What the data for this category looks like
Engineering consulting supervision business documents mainly originate from on-site supervision logs, inspection lot acceptance records, project payment approval forms, engineering change consultation forms, and similar materials. Document update rhythm follows project construction milestones. Daily supervision logs are updated every day. Phased acceptance documents are generated synchronously with acceptance milestones. Document structures are primarily mixed formats, including fixed fields such as engineering location, inspection items, measured deviation values, structured tables, and long-text supervision opinions. Field units mostly use general engineering measurement units such as millimeters, square meters, and ten thousand yuan.
What constraints these characteristics impose on the document parsing and chunking link
Mixed-format document structures require the parsing link to retain the binding relationship between tables and text, and avoid splitting structured data and associated descriptions. The combination of fixed fields and measurement units requires chunking to avoid splitting complete information of the same business unit. For example, do not split measured data of an inspection lot and supervision conclusions into different chunks. The document update feature tied to project milestones requires the parsing process to support incremental synchronization and batch archiving processing, to prevent parsing task backlogs caused by fluctuations in document update frequency. In addition, scenarios with mixed handwritten signatures and printed text put additional requirements on OCR character compatibility.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
customPdfParse | Enabled | Engineering consulting supervision documents often contain nested structured tables. Custom PDF parsing retains original typesetting and table structures |
chunkSize | 1000–1500 characters | Adapts to the length of business units in individual supervision documents, avoids splitting complete inspection lot or approval form content |
chunkSplitter | Split by paragraph + table boundaries | Adapts to mixed text and table document structures, prevents splitting core business information across paragraphs or tables |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Large archived documents such as quarterly supervision monthly reports take longer to parse. Extend timeout to avoid parsing failures |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Supports batch upload of single-project supervision archive packages, adapts to business scenarios with integrated multiple documents |
chunkSettingMode | Manual configuration | Flexibly adjust chunking parameters for different document types such as acceptance records and change consultation forms |
Reference only; verify against the target FastGPT release and deployment documentation.
Three common misconfigurations
- Issue: After importing an Excel-format supervision approval form, the knowledge base displays the table but loses table formatting during retrieval output. Cause: The
customPdfParseconfiguration is not enabled. The default parsing logic does not retain the structured nested relationship of Excel tables. - Issue: Garbled Chinese characters appear when OCR recognizes handwritten signature areas in supervision documents. Cause: No Chinese font package adapted for engineering documents is configured, leading to incorrect character mapping during mixed recognition of handwritten and printed text.
- Issue: Supervision conclusions and corresponding measured data in chunked content are split into different chunks. Cause: The
chunkSplitteris configured to split by fixed character count, and does not bind the context boundaries between tables and associated text.
How to confirm the configuration is correct
- Upload a single supervision inspection lot record sheet. Check the parsed preview page to confirm the table structure matches the original document completely.
- Trigger OCR recognition of the handwritten supervision opinion area. Check that the recognition result has no garbled Chinese characters, and all signature-related characters are complete and not missing.
- Import an Excel-format project payment approval form. Run a retrieval test for business keywords, confirm that the output contains complete tables and associated text content.
- Adjust
chunkSizeto 1200 characters, upload the quarterly supervision monthly report, confirm that chunks do not split all acceptance information of the same inspection lot.
The values listed here are common starting points. Adjust them based on the samples in use.
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.