What data for this category looks like
Daily documents for semiconductor wafer fabrication include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, yield reports, and more. Data sources cover MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Data update frequencies include per-batch wafer generation, real-time synchronization, and daily archiving. Fields and units vary across different documents, including nanoscale line widths, batch yields, equipment runtime durations, and other parameters.
What constraints these characteristics impose on model access and configuration
Multi-source heterogeneous data sources require configuring access nodes compatible with different formats. These include structured tables exported from MES, time-series data from equipment FDC modules, and visualization files such as wafer maps. Differentiated update frequencies need matching synchronization cycles. This avoids real-time data delays or expired historical data. Diverse document types and nanoscale parameters require configuring sufficient context length. This prevents truncation of critical process information. Additional parsing rule configurations are needed to ensure correct reading of visualization content such as SPC charts and defect maps.
How to set configurations
The following configuration items apply to FastGPT v4.17.0:
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the long context requirements of multi-process documents for wafer fabrication, preventing truncation of critical process parameters |
maxTokens | 4000–6000 token | Balances the accuracy and response speed of quotation calculations, and supports complex yield attribution analysis |
quoteMaxToken | 1500–2000 token | Ensures complete inclusion of attached content such as referenced SPC charts and defect maps, without impacting interpretation |
temperature | 0.2–0.4 | Ensures parameter rigor in bid documents, reducing random deviations |
AIPROXY_API_ENDPOINT | Fill in the actual address of the locally deployed model | Adapts to access requirements for locally deployed models such as QWQ-32B |
AIPROXY_API_TOKEN | Fill in the key generated during model deployment | Completes model identity verification and ensures access security |
Three common configuration errors
- After uploading attachments such as wafer batch run records and defect maps, the model does not respond, and neither simple applications nor workflow applications resolve the issue. This occurs because the
maxContextparameter is not configured correctly, and long documents are truncated before the model can read full content. - The model returns only 1 knowledge base matching text block, and cannot comprehensively output yield attribution results for multiple processes. This occurs because the
quoteMaxTokenparameter is not set reasonably, limiting the quantity and length of recalled content. - After configuring a document parsing node, the model cannot read visualization attachments such as SPC charts and wafer maps. This occurs because the
visionparameter is not enabled, and image parsing capability is not activated.
How to confirm successful configuration
- Upload a single wafer process flow card, check whether the model can correctly extract process parameters and yield data, and verify that the context length matches the configured expectations.
- Submit a test request containing multiple attachments, confirm that the model can recall and reference knowledge base text blocks that meet the configured requirements, and verify content integrity.
- Check the model access logs to confirm that
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENare configured correctly, with no connection errors or identity verification failure prompts. - Adjust the
temperatureparameter, verify the consistency of parameters in the generated bid documents, and confirm that deviations align with business requirements.
The values listed above are common starting points, and should be validated against the reader's own samples.