What Data Looks Like for This Category
Typical data for this category comes from CAD/CAM software, order BOMs, sheet metal inventory systems, laser cutter and bending machine device logs, and MES production management systems. Daily documents include part unfolding drawings, CNC cutting programs, bending process cards, nesting diagrams, first-piece inspection records, and process transfer cards. Data updates trigger when orders are released, daily production progress updates are made, or each equipment model change occurs. Document fields include bending compensation values, hole position coordinates, sheet material utilization rates, process transfer nodes, leftover material dimensions, and more, with engineering units such as millimeters, kilograms, and square meters.
Constraints for Document Parsing and Chunking
Multiple document formats are involved, including engineering drawings, CNC program text, and process cards. Parsing must distinguish structural features of different document types to avoid mixing drawing annotations and process descriptions. Data updates trigger per order or equipment model change. Chunk length must match the single-processing volume of individual documents to avoid update synchronization delays from overly large chunks. Professional engineering fields such as bending compensation values and sheet material utilization rates must retain their original units and formats. Chunking must ensure that field-associated information is not split. The multi-variety, small-batch business model requires chunking to aggregate data by order or process dimension, enabling quick association of corresponding data during subsequent ledger summary.
Configuration Settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Matches the content length of single-process descriptions, dimension parameters, and other content in sheet metal documents, avoiding splitting critical engineering fields |
chunkSplitter | Split by paragraphs and page headers/footers | Adapts to the structure of bending process cards and process transfer cards, preserving process order and associated information |
customPdfParse | Enable engineering drawing-specific parsing mode | Preserves drawing annotations and dimension information for unfolding drawings and nesting diagrams exported from CAD |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Accommodates parsing time for large nesting diagrams and multi-process process cards, avoiding timeout interruptions |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Accommodates full set of engineering documents for a single batch of orders, preventing upload failure due to oversized files |
Three Common Configuration Errors
- Issue: Uploading a PDF document exported from CAD results in parsing failure, with logs returning a connection timeout error. Cause: When deploying FastGPT locally, the marker-pdf service is not configured in the same Docker network as the main service, so the main service cannot access the locally deployed parsing service.
- Issue: Parsed ledger data cannot be exported to Excel format, with no response when the export button is clicked. Cause: The FastGPT file export module is not configured, or the field mapping rules for Excel export are not specified.
- Issue: Key bending compensation value fields are empty after parsing uploaded bending process cards. Cause: Context associated with fields is split during chunking, preventing the parsing model from recognizing complete parameter information.
How to Confirm Proper Configuration
- Upload a single bending process card PDF, and check if the parsed result includes complete process order and bending compensation value fields.
- Upload a 100 MB nesting diagram PDF, and confirm that the upload progress completes within the time set by
PARSE_FILE_TIMEOUT_SECONDS. - Attempt to export parsed ledger data to Excel format, and confirm that the exported file opens normally and includes all business fields.
- Adjust
chunkSizeto different ranges, compare field integrity of parsed results, and select a value suitable for current business documents.
The values provided are common starting points and should be measured against the reader's own samples.