What this category’s data looks like
This category’s data comes from iron making, steel making, continuous casting, quality control, and scheduling departments of integrated steel enterprises. It includes documents such as heat plans, batching lists, converter smelting records, composition reports, continuous casting process records, and slab judgment sheets. Data sources include hot metal and scrap data, L2 process control data, laboratory test data, MES system data, energy medium system data, and equipment status data. Updates trigger per heat, in real time, or per shift. This is high-temperature continuous process data, with strong coupling between heats and casting runs.
Constraints imposed on vector models and indexing
Data updates per heat, in real time, or at high frequency. Index services must support incremental synchronization and low-latency responses, to prevent production audit delays from update backlogs. Documents contain multiple field types: process parameters, composition values, defect judgments, and more. This requires support for vectorization of mixed text and structured fields, to ensure integrity of associated information. There is strong coupling between heats and casting runs. Associated data must be grouped and indexed using unique identifiers such as heat numbers and casting run numbers, to prevent cross-batch data confusion. High-temperature processes have no downtime windows. Index services must maintain high availability, to avoid disrupting production rhythms from service interruptions.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the text length of single-heat records, avoids splitting cross-heat data across chunks |
| Overlap Length | 100–150 characters | Preserves contextual association between heat plans and process records, prevents key information from being truncated |
| Similarity Threshold | 0.75–0.85 | Distinguishes similarity between different records of the same heat, avoids mismatching cross-heat data |
| Single Search Limit | Top 10 results | Matches the multi-record associated audit needs of steel smelting scenarios, does not overload search results |
defaultToken | 4096 token | Adapts to the vectorization upper limit for long documents, covers complete process record content for a single heat |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
Three Common Misconfigurations
- Phenomenon: The dataset.csv generated when importing a knowledge base only contains the index field, with no content field. Cause: The text extraction rules during document parsing are not correctly configured, and the core text fields of smelting records are not associated, resulting in the content field not being generated.
- Phenomenon: A timeout error occurs after starting the index service, and the logs show request timeout. Cause: The high-frequency update rhythm of steel smelting data is not adapted, and index synchronization threads are not allocated resources reasonably, leading to backlogged real-time update tasks timing out.
- Phenomenon: The custom index cannot match the associated fields required by the business. Cause: Core business identifier fields such as heat numbers and casting run numbers are not explicitly bound, and the default index rules are used directly, making it impossible to associate with the corresponding business data tables.
How to Verify Correct Configuration
- Import the complete process record for a single heat, check that the chunked text retains heat-related information, with no cross-heat data mixed together.
- Run a batch real-time update task, check that the index synchronization delay meets business requirements, with no obvious task backlogs.
- Search for the composition report for a specific heat, confirm that the returned results only contain associated records for that heat, with no cross-heat matching content.
- Check the vector model's running logs, confirm that the
defaultTokenandmaxTokenconfigurations are loaded correctly, with no parameter conflicts.
The values provided are common starting points and should be measured against the reader's own samples.