What the Data for This Category Looks Like
The data for this category comes from wafer fab MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Daily documents include process flow cards, single-batch wafer operation records, equipment operation recipes, SPC control charts, exception handling sheets, defect maps, and yield reports. Data update rhythm syncs with single-batch wafer production, real-time parameter collection, and daily aggregation. Fields include process number, process parameters, defect coordinates, yield values, and more. Units include nanoscale process parameters, batch numbers, timestamps, and more.
Constraints Imposed on Knowledge Base Retrieval and Recall
The data characteristics of this category impose multiple constraints on the retrieval and recall link. Multi-source heterogeneous cross-department documents and real-time incrementally updated data require the retrieval mechanism to support multi-format file parsing and incremental synchronization to avoid data lag. The spatial distribution of wafer defects and the high-precision requirements of nanoscale process parameters require precise field association for retrieval matching, to avoid incorrect attribution caused by generalized matching. The limited rework window for single-batch wafers requires the latency of retrieval return results to adapt to production rhythm, while also covering the associated information of hundreds of processes and rework paths to support rapid query for cross-module yield attribution.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Aligns with the precision requirements of nanoscale process parameters, filters irrelevant data with low matching degrees |
limit | Top 10–15 entries | Covers associated information for hundreds of processes, while controlling context length to adapt to model processing capabilities |
searchMode | mixedRecall | Combines keyword and semantic retrieval, adapts to mixed structured and unstructured business documents |
usingReRank | Enabled | Performs secondary ranking on recall results to accurately select highly relevant documents such as defect maps and yield reports |
datasetSearchUsingExtensionQuery | Enabled | Expands process-related synonyms and parameter expressions to improve recall rate for cross-process retrieval |
collectionIds | Bind dedicated wafer fabrication knowledge base collections | Only retrieves knowledge bases corresponding to process, equipment, and yield modules, avoiding interference from irrelevant data |
Three Common Misconfigurations
- An
embedding erroroccurs when uploading non-text documents such as wafer maps and SPC charts. The cause is failure to configure document parsing plugins for corresponding formats, and failure to adapt to the unique image and table document structures of semiconductor manufacturing. - The number of retrieval results is lower than expected. The cause is that the
limitparameter is set too low, failing to cover cross-process associated data. - Irrelevant non-wafer fabrication documents are mixed in retrieval results. The cause is failure to bind dedicated
collectionIdsand failing to limit the scope of retrieved knowledge bases.
How to Verify Correct Configuration
Based on FastGPT v4.17.0, configuration verification can be completed through the following actions:
- Verify the value range of the
similarityparameter, confirm that it aligns with the precision requirements of nanoscale process parameters. - Test retrieval of single-batch wafer operation records, confirm that the number of returned results matches the
limitsetting. - Check that retrieval results only include wafer fabrication-specific documents, confirm that
collectionIdsare bound correctly. - After enabling
usingReRank, confirm that secondary ranking results prioritize highly relevant documents such as defect maps and yield reports.
The values provided are common starting points and should be evaluated against the target deployment's own samples.