What the Data for This Category Looks Like
Data for this category originates primarily from wafer fab MES systems, equipment FDC modules, defect detection and metrology devices, wafer map files, process recipe libraries, and experimental data. Daily used documents include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, and yield reports. Data update cycles align with each wafer batch production run, real-time equipment parameter reporting, and daily report generation. This data supports daily work needs across process integration, module, equipment, yield, and manufacturing departments.
Constraints Imposed on Multi-Turn Dialogue and Prompt Engineering by These Characteristics
Data sources are scattered, and update cycles vary. Multi-turn dialogue must integrate document fragments from different sources based on update priority. Prompts must clearly distinguish the weight of real-time equipment parameters and historical batch records. Documents include unstructured content such as wafer maps and SPC charts. Multi-turn dialogue must support associative retrieval across document types. Prompts must limit retrieval scope to data associated with the current batch. The process chain is extremely long, and rework windows are limited. Multi-turn dialogue context length must accommodate associated information for hundreds of process steps. Prompts must clearly define the boundaries of attribution logic to avoid introducing interfering data from unrelated batches.
Configuration Settings
These configurations are written for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
messages context length | 10000–15000 characters | Accommodates the length of associated documents for hundreds of process steps in a single wafer batch, avoids truncating critical process parameters |
| Question Optimization | Enabled | Automatically corrects abbreviation ambiguity in process terminology to align with professional expression habits in wafer manufacturing |
variables binding scope | Only call MES data for the current batch | Ensures multi-turn dialogue only uses associated documents from the current production batch to avoid cross-batch interference |
chatId binding rule | Bind to current production batch ID | Ensures multi-turn dialogue context only associates documents from the current batch to avoid cross-batch confusion |
| Single retrieval limit | Top 8 entries | Balances retrieval coverage across multiple document types such as SPC charts and defect maps, avoids information overload |
| Conversation log retention period | 30 days | Matches the batch traceability cycle of wafer manufacturing to facilitate exception troubleshooting and process review |
Three Common Misconfigurations
- Phenomenon: Multi-turn dialogue output retains model-generated thinking content even if a code execution node filter is configured. Cause: The code node is not correctly mounted in the
chatId-bound dialogue flow, or the filter logic does not apply to assistant role content in themessagesarray. - Phenomenon: Knowledge base responses are truncated, only returning partial process parameters or defect analysis content. Cause: The
messagescontext length is set too low, exceeding the total length of associated documents for the current wafer batch, or the token limit of the model quantization version does not accommodate long document splitting. - Phenomenon: Prompt optimization does not take effect, and the dialogue still uses the original question wording. Cause: The question optimization configuration item is not enabled, or the optimization logic is not correctly bound to the
variablesparameter of the current dialogue.
How to Verify Proper Configuration
- Initiate a process attribution question targeting the current wafer batch, check that the output only includes MES and equipment FDC data from the current batch, with no cross-batch interfering content.
- View the log records in the dialogue interface, confirm that unnecessary thinking tag content has been filtered out of the
messagesarray, matching the configured requirements. - Trigger a retrieval request for an exception handling sheet, check that the number of returned results matches the single retrieval limit configuration, with no information overload.
- Adjust the context length configuration, verify that no truncation occurs during long document retrieval, ensuring critical process parameters are output completely.
The values provided are common starting points and should be measured against the reader's own samples.