This content is based on the configuration logic of FastGPT v4.17.0.
What Data for This Category Looks Like
Typically, the entities in sheet metal cutting and bending scenarios are non-standard manufacturers including those producing cabinet chassis and ventilation equipment. Daily documents include unfolded drawings, CNC cutting programs, bending process cards, nesting drawings, first-piece records, and process flow cards. Data mainly comes from CAD/CAM software, order BOMs, sheet metal inventory systems, laser cutters, bending machines, and MES systems. Update frequency adjusts based on orders, daily cycles, or equipment changeovers. Most document fields include sheet thickness, bending compensation values, hole coordinates, cutting dimensions, and similar items. Units are mostly common industrial units such as millimeters and kilograms.
What Constraints Do These Characteristics Impose on the "Document Parsing and Chunking" Link
The multi-variety, small-batch business model results in diverse document types and large differences in single-file size. It is necessary to adapt to multiple formats such as CAD-exported PDFs, plain-text CNC programs, and scanned drawings. The high-frequency update business rhythm requires an efficient parsing process to avoid order delivery delays caused by timeouts. The strong correlation between professional fields and process steps requires retaining contextual associations during chunking, and must not separate bending parameters from corresponding processes. In addition, some unfolded drawings are scanned versions, requiring additional processing of text and dimension information in images, which increases parsing complexity.
How to Configure Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Sheet metal documents contain professional process parameters and continuous process steps. This range balances contextual integrity and retrieval accuracy, and is calibrated via actual testing |
chunkSplitter | Split by document type and header markers | Distinguish different document types such as unfolded drawings, process cards, and CNC programs to avoid mixed arrangement of cross-category content |
customPdfParse | Enable OCR mode | Adapt to scanned unfolded drawings, extract information such as bending coefficients and hole coordinates from images |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Large nesting drawings and CNC programs take a long time to parse. This duration covers most large file parsing requirements |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Support uploading multiple large CAD drawings for a single order, avoid interception due to oversized files |
Three Common Mistakes
- After uploading documents from multiple orders, search matches process parameters from unrelated orders. Cause: Chunking rules are not configured by document type, leading to unfolded drawings and process cards from different orders being merged into the same chunk context, making it impossible to distinguish sources during retrieval.
- Knowledge base search tests fail to hit target bending compensation parameters. Cause:
chunkSizeis set outside the reasonable range, either destroying the contextual association of process steps or causing information overload within chunks, making embedding vectors unable to accurately match search keywords. - Scanned unfolded drawings return empty fields or garbled text after parsing. Cause: OCR mode for
customPdfParseis not enabled, only parsing the text layer of the PDF, unable to extract dimension and process information from images.
How to Confirm the Configuration Is Correct
- Upload a single bending process card and multiple nesting drawings, check if parsed chunks are automatically categorized by document type, with no mixed content across orders or documents.
- Upload a scanned unfolded drawing, confirm that the parsed result includes professional fields such as sheet thickness and bending compensation values, with no garbled text or empty values.
- Upload the largest size CAD drawing, check that the parsing task does not trigger a timeout error, and the status shows completed.
- Simulate a knowledge base search, enter specific bending parameters or process keywords, confirm that the search results include content from the corresponding chunks.
The values given are common starting points and should be measured against the reader's own samples.