What Data for This Category Looks Like
Data sources for this category include hot metal and scrap data, L2 process control systems, laboratory test data, MES production management systems, energy medium systems, and equipment condition monitoring data. A heat plan and batching list are generated per heat. Converter smelting records and continuous casting process data are synced in real time. Slab determination sheets and composition reports are summarized per shift. Documents include fixed fields. For heat plans, these include cast sequence number and steel grade. For smelting records, these include smelting duration and endpoint temperature. Field units include physical quantities such as degrees Celsius, tons, and meters per minute.
Constraints Imposed by These Characteristics on Vector Models and Indexes
Real-time and per-heat update frequencies require indexes to support incremental synchronization. This avoids time costs and business interruption risks from full reconstruction. Most documents are structured or semi-structured. Chunking must match document field boundaries. This prevents semantic breaks from incorrect splitting. Data includes many physical parameters. Vector models must adapt to encoding logic for numerical features. This ensures retrieval accuracy. High-temperature continuous process business characteristics require high index availability. Retrieval interruptions from configuration issues must not disrupt production scheduling.
How to Configure Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the single-document length of converter smelting records and continuous casting process records, avoids semantic splitting |
| Overlap Length | 100–150 characters | Retains context for adjacent chunks, prevents critical process parameters from being split across different chunks |
| Similarity Threshold | 0.75–0.85 | Matches matching accuracy for numerical data such as steel composition and temperature, filters low-relevance results |
| Single Search Limit | Top 10 entries | Adapts to production scheduling retrieval needs, avoids returning excessive redundant data |
defaultToken | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
maxToken | 12000 token | Adapts to the maximum token count of full continuous casting process records, prevents key content from being truncated |
Three Common Configuration Errors
- Symptom: Retrieval results are empty after custom index configuration. Cause: Chunking is not configured per the fixed field structure of smelting documents. Random splitting of long documents causes semantic breaks. This prevents matching of retrieval requests.
- Symptom: Retrieval requests time out. Cause: Incremental index synchronization is not enabled. Each retrieval triggers a full data scan. This cannot adapt to the real-time update scale of heat data.
- Symptom: Vector model calls return parameter errors. Cause: Numerical data such as composition and temperature is not normalized. This causes vector dimensions to mismatch configuration requirements.
How to Confirm Proper Configuration
- Upload a single complete converter smelting record document. Check the chunking preview interface. Confirm that key fields such as endpoint temperature and tapping volume are not split across different chunks.
- Submit a retrieval request that includes target steel grade composition parameters. Check the similarity scores of retrieval results. Confirm that scores fall within the preset threshold range.
- Simulate generating a new heat exception report. Wait one minute, then submit a retrieval request. Confirm that the report has been indexed and appears in the result list.
- Check the FastGPT v4.17.0 system logs. Confirm that token consumption from vector model calls does not exceed the
maxTokenconfiguration value. No truncation error messages appear.
The values listed here are common starting points. They should be evaluated against the reader's own samples.