What Data for This Category Looks Like
Quality data for the photovoltaic module encapsulation process comes from battery grading records, glass adhesive film and backplane batch ledgers, MES systems, string welding and lamination equipment logs, EL/IV tester outputs, insulation withstand voltage test reports, and is generated by the string welding, lamination, framing, and testing departments. Daily documents include layout drawings, material batch records, lamination process parameter sheets, EL defect images, IV test reports, packing lists, and more. Each document is linked to a single module or a batch of materials, with an update frequency of once per module or batch. Fields include batch number, current gear, lamination temperature, crack coordinates, insulation resistance, and others. Some documents are structured tables, some are images, and some are unstructured text formats.
Constraints Imposed on Document Parsing and Chunking
Quality data for photovoltaic module encapsulation has a high update frequency, and each document is linked to a single module or batch of materials. This requires chunking to precisely match specific production units to avoid cross-batch confusion. Documents contain multiple formats including structured tables, image-based EL defect maps, and unstructured process records. Mixed format parsing must be supported to ensure key information such as crack coordinates and test data in images is not lost. Fields include professional parameters such as current gear and insulation resistance. Chunking must retain the association between fields and units, and parameter groups spanning multiple fields cannot be split. Documents from different sources have large layout differences, so parsing rules must be adapted for different file types such as layout drawings and PDF process sheets.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapt to the data volume linked to a single batch of materials or single module, avoid cross-batch confusion caused by overly large chunks, while ensuring complete process parameter groups are contained within each chunk |
chunkSplitter | Split by paragraph + tab characters | Adapt to documents mixing structured tables and unstructured text, prioritize splitting by paragraph, use tab characters as separators for test report tables to ensure fields are not incorrectly split |
customPdfParse | Enable structured extraction mode | For PDF format documents such as lamination process records and batch ledgers, enable custom PDF parsing mode to retain fields like batch numbers and temperature parameters in tables, avoid losing structured information during plain text parsing |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Adapt to documents containing large-size images such as EL images and IV test reports, reserve sufficient parsing time to avoid parsing failure due to timeout |
chunkSettingMode | Manual configuration mode | For different types of documents such as layout drawings and test reports, adjust chunking parameters separately to adapt to diverse document formats in the photovoltaic module encapsulation scenario |
Three Common Configuration Mistakes
- Phenomenon: After uploading documents split using a local vector model to the server, context references show batch association errors. Cause: Chunking granularity was not adjusted based on the single-module/single-batch data characteristics of photovoltaic module encapsulation, and chunking differences across models lead to failure to precisely match production units.
- Phenomenon: Parsed chunk content has incorrect parameter splitting, such as lamination temperature and battery current gear being assigned to different chunks. Cause: The separation rules for
chunkSplitterwere not correctly configured, and splitting was not done according to professional parameter groups, leading to breakage of key associated information. - Phenomenon: Document content referenced in context cannot be rendered as Markdown format, only raw text is displayed. In FastGPT v4.17.0, passing a link to an EL test image fails to parse defect coordinate information, only returning empty content. Cause: The format retention switch for
customPdfParsewas not enabled, and configurations were not updated to adapt to the new version of parsing tools, leading to format information and image parsing failures.
How to Verify Proper Configuration
- When a lamination process record PDF is uploaded, the parsed chunks can be verified to contain complete batch number, lamination temperature, and time parameter groups with no cross-parameter splitting.
- When a structured description document for EL defect images is uploaded, context references can be checked to confirm they retain key information such as image coordinates and defect types with no format loss.
- When
chunkSizeis adjusted to different ranges and the same batch ledger is uploaded, the chunking results can be compared to confirm alignment with the association scope of single-batch data. - When an EL image link is tested, the parsed returned content can be checked to confirm it contains defect-related fields with no timeout or empty content errors.
The values provided are common starting points and should be measured against the reader's own samples.