This page’s configuration recommendations apply to FastGPT v4.17.0.
What data for this category looks like
Typical documents for this category include production process specifications, batch production records, clean zone records, sterilization batch records, inspection reports, software version records, and adverse event and traceability materials. Data sources cover design history files, device and material lots, MES systems, clean environment monitoring equipment, sterilization equipment, and inspection and calibration systems. Update frequency is once per batch or per product, with regular updates per fixed environmental monitoring plans. Documents contain fields such as device unique identifiers, batch numbers, clean zone temperature and humidity, sterilization parameters, and test results. Some documents are in scanned image format.
What constraints these characteristics impose on document parsing and chunking
Mixed structured and unstructured data from multiple sources requires retaining field associations during chunking, to avoid splitting core related information such as device unique identifiers and batch numbers. The presence of scanned documents requires the parsing step to support OCR processing. High-frequency batch updates require chunking configurations to flexibly adapt to different document sizes. Parameter fields with units require chunking to retain complete context, ensuring that complete parameter combinations can be matched during retrieval.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Matches the complete content of a single process step or single inspection record, avoids splitting core associated fields |
chunkSplitter | Separate by paragraph + heading | Retains the hierarchical structure of production process specifications, matches the natural reading logic of documents |
customPdfParse | Enable OCR mode | Adapts to scanned image format documents such as sterilization batch records and inspection reports |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Covers the parsing time required for large batch sterilization records and complete design history files |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Supports batch upload requirements for full batch production records |
chunkSettingMode | Manual configuration | Adjust chunking parameters for different document types such as process specifications and batch records |
Three common mistakes
- Knowledge base search fails to return target records: The symptom is no matching results when searching for a device unique identifier or batch number. The cause is that the complete context containing the identifier was split during chunking, leading to loss of associated information.
- Timeout error occurs when parsing large sterilization batch records: The symptom is a timeout status displayed for the task. The cause is that the
PARSE_FILE_TIMEOUT_SECONDSparameter was not adjusted, and the default threshold is insufficient to cover the parsing time of batch documents. - Only a small amount of garbled text is returned after parsing a scanned inspection report: The symptom is a large amount of garbled text or missing core content in the parsed text. The cause is that OCR parsing mode was not enabled for
customPdfParse, and text extraction was performed directly on image format content.
How to confirm the configuration is properly set
- Upload a typical sterilization batch record document, check if the parsed text retains the complete context of core fields such as sterilization parameters and batch numbers.
- Run a retrieval test for a device unique identifier, confirm that chunked content containing the identifier can be returned.
- Upload a scanned inspection report, confirm that the parsed text has no obvious garbled text and covers the core content of the report.
- Upload a large process specification document under 1000 MB, confirm that the parsing task completes normally without timeout.
The values listed here are common starting points, and should be evaluated against your own samples.