What This Category of Data Looks Like
Typical entities include injection molding, stamping, and die casting mold enterprises, as well as mold centers of manufacturing enterprises. Daily generated documents include mold design task sheets, DFM reports, part drawings, processing process cards, mold trial reports, mold modification notices, and mold history records. Data comes from customer product models, mold flow or forming simulation files, CAD/CAM outputs, machining and measurement data, and mold trial samples. Documents are updated per project. New versions are generated for each mold trial or modification. Some documents contain structured parameters, drawing images, and process descriptions.
What Constraints These Characteristics Impose on the Document Parsing and Chunking Link
Mold documents have multiple formats and structures. They include plain text process cards and PDF reports with embedded CAD drawings. Parsing must adapt to different content types. The update rhythm tied to projects and mold trial iterations requires chunking to retain document version association information, to avoid confusion across versions. Core audit parameters such as demolding, shrinkage, and assembly accuracy are included in documents. Chunking must not split key parameters from their associated context. Some documents contain large numbers of tables and images. Multi-vector parsing and image content extraction must be supported, otherwise key audit basis will be lost.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Mold documents contain a large number of process parameters and section descriptions. This length preserves the association between parameters and their context, avoiding damage to logical structure during chunking |
customPdfParse | Enabled | Mold documents often include PDF files with embedded images exported from CAD and screenshots of mold trial reports. Enhanced parsing is required to extract image and text content |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Parsing large mold project CAD assembly files takes a long time. This range covers the parsing duration of most projects |
chunkSplitter | Split by heading + paragraph | Mold documents have clear section divisions. Splitting by heading preserves the independent audit logic of each section |
chunkSettingMode | Manual mode | Document structures vary widely across different mold projects. Manual configuration can adapt to different types of audit documents |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Large mold CAD assembly files may have large sizes. This upper limit meets the upload needs of most projects |
Three Common Mistakes
- Scenario: After uploading a mold document that includes part drawings, the parsing result only shows blank content or no image descriptions. Cause: The
customPdfParseconfiguration is not enabled. The default parsing method cannot extract image content embedded in PDF files. - Scenario: After deploying MinerU locally, the enhanced PDF parsing function is enabled but parsing still fails. Cause: The API address and port of MinerU are not configured in FastGPT's parsing service, or the corresponding port is not open in the local firewall.
- Scenario: The row and column association information of tables is lost in the parsed process card documents. Cause: Multi-vector support is not enabled, and the context structure of the table is not retained during chunking, making it impossible to associate corresponding parameters during audit.
How to Confirm the Configuration Is Correct
- Upload a mold trial report PDF including CAD drawings, and verify whether the parsing result extracts image descriptions and report text.
- Check the knowledge base settings in FastGPT v4.17.0, confirm that the
customPdfParseoption is enabled, or the corresponding English parameter is configured to the enabled state. - Upload a mold CAD export file of approximately 500 MB, confirm that the parsing task is not terminated due to timeout.
- View the parsed chunking results, confirm that table data is completely split and retains row and column association information.
The values provided are common starting points and should be measured against the reader's own samples.