What this category's data looks like
Business data for photovoltaic module encapsulation comes from battery cell grading systems, MES production management systems, string welding and lamination equipment, EL/IV testers, and similar tools. Daily documents include layout drawings, material batch records, lamination process records, EL defect images, IV test reports, insulation voltage withstand records, and packing lists. Data is updated independently per module and per material batch. Fields include battery cell current grade, lamination temperature, adhesive film batch number, EL crack location, insulation voltage withstand value, and other related fields, with clear engineering units attached.
What constraints these characteristics impose on document parsing and chunking
High-frequency independent data updates for this category require chunking to strictly match the document boundaries of a single batch or single module. This prevents cross-batch data mixing that would reduce the accuracy of subsequent Q&A. Non-text documents such as EL defect images and IV test reports require associative parsing of charts and annotated text. Without this support, key process information such as crack locations and power values cannot be extracted. Structured multi-field documents such as lamination process records include tables and long-text steps. Chunking must retain field relevance to avoid splitting that breaks process logic. Large batch-uploaded batch record files require appropriate parsing timeout and file size limits.
Configuration settings
Configurations under FastGPT v4.17.0 are as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Photovoltaic encapsulation process documents include long step descriptions and structured tables. This range preserves single-segment process logic and complete field groups |
chunkSplitter | Split by paragraph + table | Adapts to the paragraph descriptions and tabular parameter records of lamination records, avoiding splitting that breaks data relevance |
customPdfParse | Enabled | Supports parsing of chart annotations and embedded text in EL defect images and IV test reports, to extract key process parameters |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Parsing large EL defect images and batch batch records takes longer. This duration covers most scenarios |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Adapts to bulk exported multi-batch material record package files, preventing upload interception |
chunkSettingMode | Manual mode | Adjusts chunking rules separately for different document types (process records, test reports) to adapt to category-specific document structures |
Three common mistakes
- After uploading a PDF containing EL defect images, the parsing result only returns text content, with no image annotations or defect location information. The cause is that the image parsing configuration item
customPdfParsewas not enabled, so embedded charts are not recognized. - After batch importing Excel-format batch records, knowledge base field matching fails. The cause is that worksheets were not split per batch, and cross-batch data mixing causes the chunking logic to fail to match process boundaries.
- A timeout error triggers when parsing a single large-size IV test report. The cause is that the
PARSE_FILE_TIMEOUT_SECONDSsetting is lower than the actual parsing time required for this type of document.
How to confirm configurations are set correctly
- Upload a single lamination process record PDF, and check that the parsing result retains tabular lamination temperature and time parameters, as well as coherent process step chunks.
- Upload a PDF containing EL defect images, and confirm that the parsing result includes the image's annotated text and defect location descriptions.
- Upload a 100MB batch batch record compressed package, and check that the parsing progress has no timeout errors and the file is fully parsed.
- Adjust the
chunkSizeparameter to the preset range, and verify that the length of chunked results matches the set range, with no excessive splitting or overly long paragraphs.
The values provided are common starting points and should be measured against the reader's own samples.