What Data for This Category Looks Like
Semiconductor wafer manufacturing data typically comes from MES systems, equipment FDC monitoring tools, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Daily documentation includes process flow cards, batch operation records, equipment recipes, SPC control charts, exception handling sheets, defect atlases, and yield reports.
Data is updated per wafer batch, supporting real-time collection of equipment parameters and daily aggregation of yield data. Single-batch data includes fine-grained fields such as nanoscale process parameters, defect spatial coordinates, and process flow records.
Constraints Imposed by These Characteristics on Multi-turn Dialogue and Prompt Engineering
Data is associated per batch and has variable update frequencies. Multi-turn dialogue must bind a unique identifier to lock the corresponding batch, preventing interference from unrelated data across batches.
Fine-grained nanoscale parameters and spatial distribution data require prompts to explicitly specify the format and units of extracted fields, avoiding deviations from generalized interpretation.
Long process chains and complex rework workflows require multi-turn dialogue to support progressive questioning by process step. Prompts must include preset guiding logic for process nodes to ensure traceability covers the full process chain.
Configuration Setup
The following configurations are adapted for FastGPT v4.17.0:
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
chatId | Bind the unique identifier of the current wafer batch | Ensure conversation context only associates with data from the corresponding batch, avoiding interference from unrelated data across batches |
messages | Limit single-turn conversation context to no more than 10 historical records | Match the characteristics of long wafer manufacturing process chains but focused single-batch data, preventing interference from redundant information |
variables | Inject current process node and defect type tags | Narrow the retrieval scope of prompts to accurately match field requirements of process documentation |
| Context recall limit | Use the top 5 matched batch data | Adapt to scenarios with large single-batch data volume, avoiding excessive unrelated information occupying system resources |
| Response format | Calibrated based on actual measurements, supports JSON format output | Facilitate automatic parsing of analysis results by downstream systems to match business requirements |
| History cleanup trigger | Execute automatically after completing the batch traceability loop | Prevent residual historical records from appearing in subsequent conversations |
Three Common Misconfigurations
- Symptom: An extra unassociated record appears in conversation history after an API call. Cause: The
messagesparameter was not correctly reset after each call, leading to residual historical records across batches. - Symptom: Like or dislike operations on conversation responses have no effect. Cause: Feedback synchronization trigger logic was not configured, and user feedback was not bound to the batch corresponding to the current
chatId. - Symptom: Conversation response results do not meet preset JSON format requirements. Cause: The prompt did not explicitly specify the JSON field structure, and output of unstructured content was not restricted.
How to Verify Proper Configuration
- Initiate a traceability conversation for a known batch, check that returned results only associate with process and defect data from that batch, with no information from other batches.
- Call the API to launch two consecutive conversations, check that the historical records of the second conversation only include interaction content from this batch, with no residual content from the first call.
- Initiate an exception analysis request, check that returned results conform to the preset JSON format, with fields including batch ID, abnormal parameters, and attribution conclusion.
- Click the like or dislike button on a conversation result, check that the backend can capture the feedback and associate it with the current conversation's
chatId.
The values provided here are common starting points and should be evaluated against the reader's own samples.