What the data for this category looks like
Business data sources for semiconductor wafer fabrication include the MES system, equipment FDC module, defect detection and metrology equipment, wafer map files, process recipe library, and experimental datasets. Data updates sync in real time with wafer lot generation. Daily process parameters refresh each day. Anomaly handling records update when events are triggered.
Daily documents include process flow cards, lot run logs, equipment recipe parameters, SPC statistical charts, anomaly handling sheets, defect atlases, and yield analysis documents. Fields include wafer lot identifiers, process step parameters, equipment operation parameters, defect spatial coordinates, and lot processing status.
Constraints for multi-turn dialogue and prompt engineering
The ultra-long process chain and high-frequency update characteristics of this category require multi-turn dialogue to bind context for a specific wafer lot. Use chatId to maintain session continuity and avoid cross-lot data confusion.
Real-time generated equipment parameters and anomaly records require dialogue context to stay within the current session’s messages array. Pass the exclusive identifier of the current wafer lot via variables to prevent redundant data from interfering with result generation.
Nanoscale parameter coupling and cross-module attribution needs mean prompts must explicitly call process documents and lot run records for the corresponding process step. Limit the number of recalled documents per request to avoid parameter overload.
The limited rework window requires dialogue responses to quickly locate key anomalies, and avoid loading excessive unrelated data.
Configuration settings
This configuration is based on FastGPT v4.17.0. Specific parameter settings are listed below:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Session ID Binding | Bind the chatId of the current wafer lot | Avoid cross-lot session data confusion, maintain continuity for single-lot dialogue |
| Context Window Size | 8000–12000 token | Match the length of process documents and run records for a single wafer lot, prevent context overflow |
| Single-recall Document Count | Top 3–5 entries | Focus on core parameters and anomaly records for the current process step, avoid interference from redundant data |
| Prompt Template Configuration | Restrict calls to the process recipe library and lot logs | Explicitly define retrieval scope as wafer fabrication-specific process and lot data |
| Result Format Control | Disable Markdown syntax | Adapt to standardized output requirements for bid documents, avoid formatting chaos |
| Session Timeout Setting | 600 seconds | Align with standard work cycles for single-lot quotation preparation, prevent session interruption |
Three common misconfigurations
- Phenomenon: A new dialogue cannot continue previous wafer lot information, and the lot identifier and process parameters must be re-entered. Cause: The
chatIdparameter is not bound. Each session generates an independent dialogue identifier without associating historical data for the current lot. - Phenomenon: AI-generated quotation content only displays in the side popup window, with no direct output in the main dialogue area. Cause: The setting to directly output results to the dialogue flow is not configured, and only the side preview mode is enabled.
- Phenomenon: The bid document returned by the AI includes Markdown headings, lists, or separators, which does not meet the typesetting requirements of bid documents. Cause: The disable Markdown option for result format control is not configured, and the prompt does not explicitly restrict output format.
How to confirm proper configuration
- Initiate a test dialogue for a known wafer lot, check whether the session automatically associates previously entered lot parameters without requiring re-entry.
- After generating quotation content, verify that results are directly displayed in the main dialogue area, with no side popup redirects.
- Review the document content returned by the AI, confirm there are no Markdown-formatted headings, lists, or separators.
- Test that the session duration exceeds the preset threshold, verify whether a timeout interruption prompt is triggered.
The values provided here are common starting points, and should be evaluated against the reader's own samples.