What data looks like for this category
The core data for this category comes from hot metal and scrap steel data, L2 process control, laboratories, MES, energy medium systems, and equipment status. Daily generated documents include heat plans, batching lists, converter smelting records, composition reports, continuous casting process records, and slab judgment sheets. Data update frequencies are per heat, real-time, and per shift. Documents contain fields such as heat identifiers, process parameters, composition values, and equipment status, with corresponding units including degrees Celsius, percentage, tons, meters per second, and others.
Constraints imposed by these characteristics on vector models and indexes
Since data is updated per heat, real-time, and per shift, vector indexes must support incremental updates to avoid resource consumption from full reconstruction. Since documents contain strongly associated information between heats and casting sequences, context from adjacent process steps must be retained during chunking to avoid loss of process logic due to breaks. Since data comes from multiple systems with diverse field types, indexes must support hybrid retrieval of structured parameters and unstructured text. At the same time, the accuracy of process data directly affects production outcomes, so the retrieval link must strictly control the recall rate of low-relevance results.
How to set configurations
The following configurations are adapted for FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the paragraph length of single-heat process records, retains context for core process parameters |
| Overlap Length | 100–150 characters | Prevents chunk breaks between heat plans and corresponding batching lists, associates information from adjacent process steps |
maxToken | 8192 token | Adapts to the maximum input limit of mainstream embedding models such as Alibaba Cloud's text-embedding-v3 |
| Similarity Threshold | 0.78–0.82, calibrated via actual testing | Filters low-relevance process documents in iron and steel smelting scenarios, avoids mismatches interfering with production decisions |
| Maximum Single Search Results | Top 15 results | Adapts to multi-step retrieval requirements for steelmaking and continuous casting, controls total context length within model limits |
| Reranking Module | Built-in semantic reranking | Optimizes sorting accuracy for multi-field mixed data, aligns with the association logic of process parameters |
Three common configuration mistakes
- Changing the embedding model results in errors or no results from existing knowledge base retrieval. Cause: The incremental index reconstruction process was not executed, and existing vector data was still generated based on the old model without synchronous updates.
- Knowledge base search result sorting does not match expected semantic relevance. Cause: The reranking module was not enabled, or the reranking module was not adapted to the multi-parameter mixed document structure in iron and steel smelting scenarios, relying only on initial recall similarity scores.
- Retrieval returns results with high semantic matching scores, but the actual document is irrelevant to the query. Cause: The similarity threshold was set too low, failing to filter low-relevance process records, or the chunk length was too long, causing redundant context to interfere with matching accuracy.
How to confirm configurations are set correctly
- Upload a single converter smelting record document, check that chunked content retains core fields such as heat numbers and composition values, with no obvious breaks.
- After changing the embedding model, perform incremental index reconstruction, check that vector generation progress for the existing knowledge base is normal, with no error prompts.
- Initiate a query containing specific process parameters, check that retrieval result sorting aligns with business logic, and adjust reranking module parameters to optimize matching effects as needed.
- View system operation logs, confirm that model access parameter configurations are correct, with no error messages such as "model not matched".
The values provided here are common starting points, and should be evaluated against the reader's own samples.