What the data for this category looks like
Inspection and process record data for semiconductor wafer fabrication typically comes from MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Data update cycles cover each wafer batch transfer, real-time equipment parameter collection, and daily yield summaries. Daily documents include process flow cards, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect atlases, and yield reports. Fields include nanoscale process parameters, wafer batch identifiers, defect spatial coordinates, yield percentages, and similar items.
What constraints do these characteristics impose on the model access and configuration link
This category has multi-source, scattered data with large differences in update frequencies. This requires the model access link to support unified parsing of multiple formats, to avoid parsing failures caused by format incompatibility. Long-sequence documents generated by ultra-long process chains and complex rework paths need sufficient context capacity, to prevent key process records from being truncated. Requirements for interpreting nanoscale parameters and spatially distributed defects require enabling visual-related configurations to support parsing of wafer maps and SPC charts. High-risk error consequences require configuring strict output verification rules to avoid incorrect audit conclusions. This configuration is based on the standard logic of FastGPT v4.17.0.
How to set the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to complete process records and associated documents for a single wafer batch, preventing truncation of key process information |
maxTokens | 4000–6000 token | Controls the length of single-round audit output, ensuring readability of defect attribution and audit conclusions |
quoteMaxToken | 1500–2000 token | Limits the length of referenced original document fragments, focusing on process parameters and records related to the current audit item |
vision | Enabled | Supports visual content parsing of wafer maps, SPC control charts, and defect atlases, adapting to interpretation needs for nanoscale defects and parameter distributions |
temperature | 0.1–0.3 | Reduces output randomness, ensuring consistency and rigor of audit conclusions, suitable for high-risk manufacturing scenarios |
AIPROXY_API_ENDPOINT | Fill in the dedicated interface address according to the deployment environment | Adapts to proxy configuration for multi-source data access, ensuring secure transmission of internal system data |
Three common configuration errors
- Setting
temperatureto 0 still results in unnecessary thinking processes in audit conclusions. This occurs when the configuration item that disables thinking output is not correctly associated, or the model's default output logic is not overridden. - Uncaught exceptions occur after accessing a local model or proxy interface, and complete error details cannot be viewed. This is caused by failing to enable detailed log output in the deployment script, or incorrect permission configuration of the proxy interface leading to data transmission interruption.
- Wafer map files cannot be parsed after enabling the
visionconfiguration. This is caused by failing to specify an applicable image format whitelist, or image resolution exceeding the range supported by the model.
How to confirm the configuration is correct
- Upload process records and defect atlases for a single wafer batch, trigger the audit process, and check whether the output content includes complete process parameters and defect analysis.
- View the model call log to confirm that the actual values of parameters such as
maxContextandmaxTokensmatch the configured values, with no abnormal truncation. - Test the configuration to disable thinking output, confirm that no additional thinking fragments are present in the returned results.
- After accessing the internal proxy interface, perform a data pull test to confirm there are no permission or connection timeout errors.
The values provided are common starting points and should be measured against the reader's own samples.