What the data for this category looks like
This category of data typically originates from wafer fab MES systems, equipment FDC modules, defect detection and metrology tools. It includes wafer maps, process recipe libraries, batch operation records, SPC charts, exception handling sheets, defect atlases, and yield reports. Data updates sync in real time with wafer batch progression and equipment operating status, with daily batch analysis reports also generated. Fields include process parameters, defect coordinates, yield values, and more, paired with nanometer-level precision units and batch-unique identifiers.
Constraints imposed on multi-turn dialogue and prompt engineering
Since the data spans multiple sources and includes nanometer-level process parameters, multi-turn dialogue must restrict context to the current wafer batch to avoid mixing historical data from unrelated batches. The real-time data update feature requires dialogue context valid duration to match batch progression rhythm, preventing outdated equipment recipes or operation records from being called. Document fields include subdivided items such as process parameters and defect coordinates, so prompts must clearly specify target extraction fields to avoid result deviations from generalized retrieval. Additionally, wafer manufacturing rework windows are limited, so dialogue must quickly converge to exception determination and attribution steps, with no excessive irrelevant discussions.
Configuration settings
This configuration is based on FastGPT v4.17.0. Specific parameters are listed below:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
variables | Bind wafer batch ID, process node ID, and defect type | Accurately filter multi-source data for the current batch to avoid cross-batch confusion |
messages | Limit context to only include interactions and business records for the current batch | Match the business scope of single-batch wafer manufacturing to narrow the retrieval radius |
chatId | Bind to the unique identifier of the wafer batch | Ensure multi-turn dialogue always associates with the same production batch to prevent context drift |
| Context window size | 1800–2200 tokens | Match the typical total length of process records for a single-batch wafer, to avoid exceeding model processing limits |
| Multi-turn dialogue timeout | 600 seconds | Match the average process handling cycle for a single wafer batch; automatically clean expired context after timeout |
| Context recall limit | Top 6–10 entries | Cover core process and exception records for a single batch, to avoid interference from redundant information |
Three common configuration errors
- Symptom: Custom variables configured in the prompt are not replaced, and returned results do not associate with target wafer batch information. Cause: Key variables such as batch ID are not passed via the
variablesparameter, or variable replacement logic is not bound to user input content inmessages. - Symptom: After configuring context recall rules, the conversation details page does not display matched wafer manufacturing documents, or displays records from unrelated batches. Cause: The
chatIdparameter is not bound to the unique identifier of the current production batch, causing the retrieval scope to cover data from other batches. - Symptom: No target fields are retrieved after executing text content extraction in the workflow, and specified replies cannot continue receiving user input after triggering. Cause: The extraction field scope is not limited to process parameters for the current wafer batch, or the extraction rule does not match nanometer-level precision parameter formats.
How to verify proper configuration
- Call the API with a request that includes the
variablesparameter, and check if returned results associate with process parameters and defect information for the corresponding batch. - Enter content containing a wafer batch ID in the conversation interface, and confirm if the prompt automatically replaces relevant business variables.
- Wait longer than the configured timeout period, and check if the conversation context automatically cleans expired historical records.
- View the conversation details page, and confirm all recalled documents are target data such as current batch process flow cards and exception handling sheets.
The values provided are common starting points and should be measured against the reader's own samples.