What this category of data looks like
Data sources for this category include hot metal and scrap charging data, real-time smelting parameters collected by L2 process control, composition reports issued by the laboratory, production scheduling information from the MES system, energy medium consumption records, and equipment status monitoring data. Daily generated documents include heat plans, batching lists, converter smelting records, continuous casting process parameter records, and slab quality judgment sheets. Data update frequencies cover each heat production cycle, real-time process nodes, and shift summary nodes. Document fields include heat number, charging weight, smelting temperature, composition detection values, continuous casting pulling speed parameters, and more.
What constraints these characteristics impose on model access and configuration
Since data is updated by heat, in real time, and across multiple nodes, model access must support incremental synchronization mechanisms to avoid excessive computing resource usage from full data pulls. Multiple types of documents contain business fields with different structures. During configuration, parsing rules for each document must be adapted to ensure core parameters such as temperature, composition, and pulling speed are correctly extracted. The timeliness of real-time process data requires model call response times to match production rhythms. At the same time, exception report drafting must associate full-link data for the corresponding heat, so the context window must cover all associated information for a single heat. The strong correlation between equipment and energy data requires retaining associations between fields during configuration to avoid logical deviations in generated process documents.
How to configure settings
The following configurations are adapted to the actual needs of iron and steel smelting and continuous casting scenarios based on FastGPT v4.17.0:
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | The total token count of documents associated with a single heat, including smelting records, composition reports, continuous casting parameters, and more, usually falls within this range to ensure full production-related data is covered |
temperature | 0.2–0.4 | Process documents and exception reports require professional rigor. A lower temperature improves content consistency and parameter accuracy |
AIPROXY_API_ENDPOINT | Local proxy address of the deployment instance | Adapts to the intranet isolation environment of iron and steel enterprises, avoiding data security risks from public network transmission |
AIPROXY_API_TOKEN | Application-specific secret key | Distinguishes call permissions for different departments such as ironmaking, steelmaking, and continuous casting, meeting data isolation requirements in industrial scenarios |
maxTokens | 1500–2000 characters | The output length of a single process document or exception report matches the standard length of industrial documents, avoiding redundant or insufficient content |
vision | Disabled | The core data in this scenario is structured text and numerical parameters, with no image input requirements. This reduces model call costs |
Three common configuration mistakes
- Phenomenon: After creating an index model, a model not found prompt appears when calling the language model. Cause: The configuration entry for the index model and language model is confused, and the deployment parameters of the index model are mistakenly used to replace the language model configuration.
- Phenomenon: No application-specific API key can be generated in the locally deployed version. Cause: The application key generation option is not enabled in the platform settings page, and a universal key is used directly for calls.
- Phenomenon: Some heat composition detection data is missing from the generated process documents. Cause: Document parsing rules adapted to the field structure of composition reports are not configured, resulting in core data not being correctly extracted.
How to verify a successful configuration
- Upload a typical document set for a single heat, initiate a model call, and check whether the returned content covers all associated information from the uploaded documents.
- Generate a single exception report, and check whether the output content includes core business fields such as heat number, smelting parameters, and abnormality judgment.
- View the platform call logs to confirm that the model response time matches the rhythm requirements of the current production link.
- Verify the permissions of the application-specific secret key, confirming that keys for different business departments can only access knowledge base data within their corresponding scope.
The values provided are common starting points and should be measured against the reader's own samples.