What this category’s data looks like
Core users are process, equipment, quality, and production departments at crystalline silicon cell manufacturing enterprises. Daily documents include process cards, diffusion and coating parameters, printing and sintering records, EL inspection images, IV grading data, and yield reports. Data sources cover incoming silicon wafers, MES systems, production equipment such as diffusion furnaces, coating equipment, and printing presses, EL and IV testing equipment, and laboratory test data. Three update frequency levels apply: per batch, per wafer inspection, and daily. Core work includes controlling process windows for texturing, diffusion, passivation, and electrode processes, identifying hidden cracks, color differences, and efficiency losses, and tracing via electrical performance grading.
Constraints Imposed on Tool Calling and Plugins Workflow
Scattered data sources and uneven update frequencies require tool calling to support real-time pulling and association of multi-source data. This prevents data gaps from reducing the accuracy of inspection results. Coexistence of multiple document types requires plugins to adapt to both structured parameter queries and unstructured image and text parsing. This covers scenarios such as process parameter verification and defect identification. High-frequency wafer-level data and process window control requirements require tool calling to accurately match the current device’s context information. This avoids interference from redundant data during judgment.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Workflow/API tool selection | auto or a designated device parameter query tool | Adapts to the precise calling needs of equipment inspection, and prevents the model from calling irrelevant tools |
| Workflow/API function-call configuration | Enabled | Supports triggering preset tools such as parameter verification and log extraction, matching the scenario needs of photovoltaic equipment maintenance |
| Maximum context length | 8000–12000 token | Accommodates process cards, text descriptions of EL inspection images, and IV grading data for a single batch, preventing truncation of critical information |
| Single tool call timeout | 300 seconds | Adapts to time-consuming scenarios such as EL inspection image analysis and device log pulling, calibrated via actual testing |
| Number of recalled results | Top 3–5 entries | Focuses on historical inspection records and process parameters for the current device, avoiding interference from redundant information |
| FastGPT Version | v4.17.0 | Adapts to the latest tool calling and plugin management logic, ensuring functional stability |
Three Common Misconfigurations
- When calling a large model that only supports stream mode, logs such as "data acquisition exception" or "error" are returned. This occurs because stream mode adaptation is not enabled in the model channel configuration, causing incompatibility between synchronous calls and the model interface.
- A configuration conflict error occurs when upgrading the PgVector plugin. This happens because the index task for the associated knowledge base is not paused before upgrading, leading to file occupation and data synchronization conflicts.
- The text extraction tool returns the
400 Invalid JSON payload received. Unknown nameerror. This occurs because the incoming text description of EL inspection images or device logs contains unescaped special characters, which do not meet JSON format requirements.
How to Verify Successful Configuration
- Initiate a query related to equipment inspection, and check whether preset parameter query and log extraction tools are correctly triggered in the tool calling logs.
- Test calling a large model that only supports stream mode, and confirm that the streaming results have no interruptions and no data exception errors.
- Trigger the text extraction tool, and check whether the returned results use valid JSON format, with no unknown field errors.
- View the version information of the PgVector plugin, confirm that it has been upgraded to the target version, and that the associated knowledge base index task is running normally.
The values listed here are common starting points, and should be evaluated against the reader’s own operational samples.