What the Data for This Category Looks Like
This category of data originates from MES systems, equipment FDC modules, defect detection and metrology tools, process recipe libraries, and experimental datasets. Data update frequencies align with per wafer batch production, real-time equipment status synchronization, and daily summary statistics. Daily documents include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, and yield-related records. Fields cover nanoscale process parameters, equipment operating values, defect spatial coordinates, batch identifiers, and recipe parameters. Units include nanometers, milliseconds, and other high-precision measurement identifiers.
What Constraints Do These Characteristics Impose on Model Access and Configuration
This category of data characteristics imposes multiple constraints on model access and configuration. First, process data spanning hundreds of steps and complex rework paths requires the model’s context window to support long-text retrieval. Second, high-precision fields such as nanoscale parameters and defect spatial distribution require the model to accurately match detailed parameters and abnormal scenarios. Real-time updated equipment and batch data requires the vector retrieval link to support high-frequency synchronization. Additionally, hidden defect risks in cleanroom environments require model outputs to strictly link to original document sources, and avoid unfounded inferences. Limited rework windows also require model response times to align with production line rhythms, and should not be excessively long.
How to Set the Configuration
The following configuration instructions are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to long process document contexts with hundreds of steps, avoids truncating core process parameters |
maxTokens | 4000 token | Controls single-round response length, ensures production line personnel can read quickly, while covering complete exception handling plans |
quoteMaxToken | 2000 token | Limits the length of a single quoted document, focuses on core process steps and parameters, avoids redundant information interference |
temperature | 0.2–0.4 | Reduces random output, ensures consistency and accuracy of process specification responses, aligns with standardized production requirements |
AIPROXY_API_ENDPOINT | Fill in the dedicated node according to the deployment environment | Ensures stability of the model call link, complies with internal network security requirements of wafer fabs |
AIPROXY_API_TOKEN | Token bound to a dedicated service account | Restricts access permissions, complies with data security specifications for cleanroom production environments |
Three Common Misconfigurations
- Symptom: After updating the version, the model output length is too short to cover complete process handling steps. Cause:
maxTokensormaxContextconfiguration was accidentally lowered, or default parameters were overwritten by the new version. - Symptom: After switching the vector model, the interface shows no progress, and the original configuration cannot be restored. Cause: The existing vector database cache was not cleared first, or the configuration save link had an exception, causing the configuration to not take effect.
- Symptom: Model files are missing in the deployment container, and the Q&A model cannot be loaded normally. Cause: The offline model package was not included in the packaging process during image building, or the mount path configuration was incorrect, causing the model to not load properly.
How to Confirm the Configuration Is Successfully Completed
- Initiate a test query containing a single core process parameter, check whether the model output links to the original document's citation source.
- Adjust the
temperatureparameter, then initiate multiple identical queries, verify whether the consistency of the output results meets expectations. - Upload a newly added process recipe document, trigger the vector index update, verify whether the retrieval link can normally recall new data.
- Check the configuration of
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKEN, initiate a test call, confirm there are no connection errors.
The values provided are common starting points and should be measured against the reader's own samples.