What the data for this category looks like
The bid and quotation documents for semiconductor wafer fabrication are primarily sourced from MES systems, equipment FDC collectors, defect inspection metrology equipment, process recipe libraries, and experimental databases. Daily documents include process flow cards, batch run logs, equipment recipes, SPC statistical charts, exception handling sheets, defect maps, and yield reports. Data update frequencies align with per wafer batch production cycles, real-time equipment parameter collection, and daily summary report generation. Document fields include wafer batch identifiers, process step parameters, equipment models, defect locations and sizes, recipe thresholds, and other items. Units involve measurement standards such as nanometers, milliamps, and seconds.
What constraints do these characteristics impose on document parsing and chunking?
This category's document traits create multiple constraints for parsing and chunking:
- Ultra-long process chains and multi-dimensional nanoscale parameter-coupled documents require retaining contextual associations between parameters, and avoid splitting process parameter pairs or cross-process logical sequences.
- Documents containing non-pure text content such as SPC charts and defect maps require support for structured chart data extraction and text linking.
- Frequently updated real-time data and large summary reports require adaptation for large-file parsing and reasonable timeout settings.
- Non-uniform document formats across multiple sources require custom parsing rules to match different document structures, ensuring complete data required for bidding is accurately extracted.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Semiconductor wafer fabrication documents contain nanoscale process parameters and process sequences. Values that are too long will lose context, and values that are too short will split parameter pairs |
chunkSplitter | Semantic block + header/footer filtering | Documents include multi-batch process flow cards and yield reports. Splitting by semantics preserves complete data for a single process or single batch |
customPdfParse | Enabled, calibrated via actual testing | Non-standard PDF formats such as SPC charts and defect maps require adaptation to extract chart data and associated text |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Parsing large yield reports and complete batch process records takes a long time. This avoids timeout interruptions |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Complete equipment FDC data or defect map documents for a single batch may exceed conventional sizes, requiring adaptation for large document uploads |
chunkSettingMode | Manual mode | Chunking strategies need adjustment for different bidding document types such as process libraries and exception handling sheets, to meet integration requirements |
Three Common Mistakes
- Phenomenon: The PDF chunking preview interface displays "Unable to read this file content". Cause: The
customPdfParseconfiguration is not enabled, so structured data and associated text in non-pure text PDFs such as SPC charts and defect maps cannot be extracted. - Phenomenon: Complex process formulas cannot be parsed normally, and returned results are empty or garbled. Cause: The deployed FastGPT version does not match the supporting PDF parsing plugin version, and the local low version does not adapt to the special formula format in the semiconductor field.
- Phenomenon: Dependency errors occur when starting the Marker parsing image in Docker. Cause: GPU resources are not mapped correctly, or the image version conflicts with the CUDA version of the host environment.
How to confirm proper configuration
- Use the deployment environment of FastGPT v4.17.0, upload a PDF of a wafer process flow card containing SPC charts and process parameter tables, and check if the chunking preview can extract chart data and text content.
- Upload an equipment recipe document containing complex process formulas, and check if the parsing result completely retains the formula structure and parameter information.
- Upload a single yield report of common business size, check if the parsing process times out, and if the chunking result covers all document content.
- Adjust the
chunkSizeparameter, compare retrieval matching degrees under different chunk lengths, and confirm that it meets the integration requirements of bidding documents.
The values provided are common starting points and should be measured against the reader's own samples.