What the data for this category looks like
Semiconductor wafer fabrication technical documents include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, yield reports, and other types. Data sources include MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Data update cycles cover per-wafer batch generation, real-time equipment parameter synchronization, and daily yield report aggregation. Documents contain fields such as process number, process temperature, deposition thickness, defect location coordinates, and some content uses professional units including nanometers, degrees Celsius, and micrometers.
What constraints these characteristics impose on the "model access and configuration" link
The multiple types of documents and scattered data sources for this category require model access to support multimodal input and cross-format parsing. Datasets with multiple update frequencies will cause fluctuations in the context quota for a single round of review, so the upper limits of context and output tokens need to be adjusted flexibly. Ultra-long process chains involve document associations for hundreds of processes, which requires the model to have the ability to independently select tools to complete cross-module review. Parsing high-precision process parameters and defect data requires high output rigor, so the randomness of model generation must be controlled. At the same time, errors may lead to serious impacts such as wafer batch scrapping, so reliable interface authentication and failure retry mechanisms must be configured to ensure the stability of the review process.
How to set the configurations
The following configuration items are adapted for FastGPT v4.17.0:
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the total context length of multi-process documents, avoiding review interruptions caused by context overflow |
maxTokens | 4000–6000 token | Matches the output requirements of a single round of cross-review, ensuring complete output of cross-process attribution analysis results |
vision | Enabled | Supports parsing of image-based documents such as wafer maps, SPC charts, and defect maps, covering all types of review materials |
| Workflow/API tool selection | auto | Lets the model independently select appropriate tools to complete cross-department and cross-process document association review |
temperature | 0.2–0.4 | Reduces output randomness, ensuring accurate parsing results for process parameters and defect data |
AIPROXY_API_ENDPOINT | Fill in the dedicated address according to the deployment environment | Adapts to access scenarios where multiple models coexist, avoiding interface conflicts |
Three common misconfigurations
- A
404 body not founderror occurs when the proxy coexists with other model proxies and the model is configured separately. Cause: The general proxy address is incorrectly filled as the model's dedicated interface address without correctly distinguishing between the general proxy address and the dedicated model call endpoint. - Knowledge base search returns content that cannot be recognized as standard JSON format when using the qwen-max series models. Cause: The model's output format constraints are not adjusted, resulting in the model generating JSON with additional comments or format deviations that cannot be properly parsed by the system.
- Tool calls in advanced orchestration do not select subsequent execution tools as expected. Cause:
toolChoiceis not configured asautoor the corresponding tool identifier is not specified, causing the model to fail to correctly match the call logic.
How to confirm the configuration is correct
- Initiate a single review test for mixed documents containing images and text, verify whether the model's returned content covers image parsing and text association results, and confirm that the multimodal configuration is effective.
- Check the interface call logs, verify whether the request for
AIPROXY_API_ENDPOINTsuccessfully returns valid data, and confirm that the interface configuration is correct. - Trigger a cross-process document association review task, observe whether the model can independently call the corresponding tools to complete data integration, and confirm that the
toolChoiceconfiguration is appropriate. - Generate multi-round review tasks, check whether the output process parameters and defect data have no obvious random deviations, and confirm that the
temperatureconfiguration meets the scenario requirements.
The values provided are common starting points and should be measured against the reader's own samples.