This page is written for FastGPT v4.17.0.
What data looks like for this category
The data for this category comes from content review backends, user reporting channels, public regulatory and policy documents, and internal sample libraries. Daily produced documents include review rule manuals, violation handling guides, appeal review records, and quality inspection review reports. Most documents contain hierarchical rule entries, sample annotation fields, handling process descriptions, and update logs. Data update frequency aligns with policy changes or occurs via daily synchronization. Individual document sizes range from hundreds to tens of thousands of characters. Most fields include standardized content such as rule numbers, effective dates, applicable scenarios, and judgment criteria.
What constraints these characteristics impose on the document parsing and chunking link
Because review rule documents have hierarchical entries and strongly linked judgment criteria, context association must be preserved during chunking. Logical units spanning multiple rules must not be split. Data updates occur frequently, so incremental parsing and chunk update support is required. This avoids full reprocessing of documents. A large number of standardized fields and sample annotations are included in documents. Parsing processes must identify structured content. This prevents splitting fields and explanatory text into different chunks. Some documents originate from scanned copies or image-format policy files. Support for non-text content parsing capability is needed. Individual document sizes vary widely. The system must adapt to input and chunking logic for different lengths.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the length of hierarchical entries in review rule documents, preserves complete logic for individual rules |
chunkSplitter | Split by headings and paragraphs | Matches the hierarchical structure of review rule documents, avoids splitting across multiple rules |
customPdfParse | Enable OCR mode | Supports parsing policy documents and sample screenshots that include images or scanned copies |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Adapts to parsing duration for large-volume quality inspection reports and historical appeal records |
chunkSettingMode | Manual configuration | Allows adjustment of chunking parameters for different types of review documents, adapts to diverse document structures |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Supports uploading batch documents of large-volume historical appeal records |
Three common mistakes
- Scenario: After uploading an image or PDF file containing images, the parsing result has no text content. Cause: OCR mode for
customPdfParseis not enabled. Default parsing only extracts text layers from PDFs, and cannot extract rule content embedded in images. - Scenario: Table dataset cannot be selected during knowledge base construction. Cause: The structured parsing switch is not enabled, or the uploaded table document format does not meet platform recognition standards. This prevents extraction of table fields.
- Scenario: No paragraph depth related settings are found when configuring chunking parameters. Cause: The chunking configuration logic in the current version differs from older versions. Switch to advanced configuration mode via
chunkSettingModeto view these settings.
How to confirm configurations are correct
- Upload a review rule PDF containing images. Check if the parsing result extracts text content from the images.
- Upload a structured table document. Confirm that the parsing result includes complete fields and content from the table.
- Switch to advanced configuration mode for
chunkSettingMode. Confirm that paragraph depth related setting options are visible. - Upload a quality inspection report over 10,000 characters. Wait for parsing to complete, then check for timeout errors.
The provided values are common starting points, and should be evaluated against the samples used in actual deployment.
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.