This document is written based on the configuration rules of FastGPT v4.17.0.
What the data for this category looks like
Daily documents for this category include winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge and pressure withstand test reports, and nameplate parameter tables. Data sources include design BOM, silicon steel sheet and insulating material data, vacuum and drying equipment, oil testing, and test stations. Documents update once per equipment unit or material batch. Some structured documents include dedicated fields: winding turns for winding records, vacuum degree and temperature for vacuum drying records, for example. Units vary by document type, such as turns, pascals, degrees Celsius, and more. Nameplate parameter tables include standardized parameters such as rated voltage and rated capacity.
What constraints these characteristics impose on document parsing and chunking
Documents from this category have diverse sources and varied formats, including PDFs with test waveform diagrams, DOCX files with assembly photos, and structured Excel spreadsheets. The parsing pipeline must support mixed text and image content. The full document package for a single equipment unit has a large volume, so the system must support large file uploads and parsing. Core data includes key quality information such as insulation parameters and high-voltage test results. Chunking must retain the association between fields and values, and avoid splitting associated content of the same test item. The update rhythm of each unit and batch requires the parsing process to adapt to high-frequency, small-batch document processing needs. Parsing accuracy directly affects the effectiveness of subsequent quality audits.
How to set configurations
| Config Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapt to field association in long texts such as insulating oil reports and test reports, and avoid splitting key test data |
chunkSplitter | Chunk by paragraph + semantics | Adapt to structured documents with multiple fields, retain associated content of the same test item, and avoid data fragmentation |
customPdfParse | Enabled | Adapt to PDF documents with test waveform diagrams and nameplate photos, and prevent failed image parsing |
UPLOAD_FILE_MAX_SIZE | 500 MB | Accommodate the full assembly and test document package for a single equipment unit, and meet batch document upload requirements |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Adapt to the complex parsing process of large test reports, and avoid task interruption due to timeout |
chunkSettingMode | Manual configuration | Adjust chunking rules for different document types, and adapt to the exclusive formats of insulation parameters and test data |
Three common mistakes
- Invalid image identifiers return after uploading DOCX or PDF files with images. Cause:
customPdfParseis not enabled, and the mixed text-image parsing pipeline is not activated. Image content cannot be correctly recognized as a result. - Timeout errors trigger when parsing large test reports. Cause:
PARSE_FILE_TIMEOUT_SECONDSis not adjusted to the applicable range. The default timeout duration is insufficient to complete complex document parsing. - Images and data in Excel spreadsheets split into different chunks. Cause:
chunkSplitteris not specified as the chunking rule for structured tables. The default chunking logic accidentally separates images and spreadsheet content.
How to confirm correct configuration
- Upload a partial discharge report PDF with test waveform diagrams, and check if the parsing result includes complete image-related descriptions and test values.
- Upload the full document package for a single equipment unit, and confirm no size limit prompts appear during upload, and the parsing task completes normally.
- View the parsing log, confirm no timeout errors occur, and the configured value of
PARSE_FILE_TIMEOUT_SECONDSdoes not trigger task interruption. - Randomly select an insulating oil report, and check that insulation parameters for the same batch are not split into different content chunks in the chunking result.
The values provided are common starting points and should be measured against the reader's own samples.