What the data for this category looks like
Data sources for carbon steel intelligent due diligence reports mainly include steel mill factory quality inspection reports, supply chain delivery ledgers, and industry association supply and demand statistics documents. Update frequency varies by data source type: factory reports are updated in real time with production batches, delivery ledgers are updated monthly, and industry statistics documents are updated weekly or monthly. A single document usually includes batch identification, production information, physical and chemical test data, dimensional parameters, and compliance statements. Batch documents are packaged and integrated by batch. Core fields include batch number, heat number, tensile strength, thickness, single coil weight, etc. The unit of tensile strength is megapascals (MPa), thickness is millimeters (mm), and single coil weight is tons (t).
What constraints do these characteristics impose on the document parsing and chunking link?
Diverse data sources lead to differences in parsing formats: structured quality inspection reports contain fixed tables, semi-structured delivery ledgers have field offsets, and unstructured industry documents have no fixed format, requiring targeted adaptation of parsing logic. Core fields are tightly bound to units. When splitting, the association between test items, corresponding values, and units must be retained to avoid field confusion. Single packages of batch documents have large volume, so support for large file parsing and batch task scheduling is required, while also avoiding context interference across batches. Some documents contain steel mill-customized test items, so the field integrity of the original text must be retained, and forced standardized splitting is not allowed.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
marker_ocr_batch_size | 4-6 | Carbon steel documents are mostly tabular structured content. Small-batch OCR reduces recognition errors and avoids memory overflow |
parse_chunk_size | 800-1200 characters | The single-block data of physical and chemical test tables in carbon steel documents is approximately 500-800 characters. Reserve context space to ensure complete field association |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | A single batch-packaged collection of carbon steel quality inspection reports usually requires a long parsing time. This value adapts to most batch scenarios |
enable_field_binding | Enabled | Fields and values in carbon steel documents are strongly correlated. Enabling this option retains the binding relationship between test items and corresponding data |
UPLOAD_FILE_MAX_SIZE | 2000 MB | Batch-packaged collections of carbon steel quality inspection reports usually reach 1-1.5 GB. Reserve sufficient upload space |
ocr_language | Chinese + steel industry terminology library | Carbon steel documents contain a large number of industry-specific terms. Loading the corresponding terminology library improves recognition accuracy |
The parameter values provided on this page are all conventional recommendations used as a starting point for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing.
Three common mistakes
- Phenomenon: After deploying version 2 of marker, the log shows an
ocr errorerror. Cause: The steel industry professional OCR terminology library was not loaded, resulting in failed recognition of section steel identifiers and test item names. - Phenomenon: After connecting to the large model, normal chat functions work normally, but an error is triggered during the file parsing link. Cause: The
enable_field_bindingparameter was not configured, causing the association between fields and values of structured tables to be lost after splitting, triggering format verification failure during large model parsing. - Phenomenon: After deploying version 4.8.22 locally, the file parsing function fails. Cause: The
UPLOAD_FILE_MAX_SIZEparameter was not configured correctly, causing batch documents to be intercepted by the system during upload and unable to enter the parsing link.
How to confirm the configuration is correct
- Upload a single carbon steel factory quality inspection report, check whether the parsed text retains core associated identifiers such as batch number and heat number, and confirm that the chunking result does not split a complete set of physical and chemical test data.
- View the content after OCR recognition, confirm that professional terms such as "tensile strength" and "yield strength" are not misrecognized or garbled, and verify the loading status of the OCR language configuration.
- Upload a batch-packaged collection of carbon steel documents, check the execution status of parsing tasks, and confirm that task interruptions are not triggered due to timeout.
- Trigger structured table parsing, check the binding relationship between fields and corresponding values, and confirm that no separation of fields and data has occurred.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.