What this category’s data looks like
Daily documents for this category include foundry process cards, batching lists, heat records, mold history records, first-piece inspection records, material test reports, and nonconforming product review forms. Data sources include ERP material master data, melting furnaces and thermometers, spectrometers, MES heat records, mold management systems, customer drawings and technical requirements. Update frequency runs per heat, per batch, or daily. Most documents contain clear numerical fields such as alloy batching ratios, melting temperatures, and pouring parameters. They also include structured and semi-structured content like defect types and mold cycle numbers.
What constraints do these characteristics place on document parsing and chunking?
Multi-source heterogeneous documents in this category have inconsistent formats. Update frequencies run per heat, per batch, or daily. These factors create requirements for the adaptability and timeliness of document parsing and chunking. Documents contain clear numerical fields such as alloy batching ratios and melting temperatures. They also include associated information like defect types and mold cycle numbers. Chunking must retain logical relationships between fields. This avoids splitting that breaks process context. Some documents include embedded images such as metallographic diagrams and mold structure diagrams. Parsing must extract image content and bind it to corresponding text. This ensures full information can be retrieved during Q&A.
Configuration Settings
This document uses FastGPT v4.17.0 for configuration.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the length of paragraphs in foundry documents such as process parameters and heat records. It avoids splitting that breaks individual process logic. |
chunkSplitter | Paragraph + semantic chunking | Adapts to semi-structured process documents. It retains semantic completeness of batching, temperature, and defect descriptions. |
customPdfParse | Enable embedded image parsing mode | Adapts to embedded images such as metallographic diagrams and mold diagrams in foundry documents. It extracts image text and associates it with context. |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Supports upload requirements for single-heat batch records or large mold history documents. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Meets parsing duration requirements for large melting records and multi-page process manuals. |
chunkSettingMode | Manual configuration | Targets the semi-structured characteristics of foundry documents. It customizes chunking rules instead of using generic templates. |
Three Common Mistakes
- After uploading a batching list with embedded images, the parsing result only includes text content. Images are not recognized. The cause is failing to enable the embedded image parsing switch for
customPdfParse, and not configuring chunking rules to associate images with text. - After uploading a large heat record PDF, the parsing task times out and throws an error. The cause is that
PARSE_FILE_TIMEOUT_SECONDSis set lower than the actual time required for document parsing. It does not adapt to the page count and data volume of foundry documents. - Split process card paragraphs lose the associated information between alloy batching ratios and corresponding heat numbers. The cause is that
chunkSplitteris configured to split by fixed character count, instead of splitting by semantic boundaries. This breaks logical binding between fields.
How to Confirm the Configuration Is Correct
- Upload a foundry process card that includes embedded metallographic diagrams. Check if the OCR text content of the images is included in the parsing result.
- Upload a heat record document with more than 10 pages. Confirm that the parsing task completes within the duration set by
PARSE_FILE_TIMEOUT_SECONDSwithout timeout errors. - Randomly select one chunked result. Verify that it includes complete associated field information such as alloy batching ratios and corresponding melting temperatures.
- Upload a mold history document under 1000 MB. Confirm that the upload and parsing processes complete normally.
The values listed are common starting points, and should be evaluated against applicable deployment-specific samples.