What this category of data looks like
Core data for this category comes from OA systems, organizational structures, policy documents, and user feedback. Updates occur with workflow changes or once per month. Daily documents include workflow configuration sheets, approval node tables, form field descriptions, and change records. Most documents are structured tables, linear step sequences, containing fields such as approval roles, permission scopes, and node timeout thresholds. Some documents include associated approval workflow screenshots and form examples.
What constraints these characteristics impose on the document parsing and chunking link
Structured tables account for a large share of documents in this category. Chunking must avoid splitting table cells or breaking row-column associations. Documents are updated frequently with workflow changes, so chunking configurations must adapt to rapidly evolving document structures. Some documents contain permission-bound fields, and the association between fields and permissions must be retained after chunking. Some documents include time-sequential approval nodes, and chunking must not interrupt the continuous logic of the nodes. Documents from multiple sources have significant format differences. Parsing logic for different export formats must be supported to avoid content loss due to format incompatibility.
How to set configurations
These configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapt to the structured content length of workflow configuration sheets and approval node tables, avoid splitting critical approval logic |
chunkSplitter | Prioritize chunking by tables | Adapt to the feature that most documents in this category have structured tables, retain table integrity and associated information |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Adapt to the parsing duration of bulk-imported workflow documents, avoid timeout interruptions due to large document volumes |
customPdfParse | Enable MinerU adaptation mode | Adapt to PDF format workflow document parsing, support complex tables and layout details |
UPLOAD_FILE_MAX_SIZE | 100 MB | Adapt to the size of bulk-imported change records and policy documents, avoid upload failures |
chunkSettingMode | Manual configuration | Adapt to frequently updated document structures, flexibly adjust chunking rules to match the latest workflow documents |
Three common mistakes
- Failed table chunking: After uploading to the knowledge base, table content is split or lost. The cause is that the table-first chunking strategy is not enabled, and the default character-based chunking is used, which destroys table structure.
- GPU recognition failure and memory overflow: After deployment, memory overflow occurs or GPU information cannot be recognized. The cause is that the GPU adaptation parameters for
customPdfParseare not configured correctly, or the CUDA version does not match the deployment environment. - Missing image domain name after Word import: When importing Word documents, images automatically carry a domain name, but the image links lose the domain name when converted to Markdown for import. The cause is that the image link processing rules during document parsing are not configured, resulting in the exported Markdown links losing source site information.
How to confirm the configuration is complete
- Upload a workflow configuration sheet containing an approval node table, check if the parsed chunking results retain the complete table structure, with no splitting or content loss.
- View the parsing log, confirm that the
PARSE_FILE_TIMEOUT_SECONDSconfiguration takes effect, with no timeout error records. - Import change records in Word format, check if image links carry valid domain names and can be properly associated with the source file.
- Test bulk upload of multiple workflow documents, confirm that the
UPLOAD_FILE_MAX_SIZEconfiguration does not trigger upload restrictions.
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.