What this category's data looks like
Data for this category comes from slab or billet batch ledgers, furnace operating parameters, mill automatic control systems, online thickness and width measuring equipment, surface inspection systems, and MES systems. Data updates per coil or individual billet. Some parameters synchronize in real time. Daily documents include rolling plans, heating schedules, rolling specifications, shape and thickness records, surface quality reports, and quality certificates. Fields include batch number, thickness (mm), width (mm), heating temperature (℃), surface defect level, rolling speed, and others. Most formats are structured tables or industry-standard documents.
Constraints Imposed on Multi-turn Dialogue and Prompt Engineering
This category’s data updates per coil or individual billet, with some parameters synchronized in real time. Multi-turn dialogue must bind exclusive context for the corresponding batch. This avoids mixing data across batches. Professional parameters have clear units. Prompts must explicitly restrict the LLM to only access rolling and inspection document content linked to the current batch. Historical dialogue content such as shape deviation and thickness values acts as reference for subsequent audits. An appropriate amount of dialogue context must be kept for parameter comparison. Inspection and audit involve serious consequences including steel coil downgrading and equipment damage. LLM output must be strictly limited to category-specific data in the knowledge base.
Configuration Setup
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
chatId | Bind the current billet batch number | Avoid confusion across batch dialogue contexts, adapts to the category’s characteristic of updating data per coil or individual billet |
messages | Retain the most recent 6 rounds of dialogue context | Enables comparison of historical thickness deviation, shape level, and other parameters, meets reference needs for multi-turn audits |
variables | Inject the current batch’s rolling plan ID and surface inspection report number | Restrict the LLM to only retrieve knowledge base documents associated with the current batch, prevent calls to irrelevant data |
| Maximum context length | 8000–12000 token | Accommodates full-process records and multi-turn dialogue content for a single steel coil, avoids truncating critical structured data |
| Single recall limit | Top 3 matching documents | Documents related to a single batch in this category have standardized structures and limited quantity. Excessive recall increases redundant computation |
| Query optimization | Calibrate based on actual testing | Convert professional rolling terminology into retrieval instructions the LLM can accurately recognize, adapts to category-specific fields and units |
Three Common Misconfigurations
- Phenomenon: The LLM outputs audit conclusions not limited to the current batch’s knowledge base. Cause: The
variablesfield is not used to bind the current billet’s batch identifier, leading the retrieval scope to cover all documents. - Phenomenon: Redundant multi-turn dialogue context slows system loading speed. Cause: The retention rounds for
messagesare not adjusted based on the independent characteristics of single-batch data in this category. The default configuration does not adapt to the data update rhythm. - Phenomenon: Markdown-formatted quality report output is truncated and cannot be fully displayed. Cause: The maximum context length setting does not match the character volume of full records for a single steel coil, leading to forced truncation of LLM output.
How to Verify Proper Configuration
- Bind the
chatIdfor a specific steel billet batch, submit an audit request, check that retrieved documents only include rolling plans, thickness records, and other content for that batch. - Input parameters for different batches sequentially to start multi-turn dialogue, check that the system can distinguish contexts and only call document data for the corresponding batch.
- Input a query containing professional rolling terminology, check that the LLM can accurately recognize and link to the corresponding fields and units.
- Generate an audit conclusion containing a Markdown table, check that the output content is complete and no truncation occurs.
The values provided here are common starting points. They should be evaluated against the user’s own sample data.