What data for this category looks like
Data comes from slab or billet batches, heating furnaces, mill automation control systems, online thickness and width measuring equipment, surface inspection devices, and MES systems. Update frequency is real-time generation per coil or per individual steel product. Daily documents include rolling plans, heating schedules, rolling specifications, flatness and thickness records, surface quality reports, and quality certificates. Each individual document includes fields such as batch number, process parameters, test values, and defect types. Units include millimeters, degrees Celsius, and special I units for flatness, among others.
What constraints do these characteristics impose on knowledge base retrieval and recall?
The real-time, small-batch high-frequency update data feature requires knowledge base retrieval to support precise association of historical process and test data by coil or steel batch, to avoid invalid recall across batches. The presence of professional fields and units requires that original parameter expressions be retained during retrieval to match business terminology, without arbitrary normalization. The core requirement of coil-level traceability requires that recall results be associated with a complete process chain per batch, covering historical data from multiple links. High-frequency real-time data also requires the retrieval link to have low latency, to support immediate response for on-site defect judgment.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Steel rolling process parameters and defect descriptions have strong professional relevance. This range filters out irrelevant general process documents and retains highly matching professional content |
limit | top 8 | Coil-level traceability requires associating multi-dimensional data (heating, rolling, inspection). 8 entries can cover core documents for a typical process chain, avoiding context overload |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
usingReRank | enabled | Semantic similarity of professional terms requires secondary calibration. Re-ranking improves matching accuracy for defect types and process causes |
datasetSearchUsingExtensionQuery | enabled | Can automatically expand related process terms based on user defect descriptions, supplementing missing professional association information |
collectionIds | bind three dedicated datasets: rolling process, surface inspection, quality certificate | Precise restriction of retrieval scope, avoiding mixing irrelevant documents from other categories |
Three common misconfigurations
- Phenomenon: Garbled text appears in retrieved document fragments. Cause: Uploaded documents such as rolling plans and surface quality reports contain professional drawing symbols or non-standard encodings, and have not been converted to standard text using FastGPT's built-in document preprocessing tool.
- Phenomenon: After calling MySQL to obtain batch data, the large language model's answer does not reference database original text fragments. Cause: The associated recall logic is not configured correctly, and database query results are not synchronized to knowledge base fragments that can be referenced by the large language model.
- Phenomenon: The number of retrieval results is far lower than expected, with only a small amount of irrelevant content returned. Cause: The
similaritythreshold is set too high, filtering out process adjustment records with slightly lower matching but critical importance, or thelimitparameter is set too low, failing to cover the data volume required for a complete process chain.
How to confirm the configuration is properly set
- Upload one rolling specification document, run a retrieval test, and check if the returned results include core parameters such as heating schedule and reduction rate in the document.
- Enter a known steel coil batch number, and verify after retrieval whether the returned results are associated with all historical test data and process records for this batch.
- After enabling
usingReRank, compare defect type matching results before and after enabling to confirm that the re-ranking function is active. - Call MySQL to query surface defect data for a specified batch, and verify whether the large language model's answer references the original text fragment content returned by the query.
The values provided here are common starting points and should be evaluated against the reader's own samples.