What this category of data looks like
Audit data for semiconductor wafer fabrication originates from sources including MES systems, equipment FDC collection terminals, defect detection metrology devices, process recipe libraries, and more. Data update cycles cover per-wafer batch generation, real-time equipment parameter synchronization, and daily yield report aggregation.
Daily documents include process flow cards, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect atlases, and yield analysis reports. Fields include wafer unique identifiers, nanoscale process parameters, equipment runtime duration, defect coordinates, and more. Units include nanometers, hours, and others. Batch identifiers use string format.
What constraints these characteristics impose on model integration and configuration
High real-time data requirements require model integration to support real-time pulling of equipment FDC data. This prevents audits from relying on expired information. Multi-source and heterogeneous document types include image-based defect atlases and numerical process parameters. These require configuring multi-modal data processing capabilities. Large data volumes updated per batch demand properly limiting context and output token caps. This avoids excessive resource usage. Cross-module yield attribution requires associating multi-process data. This means a sufficiently large context window must be configured to retain complete associated information.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Matches cross-module association needs for wafer fabrication multi-process workflows, avoids truncating critical attribution information |
maxTokens | 4000–6000 token | Adapts to the typical length of single-batch wafer data, prevents output content from exceeding model limits |
vision | Enabled | Used to process image-based documents such as defect atlases, wafer maps, and SPC charts, and extract defect spatial distribution characteristics |
temperature | 0.1–0.3 | Ensures consistency in process parameter interpretation and defect determination results, reduces random deviations |
AIPROXY_API_ENDPOINT | Fill in the actual access address based on the deployment environment | Adapts to the model invocation chain, ensures normal request forwarding across data sources |
AIPROXY_API_TOKEN | Use the key calibrated through actual testing | Ensures permission security for model invocation, prevents unauthorized access |
Three common misconfigurations
- Phenomenon: After configuration, the model only retrieves knowledge base content and cannot invoke online tools to supplement the latest equipment data. Cause: The tool invocation configuration was not adjusted based on real-time data requirements, and the tool invocation capability was accidentally disabled.
- Phenomenon: After upgrading FastGPT v4.17.0, importing a process recipe CSV file triggers a format error. Cause: The new version's CSV field parsing rules were not adapted, and the original field formats for batch numbers and process parameters did not match the updated indexing logic.
- Phenomenon: After invoking the configured model, the returned audit results do not cover all defect atlas information. Cause: The
quoteMaxTokenparameter was not adjusted, resulting in excessively truncated retrieved knowledge base fragments that cannot carry complete defect spatial distribution data.
How to confirm the configuration is correct
- Upload a defective wafer image to trigger an audit task, check if the model can extract defect positions and types from the image, confirming that the
visionconfiguration is active. - Import complete process data for a single wafer batch, view the context preview after indexing, confirming that the values of
maxContextandmaxTokensdo not truncate critical associated information. - Fill in
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKEN, then initiate a test invocation, check that the model response is returned normally with no permission or connection errors. - Adjust the
temperatureparameter to 0.2, trigger the same audit task multiple times, confirm that result differences are within an acceptable range, verifying configuration stability.
The values provided are common starting points and should be measured against the reader's own samples.