What this category’s data looks like
Data for this category comes from ironmaking, steelmaking, continuous casting, quality control and scheduling departments of integrated iron and steel enterprises. It covers documents including heat plans, batching lists, converter smelting records, composition reports, continuous casting process records, and slab inspection sheets. Data sources include hot metal and scrap data, L2 process control data, laboratory test results, MES systems, energy medium systems, and equipment status data. Update frequency is per heat, real-time, or per shift. Most documents include structured fields such as heat number, steel grade, composition parameters, temperature, and casting speed. Units follow standard industrial conventions, including tons, degrees Celsius, and meters per minute. This differs from the raw material and process logic of non-ferrous smelting categories.
What constraints do these characteristics impose on knowledge base retrieval and recall
The high-frequency update feature of this category of data requires the retrieval and recall link to support incremental synchronization and real-time query. This avoids impacting process adjustments due to data lag. The multi-source heterogeneous document structure and industry-specific fields require the retrieval mode to balance structured field matching and full-text semantic recall. The tightly coupled business logic between heats and casting sequences requires recall results to associate multiple document types from the same production batch. This prevents decision bias caused by isolated data. The business characteristic of severe consequences from errors requires retrieval and recall results to have high confidence. This reduces invalid or incorrect process references.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Industrial process documents contain many specialized terms. This range filters low-correlation results and ensures process relevance of recalled content |
limit | Top 8 results | Steel process documents often have associated data from the same batch. Too many results increase context stitching burden, while too few may miss critical parameters |
searchMode | mixedRecall | Balances structured field matching and full-text semantic recall, covering multiple document types such as batching lists and composition reports |
usingReRank | Enabled | Improves ranking priority of highly relevant specialized documents and optimizes matching effects for specialized terms |
datasetSearchUsingExtensionQuery | Enabled | Supplements professional abbreviations and context information in process questions, improving recall accuracy |
collectionIds | Bind dedicated process dataset ID | Limits retrieval scope to avoid mixing in irrelevant administrative or office documents |
Three Common Mistakes
- Phenomenon: After uploading Excel-format batching lists and composition reports, retrieval results cannot match specialized parameters in cells. Cause: Table parsing mode for structured Excel data is not enabled, and field information is lost due to processing as plain text only.
- Phenomenon: Retrieval results return a large number of non-process documents, with low recall accuracy. Cause: No reasonable
similaritythreshold is set, or no dedicated dataset ID is bound, resulting in mixing of irrelevant documents. - Phenomenon: For retrieval of multiple documents associated with a heat, results are not grouped by production batch. Cause: Hybrid mode of
searchModeis not used, and only full-text retrieval is employed, preventing precise matching of structured heat number fields.
How to Confirm the Configuration Is Correct
- Upload a batch of dedicated iron and steel smelting and continuous casting process documents, initiate a query containing steel grade and composition parameters, and verify that retrieval results only include documents within the target dataset.
- Adjust the
similaritythreshold, compare the number of recall results under different thresholds, and confirm that it meets the correlation requirements of the business. - Enable
usingReRankanddatasetSearchUsingExtensionQuery, initiate a query containing professional abbreviations, and compare result correlation before and after enabling these settings. - View retrieval logs to confirm that the number of single-recall results does not exceed the number set by
limit, complying with configuration requirements.
The values provided are common starting points and should be measured against the reader's own samples.