What the Data for This Category Looks Like
Daily data sources for this category include wafer fab MES systems, equipment FDC modules, defect detection and metrology equipment, and process recipe libraries. Daily documents cover process flow cards, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect maps, yield reports, and more. Fields include wafer batch identifiers, process parameter thresholds, defect spatial coordinates, yield proportions, and similar metrics. Units include nanoscale parameters, batch numbers, ppm, and other standard units. Data updates follow a multi-tempo schedule: batch records are generated for each wafer lot, equipment parameters are synced in real time, and yield reports are aggregated daily. This covers multi-dimensional process and exception-related information.
Constraints Imposed on Knowledge Base Retrieval and Recall
The ultra-long process chain and nanoscale parameter coupling characteristics of this category require retrieval and recall to cover associated documents across hundreds of processes, to prevent process deviations caused by taking content out of context. Real-time updated equipment FDC data and daily aggregated yield reports require index refreshes to adapt to multi-tempo updates, avoiding recall of outdated data that disrupts exception judgment. The limited rework window requires retrieval latency to be controlled within a reasonable range. Cross-module yield attribution needs also require recalling associated content across equipment and processes, to improve attribution accuracy.
How to Configure Settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
similarity | 0.85–0.92 | Meets the high-precision matching requirements of nanoscale parameters, and avoids recalling low-relevance process documents |
limit | Top 8–12 entries | Covers associated content for hundreds of processes, while controlling context length to fit model input limits |
searchMode | mixedRecall | Balances semantic matching for nanoscale parameters and exact matching for process numbers and batch numbers |
usingReRank | Enabled | Performs secondary ranking on retrieval results, corrects deviations from vector retrieval, and adapts to cross-module yield attribution association requirements |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands process terminology and exception descriptions to match professional abbreviations and industry terms in wafer fabrication scenarios |
collectionIds | Grouped by department/document type | Precise limitation of retrieval scope, avoids mixing unrelated documents from packaging and testing categories |
Three Common Mistakes
- Phenomenon: When initiating retrieval during a multi-turn conversation, recall results include process parameters from non-current wafer lots. Restarting a new conversation restores accurate results. Cause: Multi-turn conversation context accumulates redundant information from non-current process scenarios, leading to irrelevant historical conversation content being mixed in during retrieval and interfering with vector matching logic.
- Phenomenon: After replacing the embedding model, retrieval results from the existing knowledge base deviate or throw errors. Cause: The vector index for the imported knowledge base was not regenerated, and matching still uses the vector space from the old model, resulting in incompatibility between the vector spaces of the new and old models.
- Phenomenon: Retrieval results include packaging and testing documents that do not match the current wafer fabrication scenario. Cause: The
collectionIdsparameter was not used to limit the knowledge base set for retrieval, leading to unrelated cross-category documents being included in the recall scope.
How to Verify Proper Configuration
- Review retrieval logs to verify that recall results include wafer fabrication-specific documents such as process flow cards and SPC charts, with no irrelevant content from other categories.
- Initiate retrieval again after a multi-turn conversation, confirm that result accuracy does not decline significantly.
- After replacing the embedding model, perform a knowledge base index rebuild operation, and verify that retrieval result matching meets expected levels compared to before the replacement.
- Adjust the
similarityparameter, verify that recall result relevance follows preset logic as the parameter is adjusted.
The values provided are common starting points and should be measured against the reader's own samples.