What This Category of Data Looks Like
The data sources for this category cover slab or billet batches, heating furnaces, mill automatic control systems, online thickness and width measuring equipment, surface inspection devices, and manufacturing execution systems (MES). Daily documents include rolling plans, heating schedules, rolling regulations, plate shape and thickness records, surface quality reports, and quality certificates. Data is updated per coil or billet, with some parameters synchronized in real time. Data fields include batch number, thickness (unit: mm), width, surface defect level, rolling temperature, reduction amount, and others, with coil level as the core traceability unit.
Constraints Imposed on Vector Models and Indexes
The high-frequency update feature at the coil level requires indexes to support low-latency writing and real-time synchronization, to avoid lag in batch data. The structural differences across multiple document types require vector models to adapt to both structured process parameters and unstructured report text. Some scenarios also support image features from surface inspections. The core requirement of coil-level traceability requires vector indexes to bind batch numbers and coil numbers as associated primary keys. This ensures full-link data for a single coil can be accurately located during retrieval. Data volume grows linearly with production scale. A horizontally scalable index cluster must be configured to avoid overloading a single node.
Configuration Settings
This configuration is compatible with FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the typical paragraph length of rolling regulations and surface quality reports, avoiding single chunks exceeding the maxToken limit |
| Overlap Length | 100–200 characters | Covers cross-chunk data associated with batches, ensuring contextual integrity during coil-level traceability |
| Similarity Threshold | 0.75–0.85 | Distinguishes similar surface defect features from false positives, adapting to the fine difference judgment requirements of steel surface inspections |
| Single Search Limit | Top 20 entries | Covers the volume of associated data across the entire production process of a single coil, avoiding redundant retrieval results |
defaultToken | 512 | Matches the typical token count of a single structured process parameter, adapting to the vectorization of fields such as rolling temperature and reduction amount |
maxToken | 4096 | Adapts to the vectorization requirements of long documents such as complete rolling plans, avoiding truncation of key process information |
Three Common Misconfigurations
- Issue: A parameter validation failure prompt appears when configuring the vector model. The cause is failing to adjust chunk and
maxTokenparameters based on steel rolling document lengths, leading to truncation of key process information in single chunks. - Issue: The index service cannot connect to production data normally in a docker-compose deployment environment. The cause is failing to specify coil-level batches as index-associated primary keys in the configuration file, resulting in data being unable to be stored by traceability unit.
- Issue: Index construction continues to time out, and the number of retrieval results is far lower than expected. The cause is failing to adjust the batch writing strategy. One-time submission of high-frequency real-time coil-level data exceeds the cluster's processing capacity.
How to Confirm Successful Configuration
- Import the rolling plan and surface quality report for a single coil, run a vector retrieval, and check if the results include full-link associated data for that coil.
- Submit a new piece of real-time production data, wait 1 minute, then retrieve that batch number to confirm the index has been updated synchronously.
- View the vector service runtime logs to confirm there are no error messages about
maxTokenexceeding limits ordefaultTokenconfiguration errors. - Adjust the similarity threshold to verify the matching accuracy of retrieval results, confirming it meets the business requirements for steel defect judgment.
The values provided are common starting points and should be measured against the reader's own samples.