What the data for this category looks like
Business data sources for photovoltaic cell manufacturing include silicon wafer incoming quality inspection records, workshop MES systems, operation logs of diffusion furnaces, coating equipment, and printing presses, output results of EL defect detection equipment and IV performance testing equipment, and laboratory physicochemical analysis data. Data updates follow the rhythm of per production batch, per product inspection, and daily summary. Daily documents include process cards, diffusion and coating process parameter tables, printing and sintering process records, EL defect detection images, IV performance grading data, and yield analysis reports. Fields cover furnace temperature, coating thickness, grid line width, defect coordinates, conversion efficiency values, and more, with corresponding units of degrees Celsius, micrometers, millimeters, watts per square centimeter, and others.
What constraints these characteristics impose on the tool calling and plugins workflow
Multi-source heterogeneous data sources require plugins to support interfacing with devices and system protocols, to avoid data collection gaps. Mixed structured numerical data and unstructured image documents require plugins to parse structured process parameters and image defect features. The high-frequency update rhythm requires tool call trigger logic to match production nodes per batch or daily, to ensure timeliness of process documents and exception reports. The precision requirements for process parameters require plugin returned results to strictly align with business fields and units, to avoid format errors affecting subsequent process adjustments and report archiving.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Workflow/API tool selection | auto | Adapt to independent decision-making calling needs for multi-source data, matching dynamic scenarios of process drafting |
| Workflow/API function-call configuration | Enabled | Force the model to generate calling instructions per preset plugin rules, ensuring standardization of tool calls |
| Plugin trigger timing | Per batch data generation node | Match the production and inspection update rhythm of photovoltaic cell batches |
| Context window size | 8000–12000 token | Accommodate process parameter records and inspection data from multiple processes, avoiding information truncation |
| Timeout threshold | 600 seconds | Adapt to response delay ranges of device interfaces, avoiding call failures from network fluctuations |
| Number of returned entries per single call | First 3 entries | Focus on core process and exception data for the current batch, reducing redundant information interference |
Three common mistakes
- Phenomenon: Custom plugin is configured but no call is triggered, and logs show no tool calling instructions. Cause:
toolChoiceis not set toauto, or thefunctionCallmandatory calling switch is not enabled. - Phenomenon: After calling an external interface, returned result fields do not match, or EL detection image data cannot be parsed. Cause: Exclusive fields and units of photovoltaic cell manufacturing business are not aligned, and parsing adaptation rules for unstructured images are not configured.
- Phenomenon: The model does not automatically read uploaded process documents, and only generates vague drafting frameworks. Cause: Scenario matching rules for plugin triggers are not configured, and the model cannot identify the association between uploaded files and the current drafting task.
How to confirm the configuration is correct
- Test a single-batch process document drafting task, check the tool calling log in FastGPT v4.17.0, and confirm that the configurations of
toolChoiceandfunctionCalltake effect. - Call the custom plugin, verify that returned data includes exclusive photovoltaic cell fields such as diffusion furnace temperature and EL defect coordinates, and that units match business documents.
- After uploading batch inspection data, confirm that the plugin automatically triggers the call without manually activating the plugin switch.
- Simulate a scenario where device interface delay is within 600 seconds, confirm that no call timeout error occurs, and the plugin returns results normally.
The values provided here are common starting points, and should be evaluated against the user's own samples.