This configuration guide is based on FastGPT v4.17.0.
What the data for this category looks like
Daily documents for this category include assembly rules, BOM lists, welding torque records, BMS flashing records, insulation and voltage withstand reports, EOL test reports, system configuration lists, and more. These cover both structured table and unstructured text formats. Data sources include cell grading systems, MES production systems, welding equipment, torque tools, BMS software, and EOL test benches. Fields include cell capacity, torque value, test voltage, BMS version number, with corresponding units of Ah, N·m, V, and others. The update frequency is once per produced power battery pack or energy storage cabinet, or triggered each time the BMS software version changes. The number of documents updated in a single batch fluctuates with production tempo.
What constraints do these characteristics impose on the "document parsing and chunking" link
The mixed format of documents in this category requires parsing tools to support both paragraph and table recognition, and avoid splitting cross-row table content. The specificity of fields and units requires that when chunking, the context of associated parameters is retained, and cell parameters must not be split into separate chunks from corresponding BMS configurations. The frequently updated production ledgers require reasonable timeout settings for the parsing process, to avoid parsing interruptions caused by a single document containing multiple test attachments. The business logic where a single device corresponds to a dedicated ledger requires retaining the full context of a single device during chunking, to prevent cross-device information confusion that affects subsequent summary and report generation.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Matches the length of a single test record or BOM entry in documents of this category, avoiding splitting continuous content that contains associated fields |
chunkSplitter | Paragraph + table recognition | Adapts to the mixed feature of pure text test reports and structured tables in documents of this category, retaining complete table rows and paragraph logic |
customPdfParse | Enable pdf-marker v2 mode | Adapts to the complex typesetting of insulation and voltage withstand reports and EOL test reports, accurately extracting tables and annotated content |
PARSE_FILE_TIMEOUT_SECONDS | 120–180 seconds | A single ledger for this category may contain multiple test attachments, extending the timeout to avoid parsing interruptions |
chunkSettingMode | Manual configuration | Adjust chunking rules for different document types such as BOM lists and flashing records to adapt to differences between structured and unstructured documents |
Three common mistakes
- Symptom: The parsing API returns a
413 Request Entity Too Largestatus code, or parsing progress gets stuck at 90% before timing out. Cause: TheUPLOAD_FILE_MAX_SIZEandPARSE_FILE_TIMEOUT_SECONDSparameters were not adjusted. A single energy storage cabinet ledger may contain multiple EOL test PDFs, exceeding default limits. - Symptom: After splitting, cross-row material codes and specification fields in CSV-format BOM lists are assigned to different chunks, making it impossible to match complete information during retrieval. Cause: Table recognition mode for
chunkSplitterwas not enabled, and table rows were split by default paragraph chunking. - Symptom: In parsed chunks, the version number of BMS flashing records is separated from corresponding cell parameters, making it impossible to associate complete configurations during retrieval. Cause: The
chunkSizesetting was too small, splitting continuous paragraphs that contain associated fields.
How to confirm the configuration is correct
- Upload a single typical document such as the insulation and voltage withstand report and BOM list for a single energy storage cabinet, check the parsed chunk list, and confirm that table content is not split and paragraph chunks retain field associations.
- Call the batch parsing API, pass in 3 ledger documents of different types, and confirm that all tasks are completed within the time set by
PARSE_FILE_TIMEOUT_SECONDSwith no timeout errors. - Retrieve the chunked knowledge base, enter a query containing cell capacity and BMS version number, and confirm that the returned chunks contain both associated fields without missing information.
- Test multi-file upload parameters, confirm that the
contentarray inmessagescan correctly carry parsing result identifiers for multiple files, with no parameter format errors.
The values listed above are common starting points and should be evaluated against the reader's own samples.