What data for this category includes
The data for this category primarily comes from MES systems, equipment FDC modules, defect detection and metrology devices, wafer map files, process recipe libraries, and experimental datasets. Daily documents include process flow cards, batch operation records, equipment recipes, SPC control charts, exception handling sheets, defect atlases, and yield reports. Data updates follow a per-batch wafer cycle, paired with real-time parameter collection and daily summary statistics. Most fields include nanoscale process parameters, batch identifiers, defect spatial coordinates, yield proportions, and similar data points.
Constraints imposed on model integration and configuration
The multi-source, high-frequency updated datasets of this category require the model integration link to adapt to short-cycle, high-concurrency parameter call demands. Nanoscale process parameters and spatially distributed fields have professional attributes that require the configuration link to match the model’s ability to recognize specialized terminology. Batch-level and real-time updated data volumes require context management to limit total tokens per input round, preventing exceedance of model capacity limits. The rigor required for exception reports also demands strict control over model output randomness, avoiding generation of non-compliant process content.
Configuration Settings
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the total length of multi-source documents for a single wafer batch, avoiding truncation of critical process parameters and defect data |
maxTokens | 2000–3000 token | Limits the length of exception reports or process documents generated per round, ensuring compliance with factory document typesetting specifications |
temperature | 0.1–0.3 | Reduces random output amplitude, ensuring rigor of process descriptions and exception judgments, and matching the high compliance requirements of wafer fabrication |
| Workflow/API tool selection | auto | Lets the model independently choose whether to call tools to process real-time equipment parameters, adapting to dynamically updated process data |
vision | enabled | Adapts to defect recognition needs for visual documents such as wafer maps, and handles interpretation tasks for defect atlases and SPC charts |
AIPROXY_API_ENDPOINT | Access address corresponding to the deployed model | Adapts to model access channels for cloud or local private deployments, ensuring compliance of data transmission paths |
AIPROXY_API_TOKEN | Authentication key consistent with the deployed model | Ensures legitimacy of model access permissions, avoiding permission errors during testing |
Note: This document is written based on FastGPT v4.17.0
Three Common Configuration Mistakes
- Issue: The model outputs thought process content after configuration. Cause: The thought process output switch was not disabled in the model channel configuration, or the model enables thought process output by default.
- Issue: A permission error occurs when testing a locally deployed DeepSeek8B accessed via OneAPI in FastGPT. Cause: The
AIPROXY_API_TOKENconfigured in FastGPT does not match the authentication key generated by OneAPI, or outbound permission for the corresponding port is not enabled. - Issue: The model cannot recognize defect distribution after uploading a wafer map file. Cause: The
visionconfiguration item is not enabled, or the uploaded wafer map format does not comply with the specifications supported by the model.
How to Verify Successful Configuration
- Verify the
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENconfigurations. Initiate a single call in the FastGPT model test page, and check if the returned result includes authentication success logs. - Adjust
temperatureto 0.1, initiate three identical exception report generation requests, and confirm consistency of output content to verify random output amplitude meets expectations. - Upload a single wafer batch’s process document and defect atlas, test if input under the
maxContextconfiguration is received normally with no truncation prompt. - Check the
toolChoiceconfiguration, initiate a query request containing equipment parameters, and confirm if the model independently chooses to call tools or generate content directly per the configuration.
The values provided are common starting points and should be measured against the reader's own samples.