What the data for this category looks like
Core data for semiconductor packaging and testing scenarios comes from wafer test maps, MES systems, bonding and molding equipment, ATE test machines, X-ray/SAM detection equipment, and customer specification documents. Daily circulating documents include packaging process cards, bonding and molding parameter sheets, incoming wafer maps, test program version files, yield reports, and failure analysis reports. Data is updated once per production batch, per chip test, and once daily. Document structures include standardized parameter fields, timing data output by equipment, visual detection bitmaps, and tabular data.
What constraints these characteristics impose on the document parsing and chunking workflow
Mixed circulation of multiple document types requires the parsing module to support mixed text and graphics, table extraction, and bitmap association. High-frequency updated batch data requires setting reasonable timeout and concurrency processing parameters for parsing tasks. The binding relationship between standardized parameter fields and visual detection images requires retaining contextual associations during chunking, to avoid breaking the logical connection between key parameter groups and corresponding detection bitmaps. Non-standard naming formats output directly by equipment requires adapting parsing to custom field prefixes and unit matching.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the paragraph length of packaging process cards and yield reports, avoids splitting key parameter groups |
chunkSplitter | Split by paragraph + heading | Matches the hierarchical structure of documents, retains contextual associations for bonding parameters and test program versions |
customPdfParse | Enabled | Correctly extracts embedded text and annotation information from X-ray/SAM scan images, prevents image parsing failures |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Adapts to the parsing duration of daily batch-uploaded yield reports and failure analysis reports |
UPLOAD_FILE_MAX_SIZE | 500 MB | Supports uploading archived files of entire batches of wafer test maps |
trainingType | GPU-accelerated | Adapts to the parsing requirements of batch long documents, can use multiple 3090 GPUs to improve concurrency processing efficiency |
Three common mistakes
- Phenomenon: After uploading a DOCX file containing incoming wafer maps, invalid image placeholders appear in the parsing results, or the log contains Invalid image-related errors. Cause: The
customPdfParseconfiguration is not enabled, resulting in failure to correctly extract and associate embedded bitmaps. - Phenomenon: Parsing tasks do not call GPU resources, and the running duration exceeds the preset threshold. Cause: The
trainingTypeconfiguration is not set to GPU-accelerated mode, and only CPU is used for parsing, which cannot meet the processing requirements of batch long documents. - Phenomenon: Embedded device parameter images in Excel-format bonding parameter sheets cannot be parsed, and no corresponding image association content appears in the results. Cause: The embedded image extraction logic during chunking is not configured, and only plain text content is extracted while ignoring associated images.
How to confirm the configuration is correct
- Upload a single DOCX file containing incoming wafer maps, check whether the parsing results correctly display image-associated text and bitmap content.
- View the running logs of the parsing task, confirm that GPU resources are normally called, which meets the configuration requirements of
trainingType. - Upload an Excel bonding parameter sheet containing embedded parameter images, confirm that the parsing results include text descriptions and associated parameters of the images.
- Batch upload 5 packaging process card documents, check that the chunking results do not break the association between key parameters and corresponding headings.
The values provided here are common starting points and should be measured against the reader's own samples.