What this category of data looks like
Data sources cover hot metal and scrap charging data, L2 process control real-time data, laboratory composition reports, MES production scheduling data, energy medium system data, and equipment condition monitoring data. Daily documents include heat plans, batching lists, converter smelting records, continuous casting process records, slab judgment sheets, and similar items. Fields include heat number, cast number, molten steel temperature, element composition, equipment number, inspection time, and more. Update cycles cover smelting records generated per heat, real-time updated equipment status data, and slab judgment sheets generated per shift. Most data structures use structured fields, and some records include multi-dimensional associated information.
Constraints for Knowledge Base Retrieval and Recall
Multi-source, multi-update-cycle data requires the retrieval system to support incremental synchronization and filtering by data source. This prevents matching conflicts between offline documents and real-time data. Strongly coupled associations between heats and casts require precise filtering using associated fields such as heat number and cast number during retrieval. This stops the system from recalling documents from unrelated heats. High-temperature continuous production requires that the retrieval process does not occupy production system resources. Response speed must match the real-time needs of on-site scheduling. There are many structured fields with high precision requirements. Low-match redundant results must be avoided to interfere with equipment inspection and maintenance decision-making. Recalled result fields must also match the business scenario.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.82 | Business documents for iron and steel smelting and continuous casting have high field precision requirements. Low-match redundant records must be filtered to avoid interfering with decision-making references for equipment inspection and maintenance |
limit | Top 8–12 entries | The number of associated documents for a single heat or continuous casting cast is limited. Excessive recall will exceed the model context window and affect output accuracy |
usingReRank | Enabled | Equipment inspection and maintenance questions often involve multi-field associations. Reranking can improve the sorting priority of high-match documents |
datasetSearchUsingExtensionQuery | Enabled | On-site inspection personnel’s questions are often in colloquial language. Query optimization can convert colloquial expressions into keywords that match structured documents |
searchMode | mixedRecall | The query examples combine domain terms with natural-language context; compare with representative data, not as a product default. |
collectionIds | Specify by data source group | Distinguish different business data sets such as converter smelting records and continuous casting equipment inspection records, to avoid cross-category data interfering with retrieval results |
Three Common Configuration Errors
- Phenomenon: The knowledge base retrieval step returns matching results, but the final generated answer does not include the retrieved content. Cause: The
usingReRankorlimitparameters are not configured correctly, causing recalled documents to exceed the model context window, or high-match documents to be filtered out after reranking. - Phenomenon: Retrieval works normally during local testing, but no knowledge base content is returned when calling via the published channel. Cause: The associated knowledge base collection ID is not configured in the published channel, or the
collectionIdsparameter does not include the ID of the corresponding business data set. - Phenomenon: The retrieved results include paragraph internal ID fields. Cause: The paragraph ID output switch is not turned off, or non-business field metadata is not excluded during configuration.
How to Verify Proper Configuration
- Manually enter typical equipment inspection and maintenance questions, such as "Troubleshooting steps for continuous casting mold water leakage". Check whether the retrieved documents include continuous casting process records and equipment inspection records, and whether the match degree falls within the threshold range set by
similarity. - Switch to the external published channel of FastGPT v4.17.0, submit the same inspection question, and check whether the returned answer includes the retrieved business document content, which is consistent with the local test results.
- View the retrieval log, check whether the number of recalled documents is within the range set by
limit, and that no irrelevant paragraph ID fields are included. - Simulate adding a new real-time equipment status record, check whether the knowledge base can synchronously recall the new document, to verify the effectiveness of the incremental update configuration.
The values provided are common starting points and should be measured against the reader's own samples.