What the data for this category looks like
Data for this category comes primarily from land use planning approvals, construction permit documents, project feasibility study reports, mortgage assessment materials from financial institutions, and site survey records. Updates follow key project milestones. New documents are generated at milestones such as land acquisition, construction start, and pre-sale. Routine maintenance updates are completed quarterly. Most documents are 30–100 page PDFs or scanned files. Internal structure includes modules like basic project information, land parameters, planning indicators, funding details, construction progress, and risk reminders. Most fields have clear units, such as total land area (square meters), floor area ratio building area (square meters), and construction and installation cost (yuan per square meter). Many embedded tables and planning diagrams are included.
What constraints do these characteristics impose on the document parsing and chunking workflow?
The data characteristics of this category create multiple constraints for parsing and chunking. A high share of scanned documents and low-resolution files requires the parsing workflow to support general OCR capabilities. This ensures accurate recognition of both handwritten and printed text. Fields with clear units and embedded table structures require retaining the binding relationship between tables and associated text during chunking. This prevents separating units from their linked indicators. Wide variation in document length and inconsistent formatting requires splitting by project modules instead of fixed character counts. This avoids splitting complete information blocks such as land parameters and funding details. Embedded planning diagrams must be linked to their corresponding indicator paragraphs. This prevents disconnecting visual information from text chunks.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
ocrEnable | Enabled | This category of documents contains large numbers of scanned files and planning diagrams, requiring OCR to extract non-editable text content |
parseTableMode | Fully retain table structure | Core indicators such as land area and construction and installation cost are presented in tables. Retaining full table structure avoids information misalignment |
chunkSize | 800–1200 characters | Single modules in this category have high information density. This range avoids splitting the same indicator group, while adapting to knowledge base context window limits |
chunkOverlapRate | 10%–15% | Indicator association information across segments must be retained. This prevents a single core indicator from being split into two paragraphs after chunking |
parseTimeout | 600 seconds | Individual due diligence reports have many pages. Sufficient time must be reserved for OCR recognition and chunking processing |
maxUploadFileSize | 500 MB | Documents in this category often embed multiple planning diagrams. Larger file uploads must be allowed |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing settings.
Three common errors
- Phenomenon: After uploading a PDF containing planning diagrams, no image-associated text appears in parsing results. Cause: The
ocrEnableconfiguration is not enabled, so text content in images embedded in the PDF cannot be extracted. - Phenomenon: In version 4.9.10, parsed chunk hierarchy is disorganized, and core modules such as land parameters and funding details cannot be distinguished. Cause: The
maxParagraphDepthparameter is not configured correctly, so content cannot be split according to document heading hierarchy. - Phenomenon: The table dataset option cannot be selected when creating a knowledge base page. Cause: The table parsing function switch is not enabled, or the currently deployed FastGPT version does not support the table parsing module.
How to confirm configurations are properly set
- Upload a test document containing scanned files and tables. Check if the text in parsing results covers all content from the scanned pages, to confirm the OCR configuration is active.
- Randomly select a complete due diligence report. Check if chunking results are split according to document heading hierarchy, to confirm the paragraph depth configuration meets document structure requirements.
- Review parsed chunk content. Confirm that the structure and cell content of core tables are fully retained, to confirm the table parsing configuration is correct.
- Upload a test document larger than the conventional size. Confirm that upload and parsing processes do not trigger timeout errors, to confirm timeout and upload size configurations are reasonable.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.