Overview
- Client industry: Steel rolling and finishing
- Business direction: Supplier and process quality audit
- Capability area: Source citation and traceability
What Data Looks Like for This Category
This category primarily covers hot rolling, cold rolling, section steel, and steel pipe manufacturers. Daily documents include rolling plans, heating regulations, rolling specifications, plate shape and thickness records, surface quality reports, and quality certificates. Data sources include slab/billet batch information, furnace parameters, mill automatic control systems, online thickness and width measuring equipment, surface inspection devices, and MES systems. Data updates in real time at the production node for each coil or individual steel product. Data fields include batch number, thickness tolerance (mm), surface defect level, heating temperature (℃), rolling speed, and more. Each coil’s data links to upstream slab information and downstream quality inspection records.
Constraints on Source Citation and Traceability
Three core constraints apply to the source citation and traceability link from this category’s data characteristics. First, production data updates in real time per coil or individual product. Traceability must target the exact batch number of a single steel coil or product. Fuzzy time period matching cannot be used. Second, data sources are heterogeneous and multiple, including mill control systems, online inspection equipment, and MES systems. The citation link must associate fields across different systems. This ensures returned raw data corresponds to the same production node. Third, document types are diverse, and data volume per coil is large. Chunking and citation configurations must adapt to the structure and length of different documents such as rolling specifications and quality certificates. This prevents loss of critical information during chunking or citation.
How to Configure Settings
Configurations are adjusted based on FastGPT v4.17.0 to meet traceability requirements for steel rolling and finishing scenarios:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
quoteMaxToken | 800–1200 token | Long documents such as quality records and rolling specifications per coil require sufficient original citation text to be retained |
responseData | Enable batch association fields | Traceability data such as slab batches and mill parameters must be linked. This ensures citation content corresponds one-to-one with production batches |
detail | Enable full citation | Raw records such as plate shape, thickness, and surface inspection results must be fully displayed. This allows auditors to access all critical information |
| Chunk length | 1000–1500 characters | Adapts to the single-section content length of documents such as rolling plans and quality certificates. Prevents damage to data correlation during chunking |
| Single search limit | Top 6 entries | The number of upstream slab and downstream quality inspection records linked to a single steel coil usually does not exceed 6 groups. This covers all associated data |
| Reranking | Sort by production sequence | Traceability results must be displayed in production node order. This allows auditors to clearly track the production process |
Three Common Mistakes
- Phenomenon: After calling the system, returned content only includes citation fragments, with no integrated natural language answer. Cause:
responseDatais configured to only return raw data, and natural language generation logic is not enabled. - Phenomenon: Individual document citation display permissions cannot be controlled separately. The view original button status is consistent for all files. Cause: Citation visibility rules are not configured separately for different document types such as rolling specifications and quality certificates.
- Phenomenon: After the knowledge base is chunked, some quality certificate files do not generate question-answer pairs and are directly written as raw text. Cause: Chunk length does not adapt to the long-text structure of quality certificates. This triggers abnormal logic during chunking.
How to Verify Correct Configuration
- The rolling plan and quality certificate for a single steel coil are uploaded. A query for "quality status of this batch" is initiated. Returned results are checked to confirm they include raw data fragments corresponding to the coil number.
- The
quoteMaxTokenconfiguration value is checked. The returned citation text length is confirmed to meet expectations. - Citation visibility for different document types is verified. View original permissions for individual files are confirmed to be controllable separately.
- The knowledge base chunking log is checked. All rolling-related documents are confirmed to have generated correct question-answer pairs.
The values provided are common starting points and should be measured against the reader's own samples.