What the data for this category looks like
The data sources for this category include wafer fab MES systems, equipment FDC monitoring modules, defect detection and metrology equipment, process recipe libraries, and experimental datasets. Daily documents cover process flow cards, batch run records, equipment recipe parameters, SPC control charts, exception handling sheets, wafer defect maps, and yield reports. Data updates are synchronized with each wafer batch transfer, real-time equipment data collection, and daily aggregation. Data fields include nanoscale process parameters, wafer batch ID, defect spatial coordinates, equipment runtime duration, and recipe adjustment records. Units include nanometers, pascals, amperes, and other precise metrology values.
What constraints do these characteristics impose on the model integration and configuration workflow
The multi-source heterogeneous, high-frequency update, and precise metrology characteristics of this category impose multiple constraints on model integration and configuration. Multi-source data requires corresponding data access adapters to be configured to adapt to the interface formats of different systems such as MES and FDC. High-frequency updated data requires flexible adjustment of the model context window to avoid loss of critical batch information due to context overflow. Nanoscale parameters and complex spatial defect data require appropriate token allocation rules to ensure key information is not truncated. Cross-process quality attribution requires enabling tool calling functions to support association of historical data from different modules.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale for This Setting |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the context length of historical run records and defect data for multiple wafer batches, preventing overflow of critical process information |
maxTokens | 4000 token | Controls the output length of a single analysis to ensure complete output of detailed judgment results for nanoscale parameters and defect distributions |
quoteMaxToken | 1500 token | Limits the length of referenced original data fragments, prioritizing retention of core traceability fields such as batch ID and defect coordinates |
| Workflow/API tool selection | auto | Supports automatic triggering of cross-system data calls to complete cross-process yield attribution and parameter drift analysis |
temperature | 0.2–0.4 | Reduces result randomness to ensure consistency between quality traceability conclusions and process data |
AIPROXY_API_ENDPOINT | Fill in according to the deployment environment | Adapts to the large model service address deployed locally in the wafer fab, ensuring that data does not leave the cleanroom network boundary |
Three common configuration mistakes
- Unexpected formatting symbols appear at the end of model conversation response results. The cause is that the built-in trace display rule configuration is not closed, causing the system to automatically append marker characters.
- Locally deployed models cannot be properly integrated into the system. The cause is that
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENare not configured correctly, or network policies restrict access permissions for the local server. - Cross-process data association analysis results are missing. The cause is that the
toolChoiceconfiguration is not enabled, preventing the model from calling the required multi-source data interfaces and completing cross-module yield attribution.
How to confirm the configuration is complete
- Use the FastGPT v4.17.0 system management interface to check that all configured data sources can normally pull process and defect data for the corresponding batches.
- Initiate a quality analysis request for a single wafer batch, and verify that the context length of the returned result does not exceed the range set by
maxContext. - Trigger cross-process association analysis, confirm that the system can automatically call the tool interface configured by
toolChoiceand return complete attribution steps. - Adjust the
temperatureparameter and initiate multiple tests, confirm that the consistency of the results meets business expectations.
The values provided here are common starting points and should be measured against the reader's own samples.