What the Data for This Category Looks Like
Daily documents for semiconductor wafer fabrication include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, and yield reports. Data sources cover MES systems, equipment FDC, defect detection and metrology devices, wafer maps, process recipe libraries, and experimental data. Update cadence aligns with each wafer batch production run, real-time parameter reporting, and daily statistical aggregation. Document fields include wafer batch numbers, process step parameters, defect coordinates, yield values, and more. Some parameters use nanoscale units. Data relevance extends layer by layer along the process chain.
Constraints on Knowledge Base Retrieval and Recall
High-frequency real-time updated data requires retrieval to support incremental synchronization. This prevents recall of outdated process parameters or batch records. Nanoscale parameter and defect spatial distribution matching requires retrieval to cover both semantic relevance and exact keyword matching. Cross-process and cross-module yield attribution requires retrieval to pull associated documents across multiple business datasets. Limited rework windows require retrieval response speed to match production rhythm. This avoids delays that disrupt exception handling.
How to Configure
These settings are optimized for FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets high-precision matching requirements for semiconductor process parameters, filters out low-relevance equipment recipes or exception records |
limit | Top 10 entries | Aligns with the number of process steps per wafer batch, avoids context overload that reduces cross-process attribution efficiency |
searchMode | mixedRecall | Combines semantic recall and exact keyword matching for process parameters, supports retrieval of SPC charts and defect maps |
usingReRank | Enabled | Reranks recalled process documents by yield relevance, improves accuracy of cross-module yield attribution |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands query terms for process-related questions, covers exclusive terminology for equipment FDC parameter drift |
collectionIds | Divided by wafer fab business modules | Binds dedicated collections for MES, equipment libraries, defect libraries, and other systems, enables precise cross-module retrieval |
Three Common Configuration Mistakes
- Symptom: Knowledge base retrieval calls return empty results, or fail to match the latest batch run records. Cause:
datasetSearchUsingExtensionQueryis not configured, so exclusive process parameter terminology is not covered, leading to failed recall. - Symptom: Attempting to delete specified knowledge base content leaves the content intact. Cause: Correct dataset operation permission configuration is not used in FastGPT v4.17.0, or the specified deletion rule is not triggered.
- Symptom: Nested searches cannot return cross-business module process documents. Cause:
collectionIdsis not correctly configured to bind multiple business datasets, so only content from a single knowledge base is retrieved.
How to Verify Successful Configuration
- Upload a new batch of batch run records, trigger the retrieval process, and confirm the recall results include the newly uploaded process data.
- Perform a deletion operation on the specified knowledge base, and verify backend logs show the corresponding content has been cleared.
- Initiate a cross-module process attribution query, and confirm returned results include associated documents from both the equipment library and defect library.
- Test the nested search function, and verify it can simultaneously retrieve knowledge base content bound by multiple specified
collectionIds.
The values provided are common starting points and should be measured against the reader's own samples.