This configuration is based on FastGPT v4.17.0.
What the data for this category looks like
The main entities in steel rolling and finishing are hot rolling, cold rolling, section steel, and steel pipe enterprises. Daily documents include 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 measurement systems, surface inspection systems, and MES systems. Data updates are generated in real time per coil or per steel bundle. Most documents contain structured fields such as batch number, reduction rate, thickness (millimeters), and flatness index. They support full coil-level traceability.
What constraints these characteristics impose on knowledge base retrieval and recall
Real-time or coil-level updated data requires the knowledge base to support incremental synchronization. Full refreshes consume excessive resources, so incremental sync prevents this. Structured process and quality documents require the retrieval system to support both semantic matching and exact matching of structured fields. Without this support, the system cannot accurately recall rolling parameters for a specific batch. The coil-level traceability requirement lets retrieval scope be limited by batch. This prevents interference from irrelevant cross-batch data. Dense professional terminology in document content raises requirements for retrieval semantic accuracy. This stops low-relevance results from entering the review process.
How to set the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirement for rolling process parameters, filters low-relevance documents outside this category |
limit | Top 10 | Balances retrieval efficiency and recall coverage, adapts to the typical number of documents related to a single batch |
usingReRank | Enabled | Rolling documents have dense professional terminology; reranking improves the accuracy of result sorting |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands search terms such as batch numbers and specification models to match structured document fields |
collectionIds | Specify collection IDs by rolling batch | Limits retrieval scope to dedicated documents for the current production batch, eliminates cross-batch interference |
searchMode | mixedRecall | Balances semantic matching for process descriptions and exact matching for structured parameters |
Three common mistakes
- Phenomenon: Retrieval results include chemical composition control documents from steelmaking stages, which are unrelated to rolling processes. Cause:
collectionIdsis not configured to limit retrieval scope to steel rolling and finishing exclusive datasets, leading to cross-category data being recalled. - Phenomenon: The large language model response does not cite original text fragments from retrieved rolling specifications, only providing general conclusions. Cause:
datasetSearchUsingExtensionQueryis not enabled, and original text fragments from retrieval results are not spliced into the context, making it impossible to extract field content from structured documents. - Phenomenon: Uploaded plate shape and thickness record files are identified as garbled text. Cause: Correct parsing and chunking rules are not configured for structured industrial documents, leading to parsing failure of binary table content.
How to confirm the configuration is correct
- Run a retrieval test for a single batch rolling plan, verify that all returned document types are rolling-related, with no cross-category content.
- Enter a search query containing a batch number and process parameters, confirm that retrieval results include document fragments such as heating schedules and plate shape records for the corresponding batch.
- Upload structured industrial documents, verify that there is no garbled text after parsing, and chunked content can be retrieved normally.
- After enabling the reranking function, verify whether the sorting of retrieval results and process relevance meet expectations.
The values provided are common starting points and should be validated against local samples.