What the data for this use case looks like
Business data for semiconductor wafer fabrication comes from MES systems, equipment FDC monitoring data, defect detection and metrology tools, wafer maps, process recipe libraries, and experimental data. Data update frequencies include per wafer batch, real-time, and daily. Daily documents include process flow cards, batch operation records, equipment recipes, SPC charts, exception handling sheets, defect atlases, and yield reports. Data fields include nanoscale process parameters, batch numbers, operation duration, defect coordinates, and some documents are visual wafer defect images.
What constraints do these characteristics impose on the model access and configuration workflow
Nanoscale parameters and multi-process correlation of wafer fabrication data require that the model context accurately matches the scope of a single batch of business, to avoid interference from redundant data. Real-time and high-frequency updated equipment data requires model calls to adapt to short response times, to prevent disruption to production line takt time. Multi-source and heterogeneous document types include visual defect atlases, which require the model to support visual data parsing. Ultra-long process chains and limited rework windows require model output to balance accuracy and timeliness, to avoid context overflow or timeout issues.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
vision | Enabled | Required to process visual inspection and anomaly data such as defect atlases and wafer maps |
maxContext | 8000–12000 token | Adapts to the volume of multi-process correlation data for a single wafer batch, avoiding context overflow |
maxTokens | 2000–3000 token | Meets the requirements for detailed output of maintenance plans and inspection results |
quoteMaxToken | 1500–2000 token | Limits the total token count of recalled business documents, reducing invalid data usage |
temperature | 0.2–0.4 | Reduces output randomness, ensuring consistency of inspection judgments and maintenance plans |
| Workflow/API tool selection | auto | Allows the model to independently decide whether to call tools such as parameter analysis |
AIPROXY_API_TOKEN | Fill in the key assigned by the access platform | Completes identity verification for model access |
AIPROXY_API_ENDPOINT | Fill in the official address of the accessed model | Adapts to routing requirements for cross-platform model calls |
Three common misconfigurations
- Failure to parse wafer defect atlases after model access: The symptom is image parsing failure logs displayed in the interface. The cause is that the
visionconfiguration item is not enabled, and the access requirements for visual data are not met. - Token consumption statistics do not match actual usage: The symptom is abnormally high or low token statistics in the background. The cause is that
quoteMaxTokenis not configured correctly, resulting in redundant recalled data being repeatedly counted in consumption. - 401 error returned after accessing a third-party model: The symptom is request rejection. The cause is incorrect entry of
AIPROXY_API_TOKEN, or the interface address does not match the model.
How to confirm successful configuration
- Upload a wafer defect atlas to trigger model Q&A, verify that the model can correctly identify defect types and locations, confirming that the
visionconfiguration is active. - Import the operation records of a single wafer batch, enter a question related to inspection anomalies, check whether the returned results are associated with the corresponding process recipes and SPC data, confirming that the context configuration is active.
- View the model call logs in FastGPT v4.17.0, verify that the token consumption value matches the settings of
maxContextandquoteMaxToken, confirming that the parameter configuration is correct. - Enter an inspection question that requires parameter analysis, verify whether the model automatically completes parameter drift analysis, confirming that the
toolChoiceconfiguration is active.
The values provided are common starting points and should be measured against the reader's own samples.