What This Category of Data Looks Like
Data for this category originates from hot metal and scrap steel datasets, L2 process control systems, laboratories, MES, energy medium systems, and equipment status monitoring. Daily documentation includes heat plans, batching lists, converter smelting records, composition reports, continuous casting process records, and slab judgment sheets. Each document includes fields such as heat number, cast number, composition parameters, process parameters, and equipment status codes. Updates occur per heat, in real time, or per shift. Most data is stored as structured tables and text records, with a small volume of equipment alarm logs.
What Constraints Do These Characteristics Impose on Knowledge Base Retrieval and Recall
The high-frequency update requirement (per heat, real-time, or per shift) means the retrieval link must support low-latency recall to avoid disrupting real-time scheduling for steelmaking and continuous casting. Multi-source, heterogeneous document types (structured batching lists, text smelting records, equipment logs) require support for hybrid retrieval modes, balancing semantic matching and exact full-text matching. The tightly coupled business logic of heat and cast numbers means retrieval results must be aggregated by heat or cast number, to prevent cross-batch data from interfering with current process judgments. Errors in high-temperature continuous processes have severe consequences, so retrieval must strictly filter irrelevant content to ensure recalled results strongly align with the current process scenario.
How to Set Configuration
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the precision requirements for iron and steel smelting process parameters, filters low-relevance non-heat-level documents |
limit | Top 8 entries | Covers multiple types of documents associated with a single heat, avoids excessive results exceeding the context window |
searchMode | mixedRecall | Adapts to the hybrid retrieval needs of structured batching lists and text smelting records, balancing semantic and full-text matching |
usingReRank | Enabled | Filters redundant results during the recall phase, focusing on documents strongly related to the current process issue |
datasetSearchUsingExtensionQuery | Enabled | Supplements business keyword extensions such as heat numbers and cast numbers, improving precise recall |
collectionIds | Fill in according to the bound smelting and continuous casting knowledge base collections | Limits the retrieval scope, avoiding interference from cross-category data |
Three Common Misconfigurations
- Phenomenon: Online recall tests return a large number of irrelevant cross-category documents, making it impossible to accurately locate process records for the current heat. Cause:
collectionIdsis not configured, so the retrieval scope covers unrelated knowledge base collections. - Phenomenon: Retrieval latency exceeds 10 seconds, and this issue is noticeable on Tesla V100-SXM2-32GB hardware. Cause: The
limitparameter is not adjusted, and excessive recalled results increase time spent on reranking and context concatenation. - Phenomenon: After enabling
usingReRank, recalled results are not sorted by relevance, and low-relevance entries remain. Cause: Reranking model configuration is not enabled synchronously during knowledge base indexing, or reranking model call parameters are not adapted to the field characteristics of smelting documents.
How to Verify Correct Configuration
- Access the knowledge base retrieval test interface in FastGPT v4.17.0, enter a heat number keyword, and confirm that retrieval results only include documents for the corresponding heat.
- Review retrieval latency logs to confirm compliance with real-time response requirements for the business scenario. Adjust the
limitparameter to optimize based on hardware configuration. - After enabling
usingReRank, compare result rankings before and after recall and reranking, and confirm that highly relevant documents appear first. - Enter a process problem keyword, and confirm that the system automatically supplements business-related extended query terms such as heat numbers and cast numbers.
The values provided are common starting points and should be measured against the reader's own samples.