What this category’s data looks like
Semiconductor wafer fabrication process documents typically originate from MES systems, equipment FDC modules, defect detection metrology equipment, process recipe libraries, and experimental data. Daily documents include process flow cards, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect maps, and yield reports. Update frequencies cover per wafer batch, real-time, and daily. Document fields include nanoscale process parameters, wafer coordinates, batch numbers, CPK values, and similar items. Most document structures are interrelated tables, time-series data, and mixed text-image formats.
How These Document Characteristics Impact Parsing and Chunking
The characteristics of semiconductor wafer fabrication documents create multiple constraints for parsing and chunking: Ultra-long process chains coupled with nanoscale parameters require chunking to retain contextual connections between processes, avoiding excessive splitting that disrupts process logic. Multi-source real-time updated documents need to support fast batch parsing to prevent timeouts for large files. Mixed-format text-image-table documents require structured parsing capabilities to accurately extract parameter and coordinate associations. Excel-format process records must be split by individual process rows, avoiding semantic breaks caused by cross-row merging.
Configuration Settings
These configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Matches the length of a single process step or parameter group in wafer fabrication documents, avoiding splitting that disrupts process context connections |
chunkSplitter | Custom delimiter + semantic chunking | Adapts to single process rows in Excel and SPC chart parameter groups, preventing cross-row merging or semantic breaks |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Handles equipment recipe libraries or batch record documents sized 10–20 MB, preventing large file parsing timeouts |
UPLOAD_FILE_MAX_SIZE | 20 MB | Compatible with complete run record packages for a single wafer batch, meeting batch upload requirements |
customPdfParse | Enable table structured parsing | Accurately extracts numerical and coordinate fields from SPC charts and defect maps, avoiding loss of associated information due to text-only conversion |
Three Common Configuration Mistakes
- Merged rows appear in chunking results after uploading Excel-format process records. This occurs when custom delimiters are not enabled, and line-ending newline characters are not configured as the splitting basis.
- Parsing timeouts occur when uploading Word-format equipment recipe documents larger than 10 MB. This occurs when
PARSE_FILE_TIMEOUT_SECONDSis not adjusted to a duration suitable for large files, or streaming parsing optimization is not enabled. - Associations between CPK values and time series are lost in parsing results after uploading a PDF document containing SPC charts. This occurs when the table structured parsing function of
customPdfParseis not enabled, causing charts to be misidentified as plain text.
How to Verify Proper Configuration
- Upload an Excel test file with a single process step. Check if each chunked result corresponds to one independent process record per row.
- Upload an equipment recipe document of approximately 15 MB. Confirm that parsing completes within the duration set by
PARSE_FILE_TIMEOUT_SECONDS. - Upload a PDF document containing SPC control charts. Verify that the parsed result retains the association between CPK values and their corresponding time series.
- Enter
\nas the custom delimiter for testing. Confirm that the chunking result after importing Excel matches expectations.
The values provided are common starting points and should be evaluated against the reader’s own samples.