What this type of data looks like
Daily data from injection molding, stamping, die-casting mold enterprises and mold centers of manufacturing enterprises mainly comes from customer product models, mold flow/forming simulation files, CAD/CAM exported drawings, machining and measurement data, and trial mold sample records. Document types include mold design task sheets, DFM reports, part drawings, machining process cards, trial mold reports, mold modification notices, and mold history records. Data updates follow project milestones, and are synchronized after each trial mold or mold modification. Documents include professional fields such as number of cavities, material grade, demolding angle, and machining allowance, with units mostly millimeters, Rockwell hardness, hours, and similar.
What constraints do these characteristics impose on document parsing and chunking
The mixed format and professional attributes of mold-related documents impose multiple constraints on parsing and chunking. Scanned trial mold reports and part drawing PDFs contain a large number of image-based parameters and hand-drawn annotations, so OCR recognition must be supported to fully extract content. Process cards and DFM reports contain structured tables and continuous professional descriptions. Chunking must retain the complete context of a single process to avoid splitting that disrupts process logic. Frequently updated project documents require the parsing process to have stable timeout handling capabilities, while adapting to differentiated parsing logic for different format files to avoid loss of key parameters due to format incompatibility.
How to set the configuration
For FastGPT v4.17.0, refer to the following configuration for document parsing and chunking in mold design and manufacturing scenarios:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Mold process cards contain multiple sections of professional descriptions and parameters; this length retains the complete context of a single process |
chunkSplitter | Split by paragraph + table | Adapts to the mixed text and structured table format of mold documents |
customPdfParse | Enable OCR mode | Process scanned trial mold reports and part drawing PDFs, extract process parameters embedded in images |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Large CAD-exported PDFs take longer to parse; prevents timeout interruptions to the complete parsing process |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Meets upload requirements for large CAD-exported drawings and mold flow simulation files |
chunkSettingMode | Manual configuration | Adapts to differentiated chunking needs for different types of mold documents, such as using different lengths for part drawings and process cards |
Three common mistakes
- Enabling the default PDF parser fails to recognize process parameters in scanned trial mold reports: This occurs because OCR mode for
customPdfParseis not enabled, and only the embedded text layer is parsed. - Splitting process parameters incorrectly when chunk length is configured by token count: This occurs because the professional term context of mold documents is not matched, and token counting fails to retain complete process step descriptions.
- No table content appears in search results after uploading an Excel process card: This occurs because table structured parsing is not enabled, and only plain text is extracted, leading to loss of fields.
How to verify correct configuration
- Upload a scanned trial mold report PDF, and check if parsed results include parameter values and text descriptions extracted via OCR.
- Upload an Excel process card with multiple tables, verify that search results can fully return content such as machining allowance and hardness requirements within the tables.
- View the file parsing log to confirm that no timeout errors appear after the
PARSE_FILE_TIMEOUT_SECONDSconfiguration is applied. - Test search recall with different chunk lengths to confirm that the complete content of a single process is not split.
The values given are common starting points and should be measured against the reader's own samples.