What Data Looks Like for This Use Case
In semiconductor wafer fabrication scenarios, data primarily comes from 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, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect atlas, and yield reports. Data update cycles cover each wafer batch transfer, real-time parameter reporting, and daily summary statistics. Fields include process parameters, defect coordinates, yield values, and batch identifiers. Units involve nanoscale dimensions, timestamps, and batch numbers.
Constraints on Citation and Traceability
Multi-source heterogeneous data, layered update cycles, and precise process parameter characteristics for this use case impose multiple constraints on the citation and traceability link. First, nanoscale parameters and cross-process coupled process logic require precise matching of batch identifiers and corresponding process nodes during traceability, to avoid misjudgments caused by parameter drift. Second, the coexistence of real-time equipment data and daily summary reports requires support for dynamic indexing and historical data retrieval. Third, diverse document formats and field structures require the traceability module to support associated calls for structured parameter tables, visual wafer maps, and text-based handling records. This ensures that citation sources can be fully mapped to original production data.
Configuration Settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
quoteMaxToken | 1500–2000 token | Accommodates complete wafer batch run records, defect atlas, and process parameter content, to avoid truncation of citation sources |
responseData | Enabled | Retains batch identifiers and field information from original production data, to ensure traceability can be linked to specific production stages |
detail | Enabled | Displays specific process nodes and update times of citation sources, to match the cross-process audit logic of wafer fabrication |
| Chunk length | 800–1200 characters | Adapts to the chunking requirements of structured process documents and visual wafer maps |
| Overlap length | 100–200 characters | Covers continuous equipment parameter drift records, to prevent key associated information from being cut off by chunking |
| Similarity threshold | 0.75–0.85 | Precise matching of nanoscale process parameters and batch information, to avoid incorrect association of data from different wafer batches |
Common Configuration Errors
- Generated responses fail to link to specific batches and parameters of original production documents, and citation source fields are empty. This occurs when the
responseDataconfiguration is not enabled, causing the system to fail to capture identifier information from original data. - Question-and-answer response timeouts occur, with no service return for extended periods. This happens when
quoteMaxTokenis set too large, exceeding system processing limits, and the corresponding context loading threshold is not adjusted. - Incorrect wafer batch data is matched, leading to deviations in audit conclusions. This occurs when the similarity threshold is set too low, causing parameters from non-target batches to be incorrectly associated.
How to Verify Correct Configuration
A quality audit query for a specific wafer batch is initiated. The citation source module at the end of the response is reviewed to confirm inclusion of batch identifiers and corresponding document types. The content length of citation sources is checked to confirm no truncation has occurred, and that the content matches the corresponding fragment of the original document. The similarity threshold is adjusted, and the query is re-initiated. Changes in the number of matching results are observed to verify that the threshold configuration is effective. System logs are reviewed to confirm that citation source parameter fields have been correctly extracted and displayed.
The values provided are common starting points and should be measured against the reader's own samples.