What the data for this category looks like
Business data for steel rolling and finishing comes from full production workflows of hot rolling, cold rolling, section steel, and steel pipe enterprises. It includes documents such as rolling plans, heating schedules, rolling specifications, plate shape and thickness records, surface quality reports, and quality certificates. Data sources cover slab or billet batches, heating furnaces, mill automatic control systems, online thickness and width measuring equipment, surface inspection systems, and MES systems. Updates run in real time during the production cycle of each coil or individual steel billet. Each single document contains fields like batch number, process parameters, test results, and quality judgments. Units use process-specific measurements such as millimeters, meters per second, and percentages.
What constraints these characteristics impose on knowledge base retrieval and recall
Real-time data that updates per coil or individual billet requires knowledge base synchronization mechanisms to adapt to high-frequency incremental updates. This prevents retrieval of expired process documents. Professional process documents include precise fields with units, such as batch number, reduction rate, and plate shape crown. This requires retrieval to support filtering by specific fields to match process parameters for the corresponding production batch. Individual documents are long and highly specialized. Chunking must balance context integrity and retrieval accuracy, avoiding lost chunks or redundant recalled content. The business requirement for coil-level traceability requires retrieval results to link to the complete production data of the corresponding steel coil. It also requires support for filtering by document folder categories to narrow retrieval scope and improve accuracy.
How to configure the system
This configuration is adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches semantic matching for specialized process documents, avoids recalling unrelated rolling parameter records, and balances accuracy and recall coverage |
limit | Top 8 entries | Covers all relevant process documents and exception reports for a single steel coil, while controlling context length to comply with large model input limits |
searchMode | embedding | Performs semantic retrieval for specialized process parameters, prioritizes matching technical details in documents, and adapts to professional terminology for rolling processes |
usingReRank | Enabled | Re-sorts retrieval results, corrects deviations from semantic matching, and highlights core content related to exception reports and process drafting |
datasetSearchUsingExtensionQuery | Enabled | Expands process-related keywords based on the original query. For example, expand "plate shape abnormality" to "excessive plate shape crown, out-of-spec flatness" to improve recall rate for specialized documents |
collectionIds | Knowledge base collections bound to production batches | Precise retrieval scope limitation, avoids recalling redundant data across batches, and meets the business requirement for coil-level traceability |
Three common configuration errors
- Phenomenon: Unable to filter knowledge base content within a specified folder during retrieval. Retrieval results include process documents from unrelated batches. Cause: The
collectionIdsparameter is not configured to bind the knowledge base collection for the corresponding production batch, so the retrieval scope covers all available documents. - Phenomenon: Retrieval results do not strictly return the original answers from knowledge base question-answer pairs, and extra generated non-preset content appears. Cause: The restricted mode of
datasetSearchUsingExtensionQueryis not properly enabled, or unnecessary generation logic is not turned off, causing the system to deviate from preset knowledge base content. - Phenomenon: Vectorization task errors occur, with logs prompting embedding rate limit exceeded or content lost after text chunking. Cause: The number of concurrent threads is not adjusted based on data update frequency, or chunk length is set beyond the system's processing limit, leading to timeout of batch vectorization requests or truncation of tail content of long documents.
How to confirm correct configuration
- Enter the FastGPT knowledge base management interface, check whether the knowledge base collection bound by
collectionIdsincludes the current production batch's process documents and exception reports. - Initiate a test retrieval, enter keywords that specify a specific steel coil batch and process abnormality, and verify whether returned results only include relevant document content within the bound collection.
- After enabling
usingReRankanddatasetSearchUsingExtensionQuery, verify whether retrieval result sorting prioritizes core keywords related to exception reports. - Check vectorization task logs to confirm that the number of concurrent threads matches the current data update frequency, and no rate limit exceeded errors occur.
The configuration values listed above are common starting points, and should be evaluated against the reader's own samples.