Data profile for this use case
Business data for semiconductor wafer fabrication scenarios originates from wafer fab process integration, equipment, yield, and manufacturing departments. Core data sources include MES systems, equipment FDC data, outputs from defect detection and metrology tools, wafer maps, process recipe libraries, and experimental data. Daily documentation covers process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect atlases, and yield reports. Data updates align with wafer batch progression, real-time parameter changes, and daily summaries. The data includes coupled parameters for hundreds of process steps and rework path information.
Constraints for multi-turn dialogue and prompt engineering
The ultra-long process chain and nanoscale parameter coupling properties require multi-turn dialogue to strictly limit context scope. This avoids redundant information interfering with critical judgments. Real-time updated batch data and limited rework windows require dialogue response speeds to match production rhythms. Excessively long context processing slows response efficiency. Multi-source heterogeneous data formats require prompt engineering to explicitly specify data association dimensions. This prevents parameter confusion across batches and process steps. Additionally, the rigor of process handling requires that no critical exception information is missed during dialogue. Multi-round follow-up questions fill in missing batch and process step details.
Configuration settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
messages | Retain the first 10 rounds of dialogue context + the latest 2 user queries | Wafer fabrication has many process steps. Excessively long historical dialogue can interfere with accurate judgments for current process steps. Limiting rounds improves prompt focus |
variables | Bind wafer batch number, current process node ID | Accurately match run records, equipment recipes, and yield data for the corresponding batch. Prevent parameter confusion across batches |
| Maximum context length | 8000–12000 characters | Complete process documentation for a single batch is typically lengthy. This range covers key process steps, exception handling, and recipe information while controlling prompt length |
| Single recall limit | Top 6 matching results | Process parameters have nanoscale coupling. Too many recall results cause prompt redundancy and reduce inference accuracy |
| Dialogue timeout | 300 seconds | Aligns with limited rework windows in wafer fabrication. Ensures fast return of handling suggestions and avoids delaying production rhythms |
| FastGPT version | FastGPT v4.17.0 | This version supports precise binding of variables and round limit configuration for messages, and meets multi-turn dialogue requirements for this scenario |
Three common configuration errors
- Observation: Unexpected non-target title content appears in dialogue logs when calling an API-based dialogue application. Cause: Non-target context fields are not filtered in
variables, causing irrelevant metadata to be included in the prompt. - Observation: Process recommendations returned after multi-turn dialogue include redundant information from intermediate steps. Cause: Context rounds for
messagesare not limited, causing historical dialogue to interfere with judgments for current process steps. - Observation: Dialogue times out without returning results. Cause: A reasonable dialogue timeout is not set, exceeding response requirements for wafer fabrication rework windows.
How to verify correct configuration
- Submit a query that includes a specific wafer batch number. Check that returned results only relate to process documentation and run records for that batch.
- Review dialogue logs. Confirm that only target context is included in the
messagesfield, with no redundant intermediate steps or unrelated metadata. - Simulate multi-round progressive queries. Check that the prompt supplements follow-up questions to fill in missing process step or parameter details based on previous responses.
- Test timeout scenarios. Confirm that response times meet window requirements for production scenarios. Adjust thresholds based on actual business needs.
The values listed above are common starting points. They should be evaluated against the reader's own samples.