What this category of data looks like
Data sources for this category include MES systems, equipment FDC data streams, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Data update rhythms fall into three categories: per wafer batch, real-time, and daily. Daily documents cover process flow cards, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect atlases, and yield reports. Fields include wafer ID, process parameters such as etch depth and deposition thickness, batch number, defect coordinates, yield values, and more. Some parameters use nanometer-scale units.
What constraints these characteristics impose on the "tool calling and plugins" workflow
Heterogeneous multi-source data requires plugins to adapt to different interface protocols to connect to systems such as MES and equipment FDC. Differentiated update rhythms require tool calling to support dynamic triggering, to match different scenarios of real-time parameter monitoring and batch processing. Fine nanometer-scale parameters and multi-dimensional associated data require accurate mapping of fields and units during tool calling to avoid interpretation deviations. Long process chains and cross-module business logic require plugins to support associated processing of batch process data, covering full-process audits for rework paths.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Workflow/API tool selection | auto | Automatically match calling requirements for heterogeneous equipment, MES and other data sources |
| Workflow/API function-call configuration | enabled | Support structured transfer of fields such as wafer ID and process parameters, in line with business data formats |
| Tool call trigger timing | Batch completion node, real-time parameter drift threshold | Match update rhythms of per wafer batch and real-time |
| Multimodal plugin switch | Enabled | Support defect interpretation of visual data such as wafer maps and SPC charts |
| MCP function switch | Calibrated based on actual measurements | Adapt to complex scenarios of cross-department data linkage |
| Bedrock model call timeout | 600 seconds | Match reasonable duration for batch wafer data processing, based on interface specifications of FastGPT v4.17.0 |
Three common mistakes
- When calling the Claude4 Bedrock model, a software.amazon.awssdk.services.bedrockruntime.model prefixed error occurs. The cause is incorrect configuration of the model call permission policy or timeout threshold, leading to abnormal interface requests.
- The tool call returns an empty
aiPointsNotEnoughfield or an abnormal status code. The cause is incorrect mapping of defect point data fields for wafer manufacturing, leading to the system being unable to obtain sufficient interpretation samples. - Multimodal plugin calls fail, and wafer map data cannot be loaded. The cause is failure to configure the plugin's visual data reading permission, or failure to adapt to the standard format parsing rules for wafer maps.
How to confirm correct configuration
- Submit a single batch wafer inspection audit request, check the tool call log, and confirm that
toolChoiceautomatically triggers the corresponding data source interface. - After enabling the MCP function, cross-departmentally retrieve data from MES and equipment FDC, and verify that data linkage works normally.
- Upload a standard wafer map to the multimodal plugin, and confirm that defect interpretation results can be generated normally.
- Test the Bedrock model call, confirm that the request duration meets the preset threshold, and there are no permission-related errors.
The values provided are common starting points and should be measured against the reader's own samples.