What the data for this category looks like
Daily documents for semiconductor wafer fabrication include process flow cards, batch operation records, equipment recipes, SPC 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 data. Data update cadence is that exclusive operation records are generated for each wafer batch, equipment parameters are synchronized in real time, and yield reports and process adjustment documents are updated daily. Document fields include wafer batch numbers, nanoscale process parameters, defect spatial coordinates, equipment numbers, and some documents include structured parameter curves and image data.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The ultra-long process chain and nanoscale parameters of this category require retrieval to balance precise matching and breadth of coverage. Real-time and batch-level updated data from multiple sources requires the retrieval system to support precise filtering by wafer batch and equipment number, to avoid interference from invalid data across batches. Structured parameter curves and defect maps require retrieval to support both semantic matching and keyword matching, ensuring that queries about parameter drift or defect distribution can retrieve corresponding documents. The cross-module yield attribution requirement requires recall results to cover associated documents from multiple processes, while controlling the token consumption of a single retrieval to fit the time limits of the workflow.
How to set the configurations
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision of nanoscale parameters and defect descriptions in wafer manufacturing, avoiding recall of irrelevant general process documents |
limit | Top 8 entries | Covers sufficient context for cross-module yield attribution, while controlling token consumption per retrieval round |
searchMode | mixedRecall | Balances semantic matching for process parameters and precise keyword matching for wafer batch numbers and equipment numbers |
usingReRank | Enabled | Rearranges retrieved results by relevance to defect maps and yield reports, raising the priority of core defect analysis documents |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonymous queries for wafer batch numbers and process steps, such as expanding "etch process" to "plasma etching process" |
collectionIds | Bound by business module | Only retrieves knowledge base collections related to process integration, equipment, and yield, avoiding inclusion of irrelevant documents |
Three common errors
- Insufficient number of retrieval results, failing to cover defect analysis basis across all processes. This occurs because the
limitparameter is set too low, not covering process-associated documents across modules. - Retrieved content does not match the current query, incorrectly returning yield reports from other batches. This occurs because keyword matching for
searchModeis not enabled, relying solely on semantic recall which leads to batch number confusion. - Excessive retrieval time causing workflow timeouts. This occurs because the
collectionIdsscope is not restricted, retrieving the full knowledge base collection and increasing processing overhead for invalid data.
How to confirm the configuration is set correctly
- Enter a query that includes a known defect type and wafer batch number, check whether the retrieved results include the corresponding defect map and operation records for that batch.
- Check the order of retrieval results, confirm that after enabling
usingReRank, core defect analysis documents appear at the top. - Test the time taken for a single retrieval, confirm it meets the time requirements of the workflow.
- Check the retrieval scope, confirm that only knowledge base documents under the bound business module are returned, with no content from irrelevant collections included.
The values provided are common starting points and should be measured against the reader's own samples.