What data for this product category looks like
This category's data originates from design BOMs, silicon steel sheet and insulating material procurement records, vacuum drying equipment logs, insulating oil test reports, partial discharge and withstand voltage test station records, and product nameplate parameter tables. Data is updated such that each product has a dedicated set of records, and each raw material batch is paired with a synchronized ledger update. Daily documents include winding and core assembly process sheets, vacuum drying process records, insulating oil physical and chemical index reports, high-voltage test raw data, and product nameplate parameter tables. Fields include insulation moisture content, partial discharge amplitude, withstand voltage breakdown voltage, tap changer position, and similar metrics. No standardized batch fields exist for general mass assembly data.
Constraints imposed on model access and configuration
This category’s data consists of multi-source heterogeneous documents exclusive to a single product. When accessing the model, context must be restricted to only match associated data for the current batch or single product. This prevents cross-product data from interfering with analysis results. The dataset includes precise numeric fields such as insulation moisture content and partial discharge amplitude. Configurations must support numeric threshold comparison and outlier identification logic to ensure analysis accuracy. Data comes from non-standard documents including equipment logs and test reports. Extraction rules adapted to multi-format text must be configured to ensure accurate field extraction. Single product data updates frequently. The model must pull the latest test and assembly records in real time when called, to avoid using expired data for traceability analysis.
How to set up configurations
The following configurations are based on the model access logic of FastGPT v4.17.0.
| Configuration Item | Recommended Setting Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | The total length of associated documents for a single product typically falls in the thousands of tokens range. This setting covers all relevant data such as winding records and test reports, avoiding truncation of critical traceability information |
maxTokens | 4000 token | Used to generate defect analysis conclusions and traceability reports. This balances output length and content completeness to meet single-product analysis needs |
quoteMaxToken | 2000 token | Limits the length of a single document segment referenced by the model. This ensures only core paragraphs related to the current defect issue are extracted, avoiding redundant content that disrupts analysis |
| Workflow/API tool selection | auto | Allows the model to independently choose whether to call tools. This adapts to multi-source data retrieval and extraction requirements, improving traceability accuracy |
temperature | 0.2–0.4 | Reduces random output probability. This ensures analysis conclusions align with test data and assembly records, avoiding subjective bias |
AIPROXY_API_ENDPOINT | Fill in the dedicated endpoint provided by the service provider | Adapts to network path requirements for model calls, ensuring requests reach the target model service correctly |
AIPROXY_API_TOKEN | Fill in the dedicated token generated for the account | Completes identity verification for model calls, ensuring request legitimacy and security |
Three common configuration errors
- Phenomenon: A software.amazon.awssdk.services.bedrockruntime.model error occurs when calling Bedrock series models. Cause: The matching between
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENis not correctly configured, leading to identity verification failure or incorrect request paths. - Phenomenon: After configuring the
visionparameter, water content values still cannot be extracted from scanned insulating oil reports. Cause: Special OCR preprocessing rules for high-voltage test report scanned copies are not configured, resulting in incomplete image text extraction. - Phenomenon: When using some models, the workflow text extraction component returns empty fields. Cause: The model
temperaturesetting is too high, causing the output format to deviate from preset structured extraction requirements, which fails to match the component's field parsing rules.
How to confirm configuration is complete
- Initiate a simulated traceability request for a single transformer. Check if the returned results include associated document fragments such as the product's winding records and test reports, confirming the context configuration covers required data.
- Input test data with known partial discharge abnormalities. Check if the model's output analysis conclusions match the data, confirming the temperature parameter and tool call configuration perform as expected.
- View the model call log. Confirm the requested API endpoint and token match the configured values, with no identity verification errors.
- Upload a scanned copy of an insulating oil report. Check if the text extraction results include core business fields, confirming the visual configuration and preprocessing rules are active.
The values provided are common starting points and should be measured against the reader's own samples.