Reference only; verify against the target FastGPT release and deployment documentation.
What this category of data looks like
Data sources for this category include construction quality and safety supervision systems, on-site real-name registration platforms, supervision logs, and test reports. The update frequency is weekly. Commonly processed documents cover supervision plans, supervision inspection records, hidden danger rectification notices, quality problem handling forms, and completion acceptance supervision reports. Some documents use structured table formats, some are unstructured long text, and some include image resources such as construction site photos and test report attachments. Fields include related information such as hidden danger markers, responsible entities, and project locations.
What constraints these characteristics impose on the document parsing and chunking link
Documents from multiple sources have large format differences, including standardized PDFs exported from systems, on-site scanned PDFs, plain-text supervision logs, and more. The parsing module must support mixed format adaptation. In weekly batch processing scenarios, chunking parameters must balance semantic completeness and retrieval efficiency. Avoid making single chunks overly scattered or too long. Some documents include structured tables and image resources. The chunking logic must retain table hierarchical structures and associate images with their corresponding text. Parsing errors may trigger quality and safety accidents. Strict timeout and failure handling mechanisms must be configured to ensure parsing stability.
How to set the configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the semantic completeness of long text such as supervision logs and inspection records, avoiding fragmentation of single-chunk information |
chunkSplitter | Layer by paragraph + title | Retains the structured title hierarchy of supervision plans and rectification notices, maintaining the logical structure of documents |
customPdfParse | Enable mixed mode | Supports mixed parsing of system-exported PDFs, scanned PDFs, and test reports with attachments |
UPLOAD_FILE_MAX_SIZE | 500 MB | Adapts to large-volume documents such as batch completion acceptance supervision reports |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Accommodates parsing time for complex scanned documents or documents with multiple attachments |
chunkSettingMode | Manual configuration | Adjusts chunking parameters separately for different document types such as structured tables and long text |
Three common mistakes
- Phenomenon: No parsing in progress, ready, or failed status fields are returned after calling the API to create a knowledge base. Cause: Relevant configurations for knowledge base parsing status synchronization are not enabled.
- Phenomenon: The Set-Cookie field in HTTP request responses cannot be extracted by the parsing component. Cause: Custom parsing rules adapted to the extraction logic of this field are not configured.
- Phenomenon: After uploading a PDF document with construction site photos, the parsing result only contains text with no associated image content. Cause: The image embedding parsing mode of
customPdfParseis not enabled.
How to confirm the configurations are correct
- Upload a single test report that includes structured tables and images. Check if the parsing result retains the table structure and associates images with their corresponding text.
- Call the knowledge base creation API. Check if parsing status related fields are included in the returned result.
- Upload a single completion acceptance supervision report under 100 MB. Confirm that the parsing time does not exceed the set timeout threshold.
- Attempt to parse an HTTP response message containing the Set-Cookie field. Confirm that the field is correctly extracted and included in the chunked content.
The values provided are common starting points and should be measured against the reader's own 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.