What this category of data looks like
Primary users include wafer fab process integration, equipment, yield, and manufacturing departments. Data sources include MES systems, equipment FDC, defect detection and metrology devices, wafer maps, process recipe libraries, and experimental data. Daily documents cover process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, and yield reports. Data updates synchronize with each wafer batch production, real-time parameter collection, and daily report generation. Fields include nanoscale process parameters, wafer batch IDs, defect density, equipment operating temperatures, and additional metrics. Common units are degrees Celsius, inches, and per square centimeter.
What constraints do these characteristics impose on the knowledge base retrieval and recall workflow
Multi-source heterogeneous and frequently updated data requires the retrieval pipeline to support incremental synchronization and multi-format document adaptation. The coupling of ultra-long process chains and nanoscale parameters requires retrieval to cover the full process scope and accurately match parameter-related content. The production rhythm of each wafer batch requires retrieval response to align with workshop cycles, to avoid delays that impact process adjustments. Data errors can lead to serious consequences such as wafer scrap, so recall results must undergo relevance filtering to reduce false match risks.
How to set the configuration
This configuration is formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the high-precision requirement for nanoscale process parameters, filters irrelevant documents with low similarity |
limit | Top 20 entries | Covers the retrieval scope of hundreds of process steps in wafer fabrication, ensures recall of key process steps |
searchMode | fullTextRecall | Adapts to cross-field matching requirements for multi-format documents such as process flow cards and defect maps |
usingReRank | Enabled | Filters retrieval noise caused by parameter coupling, improves result relevance |
datasetSearchUsingExtensionQuery | Enabled | Supplementary extended queries for process terms, associates synonymous process expressions |
collectionIds | Bind to the dedicated wafer fabrication knowledge base | Isolates documents from other categories, avoids interference from cross-domain retrieval |
Three common misconfigurations
- Scenario: Retrieval returns "Unable to read the uploaded xlsx format equipment recipe file". Cause: No content extraction rules for spreadsheet documents are configured, or the table parsing extension configuration is not enabled. This prevents process recipe files from being parsed normally.
- Scenario: Retrieval results only return a small amount of process data, which fails to cover the full process chain. Cause: The value of
limitis set too low, failing to meet the ultra-long process retrieval requirement of wafer fabrication. - Scenario: Retrieval results include packaging and testing related documents. Cause: No dedicated
collectionIdsare bound, resulting in recall of data from cross-category knowledge bases.
How to confirm the configuration is correct
- Upload an SPC chart document with known parameters, initiate a query containing the corresponding process parameters, and verify that the returned results include matching content.
- View the retrieval log to confirm that the
usingReRankanddatasetSearchUsingExtensionQueryfunctions are triggered normally. - After binding the dedicated knowledge base, initiate a query unrelated to wafer fabrication to confirm no irrelevant documents are recalled.
- Initiate a query containing nanoscale parameters, and verify that the matching degree of the returned results aligns with the preset
similaritythreshold range.
The values provided here are common starting points, and should be evaluated against the reader's own samples.