What Data Looks Like for This Category
Data related to semiconductor wafer fabrication processes and exceptions typically comes from MES systems, equipment FDC monitoring modules, defect detection and metrology equipment, process recipe libraries, and similar sources. Daily documents include process flow cards, batch operation records, equipment recipes, SPC control charts, exception handling sheets, defect maps, yield reports, and similar materials. Fields include wafer batch numbers, process parameters such as etching rate and deposition thickness, defect spatial coordinates, yield values, and similar data points. Update frequencies cover per wafer batch run, real-time parameter reporting, and daily data aggregation.
What Constraints These Characteristics Impose on Workflow Orchestration
Multi-source heterogeneous data sources require workflows to use unified field mapping rules. This prevents data parsing failures caused by naming differences across systems. The update rhythms of batch, real-time, and daily data require workflows to support event triggering. This adapts to real-time data update cycles, and each workflow instance must bind to a single wafer batch. Complex processes with hundreds of steps and rework paths require workflows to support branch nodes. These nodes handle exception branches for different process steps. Nanoscale parameter and spatial distribution data features require nodes to have parameter validation and format conversion rules. This ensures data meets the format requirements of process documents.
How to Set Configurations
Based on FastGPT v4.17.0, the following configuration items may be referenced for setup:
| Configuration Item | Recommended Practice | Rationale |
|---|---|---|
| Global variables | Bind wafer batch number and process type in variables as fixed values | Match the exclusive data source for each wafer batch, avoid cross-batch data confusion |
| Node trigger condition | Trigger by MES batch submission event, set timeout threshold to 300 seconds | Adapt to real-time data update cycles, prevent workflow interruptions caused by waiting timeouts |
| Input field mapping | Map nodeId from device FDC data to process parameter fields | Unify field naming across different data sources, ensure accurate data access |
| Failure retry count | Set to 2 times, retry interval to 60 seconds | Address temporary device disconnections or data transfer fluctuations, reduce invalid workflow failure rates |
| Context window | Set to 6000 tokens, calibrated based on actual testing | Adapt to long-text processing requirements for process documents and exception reports |
| Result filtering | Only retain the core output fields of the first 1000 characters, discard intermediate step content | Meet the streamlined review requirements of process documents, avoid interference from redundant information |
Three Common Configuration Errors
- The AI model dropdown menu is empty in the classify node configuration panel, but the workflow runs normally after publishing. The root cause is that the classify node's model configuration relies on global variable caching, and the latest
variablesbinding relationship is not synchronized in edit mode. - The request header is missing the application token when the workflow calls an external interface, and the global variable configuration does not take effect. The root cause is that
variablesare not bound to the workflow's startup node parameters, so they cannot be loaded during runtime. - The workflow output includes the full reply from the previous AI conversation node. The root cause is that no result filtering node is configured, so intermediate step outputs are not stripped from the final result.
How to Confirm the Configuration Is Correct
- Start a test workflow, pass a simulated wafer batch number, check if
variablesload correctly. - Trigger a simulated MES batch submission event, confirm the workflow starts and executes all configured nodes as expected.
- View the workflow run logs, confirm that the field mapping corresponding to
nodeIdis correct, with no data parsing errors. - Export the final output result, confirm that it only includes core fields for process documents or exception reports, with no redundant intermediate conversation content.
The values provided are common starting points and should be measured against the reader's own samples.