What data for this category looks like
Production ledger data for semiconductor wafer fabrication comes from MES systems, equipment FDC data, output from defect detection and metrology tools, wafer maps, process recipe libraries, and experimental data. Data update frequencies cover each wafer batch production run, real-time equipment parameter synchronization, and daily aggregate statistics. Daily documents include process flow cards, batch operation records, equipment recipes, SPC control charts, exception handling sheets, defect maps, and yield reports. Fields include wafer batch ID, nanoscale process parameters, equipment runtime duration, defect spatial coordinates, and yield values.
What constraints these characteristics impose on multi-turn dialogue and prompt engineering
Multi-source heterogeneous data sources require that multi-turn dialogue must be bound with a unique identifier to associate context for the same batch, preventing cross-batch data confusion. Diverse update frequencies require dynamic loading of real-time equipment parameters and historical ledgers; prompts need to distinguish between real-time status and historical records. The business characteristics of ultra-long process chains and multiple rework paths require that dialogue context does not excessively expand, so dialogue turns and data volume must be limited. The need to interpret nanoscale parameters and defect spatial distribution requires prompts to accurately match semantic features of professional documents, avoiding ambiguous interpretations.
Configuration Settings
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Maximum context length | 10000–15000 characters | Covers the full ledger data volume for a single wafer batch, avoiding truncation of critical process records |
chatId binding rule | Bind to wafer batch ID | Associates multi-turn dialogue history for the same batch, preventing cross-batch data interference |
variables configuration | Store batch number, process node, equipment ID | Quickly locate dedicated ledgers and real-time data for the corresponding batch |
messages truncation strategy | Retain the most recent 7–9 dialogue turns | Adapts to interactive needs for tracing rework paths across multiple rounds in wafer fabrication |
| Single search limit | Top 6 entries | Matches the number of associated documents for a single process node, avoiding information overload |
| Prompt template matching | Calibrated based on actual testing | Adapts to semantic matching for specialized documents such as nanoscale parameters and defect maps |
Three Common Misconfigurations
- Dialogue returns empty results or triggers a 400 error. Cause:
chatIdis not correctly bound to the wafer batch ID, so the corresponding batch’s ledger data cannot be associated, triggering a context verification failure. - Workflow execution halts after reaching the AI dialogue step. Cause: Required parameters such as wafer batch ID are not configured in
variables, so the knowledge base search cannot match the corresponding documents, triggering the workflow termination condition. - Historical dialogue logs are automatically cleaned, making interaction processes untraceable. Cause: Dialogue log persistence configuration is not enabled, and the default cleaning rule deletes historical dialogue records.
How to Verify Proper Configuration
- Enter a known wafer batch ID to initiate a dialogue, check whether the process flow card and operation records for the corresponding batch can be automatically retrieved.
- Initiate multi-round follow-up questions about rework paths and yield attribution, check whether the dialogue context retains the batch information from previous rounds.
- Review the parameters stored in
variablesto confirm they include dedicated fields such as batch number and equipment ID, with no redundant irrelevant data. - Trigger the workflow bound to this scenario, check whether the process smoothly transitions from the knowledge base search step to the AI dialogue step without interruption errors.
The values provided here are common starting points and should be evaluated against the reader’s own samples.