What This Category of Data Looks Like
Daily documents for image annotation projects include four core types: annotation specifications, bounding box examples, quality inspection reports, and rework lists. Data sources include image samples exported from annotation platforms, category dictionaries, spot check results, and manually recorded rework logs. Update frequency follows daily cycles or individual project cycles. Document structures typically mix text descriptions and embedded visual examples. Some quality inspection reports contain structured fields such as bounding box coordinates and annotation categories, which differ from purely text documents used in text annotation projects. Content focuses on visual target position determination and category annotation.
What Constraints Do These Characteristics Impose on Document Parsing and Chunking
Because documents mix text and visual examples, parsing and chunking must retain the association between text descriptions and corresponding bounding box examples within the same chapter. This prevents splitting that fails to meet the context requirements of compliance reviews. Structured data from multiple sources requires parsing tools to support exported files in different formats, and adapt to the high-frequency update rhythm of individual projects. This avoids losing fields from rework lists or spot check results due to rigid chunking logic. At the same time, compliance reviews require parsing results to fully retain the hierarchical relationship of annotation categories, and prevent chunking from destroying the complete structure of category dictionaries.
How to Configure
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Image annotation documents include text descriptions and embedded bounding box examples. This value preserves complete context for each chunk, and avoids splitting examples from their corresponding descriptions |
customPdfParse | Enabled | Adapts to embedded bounding box example screenshots in annotation specifications, prevents loss of visual association information during parsing, and meets the complete content requirements of compliance reviews |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Single batches of rework lists or quality inspection reports have large data volumes. This setting reserves sufficient time for structured field extraction and chunk processing |
chunkSplitter | Split by paragraph + visual block | Matches the chapter structure of annotation documents and the layout of embedded examples, ensuring that descriptive content for the same visual example is not split across chunks |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Supports batch upload of complete dataset files for single-batch spot check results, adapting to project-level batch parsing needs |
chunkSettingMode | Manual configuration | Adjust chunking parameters for three different document types: annotation specifications, quality inspection reports, and rework lists, to accurately match content verification requirements for compliance reviews |
Three Common Misconfigurations
- Connection failure error when calling marker-pdf in Docker deployment environments | Local service port mapping for
customPdfParseis not correctly configured, preventing FastGPT from connecting to the local parsing container - Bounding box example associated descriptions are missing from parsed results after document upload | The
chunkSizeparameter is set too small, splitting embedded visual examples and their text descriptions into separate chunks, losing the contextual association required for compliance reviews - Parsing tasks terminate early due to timeout | The
PARSE_FILE_TIMEOUT_SECONDSvalue is set too low, failing to accommodate the full parsing duration of large single-batch quality inspection reports
How to Confirm Successful Configuration
- Upload an annotation specification document, and check if the parsed results retain the association between chapter titles and corresponding bounding box examples, with no chaotic content splitting
- Review parsing logs in the FastGPT v4.17.0 backend, and confirm no timeout errors related to
PARSE_FILE_TIMEOUT_SECONDSoccur - Test exporting parsed results to Excel format, and confirm all annotation fields and bounding box coordinate information are fully extracted and retain their original format
- Verify that custom JS parsing methods can be called normally, and perform format verification on parsed bounding box data to ensure compliance with field requirements of the annotation project
The values listed above are common starting points. They should be evaluated against the user’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 2026-10-19.