What the data for this category looks like
Data for this category mainly comes from hot rolling, cold rolling, section steel and steel pipe production processes, and covers documents such as rolling plans, heating schedules, rolling specifications, plate shape and thickness records, surface quality reports, and quality certificates. Data sources include slab/billet batch information, heating furnace operating parameters, mill automatic control data, online thickness and width measurement results, surface inspection results, and MES system records. The data update frequency is real-time per coil or per piece. Document fields include batch numbers, process parameters, test values, defect locations and grades, and units are mostly industrial standard units such as millimeters, degrees Celsius, and square meters.
What constraints do these characteristics impose on the "source citation and traceability" link
The characteristics of real-time coil-level updates, multi-source industrial data correlation, and strong batch binding for this category impose multiple constraints on the source citation and traceability link. First, it is necessary to accurately match slab/billet batch numbers to avoid citing unrelated process or test data across batches, and prevent deviations in traceability results. Second, real-time online test data must be synchronously linked with offline quality certificates and other documents to ensure that traceability results cover the entire process from raw materials to finished products. Third, the core requirement of coil-level traceability requires that citation results must be bound to specific coil numbers, cannot generalize generic process specifications, and need to clearly distinguish general rules from actual execution parameters for individual coils.
Configuration Settings
Configuration references under FastGPT v4.17.0 are as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
quoteMaxToken | 8000–12000 token | Matches the total length of full-process documents for a single coil, avoids truncating critical batch and test data |
| Single search limit | Top 8 results | Covers the number of multi-source test and process records under the same batch, avoids missing key traceability information |
responseData | Enable source file name and coil number fields | Directly extract bound coil numbers and source identifiers from documents, no additional parsing required |
detail | Full mode | Retains the original document's field hierarchy, making it easier to verify traceability details such as defect locations and process parameters |
| Maximum context | 600 seconds | Adapts to the rhythm of real-time industrial data updates, ensuring retrieval calls the latest batch-associated data |
| Reranking | Sort by batch priority | Prioritizes returning documents from the same batch, reducing traceability matching time |
Three Common Misconfigurations
- Retrieval results only return 1 document. The cause is that the single search limit value was not adjusted. The default configuration restricts the number of recalled results, making it impossible to cover multi-source data from the same batch.
- The knowledge base is empty when called, requiring manual selection of the target dataset. The cause is that no default bound coil number-associated knowledge base was configured, so the system cannot automatically match the corresponding batch's document set.
- Cited source documents cannot be downloaded. The cause is that full field extraction for
responseDatawas not enabled, and original document storage paths and access permission identifiers were not retained.
How to Confirm Proper Configuration
- Initiate a traceability query for a specific coil number, and check whether the returned results include all associated documents for that coil number.
- View the source fields of the returned results, and confirm that complete file names, coil numbers and data source identifiers are included.
- Test adjusting the
quoteMaxTokenvalue, and verify that all content of long documents can be fully returned. - Check the sorting logic of retrieval results, and confirm that documents from the same batch are displayed first.
The values provided are common starting points and should be evaluated against the reader's own samples.