This content is based on the configuration logic of FastGPT v4.17.0.
What Data for This Category Looks Like
Daily documents for general mechanical equipment assembly scenarios mainly include assembly process cards, BOMs, torque records, gap adjustment records, commissioning reports, factory inspection sheets, packing lists, and more. Data typically comes from PLM/ERP, part inspection data, assembly tool collection, test benches, MES, and customer technical agreements. The update frequency is once per device or per order. Most document structures include hierarchical process descriptions and precise numerical fields, such as torque values (N·m), gap values (mm), and coaxiality tolerances (μm). Some documents include tabulated bill of materials or inspection records.
What Constraints Do These Characteristics Impose on Document Parsing and Chunking
Engineer-to-order assembly models result in document structures that vary per order, requiring flexible adaptation to different process descriptions and field combinations. Context integrity is required to link precise numerical fields, so chunking must not break the binding between processes and corresponding inspection data. High-frequency per-device or per-order updates demand fast response from the parsing workflow to avoid delaying assembly progress. Large document packages require reasonable upload and parsing limit settings to ensure full documents can be imported normally.
How to Set the Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Aligns with the paragraph length of single processes in assembly process cards, retains context for key numerical fields such as torque and gap values |
chunkSplitter | Split by paragraph + title | Matches the structure of process titles and operation steps in process cards, prevents splitting from breaking linked descriptions of coaxiality and preload requirements |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | Covers parsing time for combined documents of single assembly process cards and BOMs, avoids parsing timeout for large-volume packing lists |
customPdfParse | Enabled | Preserves table structures in PDFs, accurately extracts hierarchical fields of BOMs and numerical items from inspection sheets |
UPLOAD_FILE_MAX_SIZE | 500 MB | Matches the upload limit for full assembly document packages (including multiple inspection records) for a single device |
chunkSettingMode | Manual configuration | Adjusts chunking parameters for document structures of different orders, meets the needs of multi-variety, small-batch assembly scenarios |
Three Common Mistakes
- Phenomenon: Table fields are lost in parsed results after uploading an assembly process card PDF. Cause: The
customPdfParseconfiguration is not enabled. The default parsing mode cannot retain the hierarchical structure of tables, leading to loss of numerical fields in BOMs and torque records. - Phenomenon: Image parameters cannot be passed correctly when calling the parsing API, and the interface returns a parameter error. Cause: Image files are not encapsulated in multipart/form-data format. Submitting images as JSON strings violates interface requirements.
- Phenomenon: Chunked data stored in the knowledge base cannot be linked to inspection information for the corresponding order. Cause: The
chunkSizesetting is too small, splitting a single gap adjustment record into multiple chunks and breaking the contextual link between gap values and device IDs.
How to Confirm the Configuration Is Correct
- Upload a typical assembly process card PDF, check if the parsed result retains core fields such as process titles, torque values, and gap requirements.
- Call the file upload interface, upload an assembly document package under 500 MB, confirm that the upload and parsing processes do not throw timeout errors.
- Review the chunked text blocks, confirm that each block contains complete operation steps or inspection records, and no key numerical values are split.
- Cross-check the chunked data stored in the knowledge base, confirm that the field matching between document content and original files meets expectations.
The values provided here are common starting points and should be measured against the reader's own samples.