This document was written based on FastGPT v4.17.0
What data in this category looks like
Daily documents for this category include annotation specifications, bounding box examples, quality inspection reports, and rework lists. Data sources include image samples, files exported from annotation platforms, category dictionaries, and random inspection results. Typical users are annotation teams, computer vision algorithm teams, and quality inspectors. Data is updated daily or per annotation batch. Document structures include plain text rules, formatted tabular reports, and annotation examples combining text and images. Fields include target categories, bounding box coordinates, random inspection pass rates, rework items, and more. There is no unified fixed format, and it is adjusted according to project requirements.
What constraints do these characteristics impose on the document parsing and chunking link
Daily documents have diverse table formats, including statistical data in quality inspection reports and task details in rework lists. When parsing, the row and column correspondence of tables must be retained to avoid data misalignment after splitting. The metadata of image samples is bound to annotation rules. When chunking, fields related to visual targets must be retained to prevent loss of key annotation information. Frequently updated annotation data requires the parsing process to be efficient, to avoid task interruptions caused by timeouts. At the same time, document structures vary greatly across different projects, so flexible chunking configuration must be adapted, and general fixed splitting rules cannot be used.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Annotation documents include long texts such as bounding box rules and category lists. This range balances the information density of a single chunk and retrieval accuracy |
chunkSplitter | Split by paragraph + table boundary | Daily documents contain a large number of tabular quality inspection reports and rework lists. Splitting by table boundaries can retain data correspondence |
customPdfParse | Enabled | Annotation specifications are mostly formatted PDFs. Enabling this option preserves font hierarchy, table structure, and special annotation fields |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Packed files of single-batch annotation image samples are usually large, and this setting adapts to batch upload requirements |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Parsing large-scale annotation datasets requires long processing time to avoid task interruptions due to timeouts |
chunkSettingMode | Manual mode | Document structures vary greatly across annotation projects. Manual configuration can accurately match the splitting needs of different document types |
Three common mistakes
- Symptom: After uploading an Excel-format tabular dataset, complete table content cannot be returned during retrieval. Cause:
chunkSplitteris not configured to split by table boundaries, causing table data to be split into scattered text chunks and losing row and column correspondence. - Symptom: Bounding box coordinate information is missing when retrieving and outputting annotation example documents displayed in the knowledge base. Cause:
customPdfParseis not enabled, and format fields and special annotation content in the document are lost during parsing. - Symptom: Parsing tasks throw timeout errors after uploading batch annotation files. Cause:
PARSE_FILE_TIMEOUT_SECONDSis not adjusted to an appropriate duration, causing large file parsing to be interrupted before completion.
How to confirm the configuration is correct
- Upload a single annotation specification PDF, check whether the parsed text chunks retain complete paragraphs of category definitions and bounding box rules, with no content truncation or misalignment.
- Upload an Excel-format quality inspection report, perform a retrieval test, and verify whether the returned results include complete tabular row and column data, with no scattered cell content.
- Adjust
PARSE_FILE_TIMEOUT_SECONDSto 300 seconds, upload a single batch annotation file, and confirm that the parsing task has no timeout errors. - After switching the vector model, import the original chunked data, perform a retrieval test, and confirm that the matching results meet the retrieval needs of the annotation project.
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.