What data in this category looks like
Daily documents for image annotation projects include four categories: annotation specifications, bounding box examples, quality inspection reports, and rework lists. Data sources include image samples, annotation platforms, category dictionaries, and spot check results. Document update frequency is daily or per project batch. Annotation specifications contain fields such as category names and bounding box determination rules. Quality inspection reports include structured content such as batch numbers, problem items, and rectification requirements. Rework lists are associated with problem details and processing progress of corresponding annotation tasks.
Constraints imposed on document parsing and chunking
Since documents contain text content bound to visual rules, parsing and chunking must retain the association between categories and bounding box rules. This prevents inability to match corresponding examples after splitting. Frequently updated documents require parsing processes to adapt to batch processing. This avoids task blocking caused by overly long single-file parsing times. Structured quality inspection reports and rework lists contain multi-dimensional fields. Chunking logic must balance field integrity and retrieval relevance. This ensures associated information can be fully extracted during subsequent matching. Unlike text annotation projects, documents for image annotation projects must be linked to visual position information. Chunking must avoid destroying this association.
Configuration settings
This document is written based on the configuration logic of FastGPT v4.17.0. The following is the configuration plan adapted for image annotation projects:
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapt to the text length of annotation specifications and quality inspection reports, avoid splitting key rules and corresponding example descriptions |
chunkSplitter | Split by paragraph + bounding box keywords | Match the content structure of "category-bounding box" binding in image annotation documents |
customPdfParse | Enabled | Support parsing PDF documents with embedded annotation example images, retain visual association information within the document |
PARSE_FILE_TIMEOUT_SECONDS | 90–120 seconds | Adapt to the parsing time of batch quality inspection reports, avoid timeout errors |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Adapt to the upload requirements of single large-scale annotation dataset documents |
chunkSettingMode | Manual configuration | Adjust chunking logic for non-standard structure of annotation documents |
Three common configuration errors
- The symptom is a platform timeout error. This occurs when the
PARSE_FILE_TIMEOUT_SECONDSconfiguration is not adjusted, and the default short duration is used. This cannot adapt to the parsing time of batch quality inspection reports. - The symptom is that the number of returned results when retrieving annotation data is lower than expected. For example, complete content is not matched when retrieving annotation data for a foreign usage location. The cause is that associated fields were split during chunking, resulting in inability to fully extract corresponding information during retrieval.
- The symptom is that bounding box example descriptions in the document are not extracted after uploading a PDF. The cause is that the
customPdfParseconfiguration is not enabled. The default parsing only extracts plain text content, ignoring embedded visual association content.
How to verify correct configuration
- Upload a single annotation specification PDF, check parsed chunked content, confirm category rules and corresponding example descriptions are not split.
- Submit a batch quality inspection report task, check parsing logs, confirm no timeout error is triggered.
- When retrieving annotation data, verify returned results include expected field content, such as usage location, bounding box parameters, etc.
- Adjust the value range of
chunkSize, compare chunking integrity under different configurations, confirm alignment with the structural requirements of the current document.
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 2026-10-19.