What This Category of Data Looks Like
Semiconductor wafer fabrication inspection and process record data comes from multiple sources. These include MES systems, equipment FDC monitoring data, defect detection and metrology tools, wafer map files, process recipe libraries, and experiment records. Data update cycles cover three scenarios: each wafer batch generation, real-time equipment parameter uploads, and daily yield report aggregation. Daily documents include process flow cards, batch run records, equipment recipe parameters, SPC control charts, exception handling sheets, defect maps, and yield analysis reports. Data fields include wafer batch numbers, process node parameters, defect coordinates, yield ratios, and more. Some parameters require nanometer-level precision.
Constraints on Source Citation and Traceability
Ultra-long process chains and multiple rework paths require traceability to tie to unique wafer batch numbers and corresponding process nodes. This avoids cross-batch confusion. Nanometer-level process parameters and defect spatial distribution data require citation and traceability to accurately match field precision and coordinate information. Fuzzy matching cannot be used. Real-time updated equipment FDC data and daily aggregated yield reports require indexes to support a combination of incremental updates and full retrieval. Wafer fabrication inspection runs through the entire production process. Traceability must cover every operation record from material input to before final product. It must also support associated retrieval of multiple document types.
How to Configure
The following configurations follow the official setting specifications for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Single Search Limit | Top 8 results | Wafer fabrication has a large number of process steps. Too many search results will increase context redundancy and reduce attribution analysis efficiency |
| Reranking | Enabled | Process data has strong hierarchical associations. Reranking prioritizes returning process and batch records most relevant to the current query |
quoteMaxToken | 1200–1500 tokens | Process parameters and defect description texts for wafer fabrication are lengthy. Sufficient tokens must be reserved to fully display traceability data |
responseData | Return cited entry ID and source metadata | Meets engineers' need for traceability of specific document paths and batch numbers. It matches the precise positioning requirements of the business scenario |
| Maximum Context Length | 8000–10000 characters | Covers complete records of a single batch with multiple processes. It avoids truncation of key traceability information |
| Citation Display Format | Sorted by source document category | Allows engineers to quickly locate associated inspection and record data by process and equipment type |
Three Common Misconfigurations
- Symptom: The citation list returned in a conversation only shows document names. It does not include unique identifiers such as wafer batch numbers and process step numbers. Cause: The
responseDataparameter is not configured to return complete metadata fields of source documents. Precise batch traceability is not possible. - Symptom: No cited document ID or path is returned after calling the conversation interface. Cause: The return switch for
responseDatais not enabled. The default configuration only returns generated response text. It cannot meet engineers' traceability and positioning needs. - Symptom: Retrieved citation data includes irrelevant cross-batch records. This leads to deviations in yield attribution analysis. Cause: No wafer batch filtering conditions are bound, or the single search limit is set too high. Precise screening using the hierarchical association of process data is not performed.
How to Verify Proper Configuration
- Initiate an inspection query targeting a specific wafer batch. Check if the returned citation list includes running records, process parameters, and other documents for that batch. Confirm the list also includes corresponding batch numbers and process information.
- Review the response results of the conversation interface. Confirm that the
responseDatafield exists and includes the ID, source path, and metadata information of cited documents. - Adjust the value of the
quoteMaxTokenparameter. Observe the truncation of citation texts. Verify that process parameters and defect description content are fully displayed. - Simulate a query for an abnormal process step. Verify that the retrieval results only return associated documents for the target batch. No cross-batch irrelevant data should appear.
The values provided are common starting points and should be measured against the reader's own samples.