What this category of data looks like
Data comes from urban investment platform internal project declaration systems, documents returned by administrative approval departments, and internal archived case materials. Updates align with individual project progress timelines. Single documents are mixed structured and semi-structured files, typically tens to over one hundred pages long. They include fixed fields such as project approval number, full name of the urban investment entity, declared funding amount, approval document number, and land area. Common units include ten thousand yuan, square meters, and mu, standard engineering and financial measurement units.
Constraints for document parsing and chunking
Continuous distribution of fixed business fields means chunking cannot split critical information. Splitting critical information will disrupt subsequent retrieval matching. Semi-structured approval attachments are mostly scanned documents, requiring precise OCR parsing to restore original data. Single documents have multiple pages, so chunking must balance context integrity and retrieval granularity. Some documents include cross-page financial and engineering data. Chunking logic must preserve cross-page associated information to avoid splitting critical data.
Configuration settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Aligns with the length of business fields in single declaration materials and cross-page data association needs, avoids splitting critical business units |
chunkSplitter | Split by paragraph + heading | Matches the fixed template structure of urban investment declaration documents, preserves association between fields and context |
customPdfParse | Enable table OCR and structured extraction | Restores semi-structured data such as engineering budget tables and approval attachment tables within documents |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Addresses parsing time for single documents over one hundred pages, prevents timeout interruptions |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Adapts to upload size limits for single multi-attachment documents on urban investment platforms, prevents upload failures |
chunkSettingMode | Manual configuration | Adjusts chunking rules for fixed field distribution in urban investment documents, meets retrieval needs of business scenarios |
Three common configuration errors
- Symptom: Chinese garbled characters appear after OCR parsing, or table content is not fully extracted. Cause: Precise OCR configuration for Chinese scanned documents is not enabled, or engineering fonts and table border styles in urban investment documents are not supported.
- Symptom: After importing an Excel table dataset, retrieval cannot match the field correspondence within the table. Cause: Table structured extraction configuration is not enabled, or table data is split into independent small units during chunking, breaking field association logic.
- Symptom: Table content is missing or displayed abnormally in the output results after parsing a single declaration material. Cause: Chunk length is set too small, splitting complete cross-page tables into multiple fragments, making it impossible to fully read table data during retrieval.
How to verify correct configuration
- Upload a typical urban investment declaration PDF document, check the parsed text content to confirm that critical business fields are not split.
- Import an Excel format declaration table dataset, run a retrieval test to confirm that field correspondence within the table can be matched normally.
- Submit a long document over one hundred pages, wait for parsing to complete, confirm that no timeout error occurs.
- Initiate a retrieval for table content within the document, confirm that relevant data within the table can be returned completely.
The values provided here are common starting points, and should be evaluated 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.