What the data for this category looks like
Data sources include silicon wafer incoming material ledgers, workshop MES systems, operation logs from diffusion furnaces, coating equipment, and printing presses, test results from EL and IV testing equipment, and laboratory quality inspection records. Update frequencies cover every production batch, every component test, and daily summary processes. Daily documents include process cards, diffusion and coating process parameter tables, printing and sintering operation records, EL defect detection images, IV performance grading data, and yield summary tables. Fields include process temperature, coating thickness, printing offset, defect coordinates, electrical performance values, and more, with corresponding detection times and batch numbers attached.
What constraints these characteristics impose on the model access and configuration workflow
Mixed data types (structured process parameters, unstructured EL detection images) require the model to support both text and visual reasoning, so corresponding visual parameters must be configured. High-frequency updated multi-batch data creates large context load, so the context scale for single-round processing must be limited. Semantic consistency of professional process terms requires model output to remain precise, so generation parameters need to be adjusted to balance flexibility and accuracy. Dispersed data sources bring multi-format inputs, so access configuration must be compatible with different types of input files to avoid parsing failures.
How to set the configuration
This configuration guide is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the total text length of multi-batch process ledgers and test reports, avoids context overflow |
vision | Enabled | Required to process unstructured image data such as EL detection images, supports visual reasoning |
temperature | 0.1–0.3 | Summary reports of professional process data need to maintain term accuracy, reduces random generation deviations |
AIPROXY_API_ENDPOINT | http://192.168.1.100:8000/v1 | Adapts to the self-deployed DeepSeek model service address via vllm, ensures correct request routing |
AIPROXY_API_TOKEN | Access key for self-built models | Validates request legitimacy, prevents unauthorized access |
maxTokens | 2000–3000 token | Controls the generation length of a single production report, avoids redundant output content |
Three common configuration mistakes
- Enabling the
visionparameter still fails to load EL detection images: The cause is that the visual model support option was not checked in the model access configuration, or the deployed model does not support visual reasoning capabilities. - The model test returns
connector error: The cause is that the address configured inAIPROXY_API_ENDPOINTis inaccessible, or theAIPROXY_API_TOKENkey is invalid, causing the request to fail to reach the self-built model service. - The generated report lacks IV grading data: The cause is that the
maxContextconfiguration is too small, failing to fully load the original IV test data for the batch, so the model cannot extract the corresponding fields.
How to confirm the configuration is properly set up
- Submit a model test request, upload a single EL detection image and a process parameter file, check whether it can be parsed normally and generate a report containing image descriptions.
- Verify the
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENconfigurations, test the connectivity of the self-built model service via tools, confirm there are no connection errors. - Import structured ledger data from multiple batches, check whether the generated report covers all core process fields and test results, with no missing or redundant content.
- Adjust the
temperatureparameter to the set range, compare two generated reports, confirm term consistency and no obvious deviations.
The values provided are common starting points and should be measured against the reader's own samples.