What This Category of Data Looks Like
This category of data comes from ERP material master data, melting furnaces and thermometers, spectrometers, MES furnace batch records, mold management systems, customer drawings and technical specifications. The update frequency is per furnace batch, per batch, or daily. Daily documents include foundry process cards, batching lists, furnace batch records, mold history records, first article inspection records, material test reports, and nonconforming product review forms. Fields include alloy composition proportion, melting temperature, pouring parameters, furnace batch number, mold times, and more. Units involve degrees Celsius, percentage, kilograms, and other standard units.
Constraints Imposed on Document Parsing and Chunking
Foundry documents have both structured table and semi-structured text characteristics, and include unique associated identifiers such as furnace batch number and mold times. When chunking, context association between fields must be retained to avoid splitting that breaks data correspondence. Document update frequency is high, and single-batch document content changes with production adjustments. Fixed chunking rules must be avoided. Multi-source data must be integrated into bid documents. The parsing link must support cross-document field matching, and accurately extract core parameters such as melting and pouring to provide accurate basis for subsequent quotation and bid verification.
Configuration Settings
| Configuration Item | Recommended Approach | Basis for This Approach |
|---|---|---|
chunkSize | 800–1200 characters | Foundry documents contain multi-field associated content. Excessive length will lose context, while insufficient length will break the association between furnace batch and mold times |
chunkSplitter | Chunk by paragraph + semantics | Retain field association of batching lists and furnace batch records, and adapt to mixed semi-structured and structured document types |
customPdfParse | Enable structured parsing mode | Adapt to the fixed table structure of foundry process cards and material test reports, and accurately extract preset fields |
PARSE_FILE_TIMEOUT_SECONDS | 120–180 seconds | Single multi-furnace batch documents take a long time to parse. Avoid parsing failure caused by a threshold that is too low |
chunkSettingMode | Manual configuration | Adjust chunking parameters for different document types such as furnace batch records and customer drawings to meet bid document integration requirements |
This configuration is based on FastGPT v4.17.0.
Three Common Errors
- Phenomenon: After uploading a batching list document, the furnace batch number and corresponding alloy batching ratio are separated in the chunking result. Cause: The
chunkSizesetting is too small, which splits associated fields of the same furnace batch and destroys data integrity. - Phenomenon: After importing images from foundry process cards, images cannot be loaded normally in the knowledge base. Cause: The structured parsing mode of
customPdfParseis not enabled, and the associated metadata of images and documents is not retained, resulting in loss of image-related information. - Phenomenon: Large multi-furnace batch record documents fail to parse, returning timeout-related errors. Cause: The
PARSE_FILE_TIMEOUT_SECONDSsetting is too low, which does not adapt to the parsing duration requirements of multi-furnace batch documents.
How to Confirm Configuration is Correct
- Upload a single furnace batch record document, check whether the chunking result retains the associated content of furnace batch number and corresponding batching ratio.
- Upload a material test report with tables, confirm that table fields are complete without missing or misalignment after structured parsing.
- Test a multi-furnace batch document with more than 10 pages, confirm that the parsing process does not trigger a timeout error.
- Upload a mold history document with embedded images, confirm that images can be loaded normally and associated document metadata is not lost.
The values provided here are common starting points and should be measured against the reader's own samples.