What the data for this category looks like
Production data sources for photovoltaic cell manufacturing include silicon wafer incoming inspection records, workshop MES systems, operating parameters of diffusion furnaces, coating equipment, and printers, EL defect detection images, IV electrical performance test data, and laboratory yield reports. Data update frequencies fall into three categories: per-wafer inspection, per-batch synchronization, and daily aggregation. Document structures include structured process parameter tables and unstructured visual inspection images. Fields include batch unique identifier, wafer number, diffusion temperature, coating thickness, conversion efficiency, crack coordinates, and others.
What constraints do these characteristics impose on the "tool calling and plugins" workflow
The multi-source and heterogeneous data structure requires tool calling to support both structured parameter retrieval and unstructured image parsing. The high-frequency update rhythm at the wafer and batch level requires tool triggers to be bound to batch numbers and wafer numbers as unique identifiers, to avoid cross-batch data confusion. Cross-department collaboration traceability requirements mean plugins must synchronously call interfaces from three data source types: process, equipment, and quality. Real-time defect identification scenarios require support for both anomaly-triggered and fixed-timed calling logic, to adapt to the rhythm requirements of different quality monitoring workflows.
How to set the configuration
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Tool call trigger timing | Trigger by batch number + wafer number, or scheduled pull every 10 minutes | Matches the per-wafer traceability and daily aggregated production rhythm of photovoltaic cell manufacturing |
| Workflow/API tool selection | Auto (auto) | Adapts to plugin calls for multiple data source types, no manual tool function specification required |
| Plugin timeout period | 600 seconds | Covers the average time required for equipment data retrieval and EL detection image OCR parsing |
| Parameter passing rule | Retrieve batch number and wafer number from workflow variables, use default empty values if missing | Adapts to real-time incoming production data and exception scenarios |
| Number of returned results | Top 10 entries | Matches the traceability data volume of a single batch, avoids context overflow |
| Result cache duration | 24 hours | Matches the update and aggregation rhythm of daily yield reports, avoids repeated retrieval of historical data |
This document is configured based on FastGPT v4.17.0.
Three common configuration errors
- Cross-domain configuration error: The symptom is a 403 or CORS-related error returned by the browser when calling the plugin API. The cause is that the intranet or public network domain whitelist of the enterprise production system was not added in the FastGPT plugin configuration.
- Improper
functionCallparameter configuration: The symptom is that the plugin cannot correctly parse the crack coordinates in the EL detection image. The cause is that the called tool functions were not specified as the image OCR and defect recognition module. - Missing parameter fallback logic: The symptom is that the plugin returns empty fields when MES system data is not synchronized. The cause is that the logic to retrieve parameters from default values or alternate data sources was not configured.
How to verify a correct configuration
- Pass test batch numbers and wafer numbers to trigger tool calls, check if diffusion parameters, EL detection images, and IV test data for the corresponding batch are retrieved.
- View the FastGPT plugin call logs, confirm that the
toolChoiceparameter is correctly set to automatically adapt to multiple tools. - Simulate the scenario where MES data is not synchronized, check if the plugin returns default placeholder values or triggers alternate data source data retrieval.
- Verify cross-domain configuration: Use the enterprise intranet browser to call the plugin interface, confirm that no CORS-related errors occur.
The values provided are common starting points, and should be evaluated against the reader's own samples.