What the data for this use case looks like
Data for this use case comes from wafer fab MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Update cycles cover each wafer batch transfer, real-time parameter collection, and daily summary statistics. Daily documentation includes process flow cards, batch operation records, equipment recipes, SPC control charts, exception handling sheets, defect atlases, and yield reports. Fields include wafer batch identifiers, process parameter thresholds, defect spatial coordinates, yield proportions, and similar metrics. Units include nanoscale process parameters, degrees Celsius, output per batch, and other relevant units.
What constraints these characteristics impose on multi-turn dialogue and prompt engineering
The coupling of ultra-long process chains and nanoscale parameters requires multi-turn dialogue to accurately link batch-level and real-time data, avoiding cross-dimensional confusion. Frequently updated real-time parameters must have their time windows limited within dialogue context to prevent expired data from being used. Exception report drafting requires linking defect atlases and SPC chart spatial distribution features. Prompt engineering must clearly define field extraction rules. Limited rework windows require dialogue responses to align with process transfer rhythms, with no excessive waits for context aggregation. Data sources are scattered across departments, so prompt engineering must clearly specify permission ranges and source identifiers for data calls.
How to Configure
This document is written for FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Context Window | 10000–15000 preceding characters | Covers the full process path and associated exception data for a single wafer batch |
messages parameter | Calibrated based on actual testing | Adapts to continuous questioning logic for batch numbers, parameter drift, and defect types during multi-turn dialogue |
| Variable Binding | Bind variables to wafer batch ID and process node | Accurately locates operation records and recipe data for the target batch |
| Maximum Response Duration | 60 seconds | Matches wait windows for cleanroom process transfers, avoiding dialogue interruptions due to timeout |
| Single Recalled Document Count | Top 8 entries | Covers SPC charts, equipment recipes, and historical exception handling data for a single process step |
| Prompt Template Constraint | Restrict calls to only current batch data | Prevents introduction of process parameters from unrelated batches that could cause incorrect attribution |
Three Common Configuration Errors
- Failure to trigger preset HTTP requests: Symptoms include no log output when calling online search or external data interfaces, with AI-generated content returned directly. The root cause is that
chatIdis not configured in the workflow to bind the current batch context, so requests do not carry correct retrieval parameters. - CORS error in frontend interface calls: Symptoms include cross-origin request blocked error messages in the browser console. The root cause is that the whitelist for the corresponding frontend domain name is not added in the privately deployed service configuration.
- Inaccurate token consumption calculation: Symptoms include a mismatch between counted token numbers and actual generated content. The root cause is that original parameter text from the equipment FDC module and wafer map metadata are not included, leading to incomplete context truncation.
How to Verify Proper Configuration
- Initiate a dialogue containing a known wafer batch ID, and confirm returned content only links operation records and recipe data for that batch.
- Review workflow logs to confirm preset HTTP requests are triggered and corresponding data entries are returned.
- Test cross-node questioning to confirm context retains process parameters and exception type information from prior dialogue turns.
- Cross-check token consumption statistics to confirm all called document text and dialogue history are included.
The values provided are common starting points and should be measured against the reader's own samples.