What this category of data looks like
Shipbuilding declaration materials and file review data comes primarily from shipyard compliance declaration documents, classification society inspection files, construction progress archives, and design drawing attachments. Data is updated in batches when new ship projects launch, and supplemented quarterly during routine compliance checks. Document formats include fixed-format Word declaration forms, PDFs with vector drawings, structured Excel ledgers, handwritten scanned copies, and more. Fields include ship type number, IMO identification code, construction materials, tonnage parameters, with units such as meters, tons, and man-hours.
What constraints these characteristics impose on document parsing and chunking
Vector drawings in shipbuilding documents require parsing tools to support image extraction and text association, to avoid losing design details. Structured Excel ledgers hold multi-row, multi-column parameter data. Chunking must not break the table’s row and column structure, or complete parameter association will fail during retrieval. A single full-ship construction archive can reach hundreds of pages. Chunking must retain context association to prevent truncation of technical logic. Batch declaration file processing requires parsing tools to support high-concurrency scenarios, as well as large file uploads.
How to configure settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Shipbuilding documents contain many technical parameters and long paragraphs. This range avoids splitting that disrupts technical logic |
chunkSplitter | Split by paragraph + heading | Shipbuilding documents have clear chapter hierarchies. This splitting rule retains context association |
customPdfParse | Enabled | Shipbuilding PDFs include classification society stamped drawings and embedded vector text. Default parsing cannot extract content completely |
UPLOAD_FILE_MAX_SIZE | 2000 MB | A single full-ship construction archive may exceed 1 GB. Large file upload support is required |
PARSE_FILE_TIMEOUT_SECONDS | 1200 seconds | Long document parsing takes significant time. This setting avoids task failure from mid-process timeout |
chunkSettingMode | Manual configuration | Shipbuilding documents have wide structural differences. Chunking rules need adjustment for different document types |
Reference only; verify against the target FastGPT release and deployment documentation.
Three common mistakes
- Symptom: After uploading a ship design PDF, embedded hull structure drawings do not appear in the interface. Cause: The
customPdfParseconfiguration is not enabled. The default parsing engine only extracts plain text and does not handle embedded vector images. - Symptom: After uploading a ship construction ledger Excel file, parsed content has randomly truncated table data. Cause: The
chunkSplitteris not configured to split by table boundaries, leading to disrupted structured data. - Symptom: All parsing tasks time out when uploading 10 or more declaration files in batch. Cause: The
PARSE_FILE_TIMEOUT_SECONDSsetting is too low, and does not accommodate the time required for long document parsing.
How to confirm configurations are correct
- Upload a test PDF containing hull drawings, and check if the parsed result includes drawing text annotations and image previews.
- Upload a ship construction ledger Excel file, and verify that parsed chunks retain the table’s row and column structure and full field integrity.
- Upload a test archive with a single size of 1500 MB, and confirm the upload and parsing process does not trigger a file size limit error.
- Check the actual active value of the
chunkSizeconfiguration item, and verify that post-chunk text length falls within the preset range.
The values listed here are common starting points. Always test against your own specific samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.