imposed on Model Access and Configuration
Full-link data for a single product requires the model to associate accurate batch identifiers to avoid mixing data across products. Frequently updated test and material data requires configuration to support real-time or near-real-time data synchronization, preventing the model from calling outdated information. Configuration must support mixed-format knowledge base retrieval for multiple types of structured and semi-structured documents, including test reports and assembly records. Limit the per-round context length to avoid timeouts caused by excessive data volume for a single product. Partial discharge anomaly data from high-voltage tests requires the model to accurately match anomaly thresholds, so the relevance of retrieval results must be higher.
How to Set the Configuration
The following configurations are formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the context length of full assembly and test documents for a single product, avoiding truncation of critical test parameters |
maxTokens | 2000–3000 token | Controls the output length of single-round responses to ensure complete test clauses and maintenance steps are included |
quoteMaxToken | 1500–2000 token | Limits the total reference length of retrieved documents, ensuring the model can associate multiple test reports and assembly records simultaneously |
temperature | 0.1–0.3 | Reduces randomness for rigorous scenarios such as high-voltage test parameters and maintenance specifications, ensuring result consistency |
AIPROXY_API_ENDPOINT | Fill in the dedicated interface address according to the deployment environment | Ensures compliance of the model call link, adapting to enterprise internal network access restrictions |
AIPROXY_API_TOKEN | Bind dedicated service keys | Isolates model call permissions for different product batches to prevent unauthorized data access |
Three Common Misconfigurations
- Symptom: Single-round model inference takes too long, exceeding standard business thresholds. Cause: The
maxContextparameter is not set appropriately, loading too many associated documents for a single product, triggering additional computational pressure from the reranking model. - Symptom: Model responses only include concise conclusions without referencing specific test clauses. Cause: The
quoteMaxTokenvalue is set too low, truncating complete retrieved document content and preventing extraction of specific clauses from test reports. - Symptom: Model call fails with an API authentication error prompt. Cause: Duplicate configuration of model keys in FastGPT and OneAPI, causing key conflicts that prevent normal authentication.
How to Confirm Proper Configuration
- Initiate an inspection query for a single transformer. Verify whether the returned results include the insulating oil report parameters for the corresponding batch, confirming the accuracy of retrieved data.
- Adjust the
temperatureparameter, then initiate two identical queries. Verify the consistency of the two responses, confirming the parameter takes effect. - View the model call logs. Confirm that the actual used value of
maxContextdoes not exceed the preset range, verifying that the context loading logic works correctly. - Submit a query containing partial discharge anomaly keywords. Verify whether the returned results are associated with the high-voltage test report for the corresponding batch, confirming that the retrieval link matches the business scenario.
The values provided are common starting points and should be evaluated against the reader's own samples.