What the data for this category looks like
Data for this category typically comes from MES systems, equipment FDC monitoring modules, defect detection and metrology equipment, process recipe libraries, and experimental databases. Daily documents include process flow cards, wafer lot run records, equipment operation recipes, SPC statistical charts, exception handling sheets, wafer defect maps, and yield reports. Data update frequencies cover each wafer lot transfer, real-time equipment operation status, and daily yield summaries. Data fields include unique wafer lot identifiers, process node parameters, equipment operation duration, defect spatial coordinates, yield loss proportion, and more. Some parameters use dedicated units such as nanometers, revolutions per minute, and milliseconds.
What constraints these characteristics impose on workflow orchestration
Data for this category is multi-source, dispersed, and has significantly different update rhythms. This requires workflows to support asynchronous pulling and real-time alignment of multi-source data. Ultra-long process chains and multiple rework paths require workflows to support backtracking jumps and branch orchestration across process nodes. The unique identifier attribute of wafer lots requires workflows to bind the lot ID in variables as the core context for transfer. The strong coupling of nanoscale process parameters requires workflows to configure multi-parameter joint verification logic. Limited rework windows also require workflows to include a timeout trigger mechanism to avoid invalid waits occupying production resources.
How to set the configurations
These configurations are designed based on the workflow function of FastGPT v4.17.0:
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
variables | Bind the wafer lot ID as the core context variable | Match the unique identifier attribute of wafer lots to ensure accurate cross-node data association |
| Node timeout setting | 600–1200 seconds | Adapt to the average processing duration of a single process in wafer fabrication to avoid invalid timeouts triggered by equipment delays |
| Context cache size | 1000–1500 MB | Adapt to the storage requirements of multi-source process documents and lot data to avoid cache overflow |
| Tool call branch | Trigger one of two options based on defect type | Avoid parameter conflicts caused by concurrent calls, and align with the single judgment logic of process analysis |
nodeId | Bind the unique identifier of key process nodes | Enable the construction of cross-process traceability links and support node backtracking for defect analysis |
Three common misconfigurations
- Phenomenon: The workflow returns empty results after execution, or the error log shows
empty response, with no clear exception prompt on the interface. Cause: No batch ID verification invariablesis configured, leading to failed multi-source data association, or no fallback branch for timeout triggers is set. - Phenomenon: Tool call nodes execute concurrently, with multiple process analysis branches triggered simultaneously. Cause: No one-of-two branch trigger rules based on defect type are set, causing the workflow to match multiple judgment conditions at the same time.
- Phenomenon: Precise target node location fails during cross-workflow calls, leading to call confusion. Cause: No unique identifier is bound via
nodeId, causing node parameter confusion across multiple workflows.
How to confirm the configuration is correct
- Verify that the lot ID bound in
variablescan be fully transferred during cross-node calls. This can be checked by printing node logs to view variable content. - Test the trigger logic for different defect types to confirm that only the corresponding branch executes, with no concurrent calls.
- Adjust the node timeout setting to simulate equipment delay scenarios, and confirm that the fallback branch for timeout triggers starts normally.
- Check the process node identifier bound via
nodeIdto confirm that the cross-process traceability link can jump normally.
The values provided are common starting points and should be measured against the reader's own samples.