What the Data for This Category Looks Like
Daily documents for this category include process sheets, layout plans, cutting orders, process quality sheets, packaging and size matching sheets, and similar materials. Data sources cover order and size tables, CAD patterns, fabric and accessory inspection records, workshop hanging system data, and customer sample sealing information. Data updates align with each order, each garment style, and each production batch. Most documents include fields such as size parameters, fabric specifications, process requirements, and packaging ratios. Some CAD-exported layout plans contain pixel-level cutting piece coordinate information. Supported file formats include Excel, scanned PDFs, and CAD native files.
Constraints Imposed on Document Parsing and Chunking
Multi-source, heterogeneous document formats require the parsing module to support CAD-exported files, scanned PDFs, and Excel spreadsheets. Frequent order-level updates require chunking logic to support fast incremental synchronization. Precise fields such as cutting piece coordinates and size ratios must avoid cross-document chunk breaks. Multiple process entries in a process quality sheet must be split by process nodes. This prevents random chunking from breaking process relevance. Mixed uploads of scanned documents and CAD files require the parsing process to distinguish text and graphic content. This avoids misidentifying cutting piece coordinates as invalid text. The granularity of individual chunks must also be controlled to match the precise retrieval needs of process Q&A.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Matches the single-process text length of process sheets and process tables, avoids splitting cross-process content |
chunkSplitter | Split by paragraph + table header | Adapts to the header + process entry structure of process sheets, preserves field relevance |
customPdfParse | Enable OCR + text priority mode | Compatible with scanned process manuals and CAD-exported PDFs, prioritizes extracting structured text |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Supports batch uploads of layout plan CAD files and historical order documents for single batches |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Adapts to the parsing time of large CAD files, avoids timeout interruptions |
chunkSettingMode | Manual configuration | Adjust chunking rules for different document types (Excel/scanned PDF/CAD) to adapt to category-specific document structures |
Three Common Configuration Errors
- Phenomenon: Size fields cannot be retrieved after uploading an Excel process sheet. Cause: Excel chunking by table header row is not configured, causing cross-header fields to be merged into a single chunk, preventing accurate field matching during retrieval.
- Phenomenon: No valid text content appears after uploading a scanned layout plan. Cause: OCR mode for
customPdfParseis not enabled. Only native PDF text is recognized, so text information in scanned documents cannot be extracted. - Phenomenon: Cutting piece diagrams are lost after parsing a graphic-and-text combined process manual. Cause: Graphic and text chunks are not distinguished. When images are embedded in text chunks, only text content is returned during retrieval, and the corresponding cutting piece diagrams cannot be associated.
How to Verify Correct Configuration
For FastGPT v4.17.0, use the following steps to check configuration effectiveness:
- Upload a single Excel process sheet, retrieve the "shoulder width" field, and confirm the returned result contains the complete corresponding process content.
- Upload a scanned first-piece sample sealing record, view the parsed text chunks, and confirm they include the appearance description text of the sealed sample.
- Upload a single CAD-exported layout plan PDF, check that parsing time stays within the preset threshold, with no timeout errors.
- Upload a graphic-and-text combined process manual, confirm that the chunking result retains the association markers between text and corresponding images.
The values provided are common starting points and should be measured against the reader's own samples.