What this category of data looks like
This category of data comes from hot metal and scrap steel ledgers, L2 process control systems, laboratory test results, MES production systems, energy medium and equipment condition monitoring systems. Daily documents include heat plans, batching lists, converter smelting records, composition reports, continuous casting process records, and slab judgment sheets. Fields include unique heat identifiers, process parameters, test results, and judgment conclusions. Update frequencies cover real-time generation per heat, second-level real-time data updates, and shift-level summary statistics.
What constraints do these characteristics impose on the "knowledge base retrieval and recall" link
The high real-time performance and multi-source dispersed nature of this category of data require that the data source synchronization frequency of knowledge base retrieval matches the update rhythm of per heat, real-time, and shift levels, to avoid recalling outdated data. Structured process and test documents require retrieval models to accurately match the semantic logic associated with parameters. Documents generated per single heat have small volume but large batch volume. The context length of a single recall must adapt to the long-link logic of the steel production process, while controlling the number of recalled entries to avoid exceeding the large model context window. In addition, the demand for cross-system data integration requires that the retrieval scope can be accurately limited to a specified data source set.
How to set the configurations
This setup is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Match the semantic association accuracy of steel process parameters and test results, avoid recalling irrelevant cross-heat data |
limit | Top 8 entries | Adapt to the volume of associated documents for a single heat production process, control the context window usage of the large model |
searchMode | mixedRecall | Balance keyword matching for structured process parameters and semantic association recall, adapt to multiple types of documents |
usingReRank | Enabled | Filter low-relevance recall results, improve matching accuracy in steel production scenarios |
datasetSearchUsingExtensionQuery | Enabled | Convert professional terms from steel production into standardized retrieval terms, adapt to scenario-specific expressions |
collectionIds | Bind the IDs of the three collections: converter smelting, continuous casting process, and laboratory testing | Limit the retrieval scope to the exclusive document collection of this category, avoid mixing in irrelevant data |
Three common mistakes
- Phenomenon: Target documents can be retrieved in the knowledge base backend, but the large model does not reference the recalled results when generating responses. Cause:
datasetSearchUsingExtensionQueryis not enabled, or thelimitvalue is set too high leading to context overflow, so the large model cannot load all recalled data. - Phenomenon: Retrieval results include documents from non-this category, such as production records from other industrial categories. Cause: The
collectionIdsconfiguration does not bind exclusive data source collections, so the retrieval scope covers irrelevant knowledge base collections. - Phenomenon: Retrieval results do not match the process requirements of the current heat, and outdated historical data is recalled. Cause: No retrieval rule for filtering by update time is configured, or the data source synchronization frequency is lower than the data update rhythm, leading to recalled expired documents.
How to confirm the configuration is correct
- Enter the knowledge base management interface, check whether
collectionIdsis bound to exclusive data source collections such as converter smelting, continuous casting process, and laboratory testing. - Launch a retrieval test for heat process parameters, check whether the number of recalled results matches the
limitconfiguration value. - Compare the backend retrieval results with the referenced content generated by the large model, confirm that the recalled data is correctly loaded into the context window.
- View the data source synchronization logs, confirm that the synchronization frequency matches the per heat, real-time, or shift-level update rhythm.
The values provided are common starting points and should be measured against the reader's own samples.