What the data for this category looks like
Core data for steel rolling and finishing comes from slab or billet batches, heating furnaces, mill automation controls, online thickness and width measurement systems, surface inspection systems, and MES systems. Daily documents include rolling plans, heating schedules, rolling specifications, flatness and thickness records, surface quality reports, quality certificates, and more. Data updates occur in real time per coil or per individual steel product. Documents include specific fields and units, such as rolling temperature (℃), reduction rate (%), thickness tolerance (mm), surface defect grade, and others. Some documents are in PDF format with structured parameter tables.
What constraints do these characteristics impose on the document parsing and chunking link
Real-time or coil-level updated data requires parsing and chunking processes to support high-frequency document handling, to avoid delaying batch scheduling due to parsing delays. Coexisting multi-type documents require adapting parsing logic for both long-text rolling specifications and structured parameter table records, to avoid splitting critical parameter groups. The requirement to retain specific fields and units means the chunking process must not break the association between units and values, to prevent parameter ambiguity during Q&A. The coil-level traceability business requirement means chunked content must be linked to corresponding batch identifiers, to avoid parameter confusion across batches.
How to set the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Balances context completeness and chunk granularity for long rolling specification text, avoids splitting critical parameter groups |
chunkSplitter | Mixed paragraph + fixed length | Ensures reasonable chunking for both structured documents such as rolling plans and heating schedules, and long-text specifications |
customPdfParse | Enabled | Required to parse PDF documents with parameter tables such as quality certificates and flatness records, retains original layout and fields |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Large rolling specification PDFs or batch document sets take longer to parse, prevents timeout interruptions |
chunkSettingMode | Manual configuration | Adjusts chunking parameters for different document types (such as specification documents, record documents) to fit structured data in steel rolling scenarios |
This configuration applies to FastGPT v4.17.0
Three common errors
- Phenomenon: Context references do not retain the original document layout, only plain text or incorrect formatting markers are displayed. Cause: The
customPdfParseconfiguration is not enabled. Default parsing does not retain table and format fields, leading to misalignment of units and values in rolling parameter tables. - Phenomenon: External links are submitted, but rolling plan document content cannot be parsed, only empty results are returned. Cause: The document format pointed to by the link is not confirmed to be PDF or structured text, parsing rules for external link documents are not adapted, or the timeout setting is too short, causing parsing interruptions.
- Phenomenon: Single chunk ID cannot be copied in the knowledge base chunk list. Cause: Chunk ID copy permission is not enabled in system settings, or the front-end rendering logic does not load the corresponding interactive component, making it impossible to obtain chunk identifiers linked to batches.
How to confirm the configuration is properly set
- Upload a PDF document of hot-rolled flatness and thickness records, check if the parsed text retains the values and units of thickness and flatness.
- Submit a query containing rolling temperature parameters, confirm that the chunked content referenced in the context includes a complete heating schedule parameter group.
- Upload a single rolling specification document not exceeding 1000 MB, verify that the parsing process does not trigger a timeout error.
- Enter the knowledge base chunk list, try clicking any chunk ID area, confirm that its content can be copied.
The values provided are common starting points and should be measured against the reader's own samples.