What data looks like for this category
Ledger data for rail transit equipment mainly comes from production assembly BOM documents, daily inspection records, monthly maintenance reports, annual operation ledgers, and similar materials. Data update cycles cover daily, monthly, and annual frequencies: daily inspection records are added daily, monthly maintenance reports are archived monthly, and annual operation ledgers are aggregated per fiscal year. Documents contain both structured fields and unstructured text. Structured fields include equipment ID, model, total running kilometers, maintenance man-hours, component serial numbers, with units of km, h, and units respectively. Unstructured content includes fault phenomenon descriptions, on-site handling records, supplier delivery notes, and similar content. Some documents are scanned PDFs with handwritten annotations and tables.
Constraints on document parsing and chunking
The mixed document structure of rail transit equipment ledgers requires the parsing process to retain both structured fields and unstructured context, to avoid splitting that breaks the complete association of single maintenance records. High-frequency updated daily inspection documents and ultra-long annual ledgers set clear requirements for upload file size limits and parsing timeout settings. Scanned PDF documents require dedicated parsing tools to extract table and annotation information, otherwise key structured data will be lost. Associated reference fields in ledgers, such as linked work order numbers and supplier contract numbers, must retain their association relationships, rather than only extracting plain text content.
Configuration settings
| Configuration item | Recommended value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Matches the field length of single ledger records for rail transit equipment, avoids splitting that breaks the complete association of maintenance processes and component lists |
chunkSplitter | Split by paragraph + semantic meaning | Adapts to mixed structured tables and unstructured fault analysis text in ledgers, retains context coherence |
customPdfParse | Enable pdf-marker parsing | Supports extraction of tables and handwritten annotations from scanned PDFs, resolves the issue of structured data loss with default parsing |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Adapts to the page count of single annual ledger documents, avoids parsing timeout for large documents |
UPLOAD_FILE_MAX_SIZE | 500 MB | Supports batch upload of large assembly ledgers and annual operation ledgers |
chunkSettingMode | Manual configuration | Adjusts chunking parameters for the special structure of rail transit equipment documents, adapts to mixed content scenarios |
Three common misconfigurations
- Issue: No custom parsing option appears when uploading a PDF to the FastGPT knowledge base. Cause: The locally deployed pdf-marker has not been configured with cross-container network connectivity, and has not established port communication with the FastGPT service.
- Issue: The references field is empty after parsing. Cause: The
customPdfParsefull parsing mode is not enabled. The default parsing only extracts plain text content, and does not retain associated reference fields. - Issue: Chunked content is overly fragmented, and continuous maintenance step descriptions are lost. Cause: The
chunkSizesetting is too small, and does not match the field length of single ledger records, resulting in splitting of complete fault troubleshooting paragraphs.
How to verify correct configuration
- Upload a standard rail transit maintenance PDF document, and check in the FastGPT v4.17.0 knowledge base whether structured fields such as equipment ID and total running kilometers in tables are retained after parsing.
- View parsing task logs, confirm that large documents do not show timeout errors after the
PARSE_FILE_TIMEOUT_SECONDSsetting is applied. - Randomly sample chunked content after parsing, check that complete fault troubleshooting paragraphs are included, with no incomplete sentences split mid-passage.
- Call the FastGPT document parsing API, test uploading a small inspection document, confirm that the returned number of chunks matches the document content, and there are no abnormal empty chunks.
The values listed here are common starting points, and should be validated 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.