What the data for this category looks like
The data for this category covers daily outputs from departments including process integration, equipment, and yield rate at wafer fabs. It includes documents such as process flow cards, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect maps, and yield reports.
Data sources include MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Updates sync with each wafer batch production, real-time parameter collection, and daily aggregation.
Fields include process number, wafer batch ID, parameter drift threshold, defect coordinates, and more. Units are mostly nanoscale parameters, batch numbers, and timestamp formats.
Constraints Imposed on Source Citation and Traceability
The multi-source heterogeneous data and high-frequency update rhythm of this category require that citation traceability must be bound to a unique wafer batch ID and process number, to avoid mixing data across batches.
The coupling of ultra-long process chains and nanoscale parameters requires that traceability results must accurately match the parameter thresholds and defect coordinates of the corresponding process, and cannot rely on fuzzy matching.
Additionally, cross-department document association requirements mean the traceability module must simultaneously associate MES records, equipment FDC data, and defect maps, to ensure the citation chain of exception reports fully covers the entire production process.
How to Configure the Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
quoteMaxToken | 800–1200 token | Matches the average length of a single wafer process document, to avoid truncating core process parameters |
responseData | Return only associated fields | Focuses on traceability key fields such as wafer batch ID and process number, reducing invalid data loading |
detail | Enable full traceability mode | Retains the original file path and update timestamp of the data source, meeting compliance requirements for production traceability |
| Single search limit | Top 6 entries | Covers core process data involved in a single batch production, avoiding interference from too many low-relevance results |
| Context window | 1500–2000 token | Adapts to the stitching needs of multi-source heterogeneous data, ensuring the citation chain of exception reports is complete |
| Reordering logic | Sort by batch priority | Prioritizes matching data sources from the current production batch, improving traceability accuracy |
These configurations are based on the functional logic of FastGPT v4.17.0.
Three Common Configuration Mistakes
- Phenomenon: The number of returned citations is less than the configured single search limit, and does not cover core process data for the current production batch. Cause: The wafer batch ID is not used as a filter condition, leading to search results mixing invalid data from historical batches.
- Phenomenon: The sorting of citation results does not meet business expectations, and does not prioritize matching parameters from the current production batch. Cause: The reordering logic set to sort by batch priority is not configured, and only general similarity sorting is used.
- Phenomenon: Generated process documents or exception reports include unclosable knowledge base citation links. Cause: The citation display switch is not turned off in the workflow's output node, causing all generated content to automatically include traceability identifiers.
How to Verify Correct Configuration
- Import a simulated wafer batch run record, initiate a search request, and check if the returned citation results only associate documents from this batch.
- Generate a draft exception report, and verify that the citation traceability includes the original data's file path and update timestamp.
- Test search requests for different production batches, and confirm that results prioritize matching the configured batch priority sorting.
- Turn off the citation display switch, generate a process document again, and check that the generated content does not include any knowledge base citation links.
The values given are common starting points and should be measured against the reader's own samples.