What Data for This Category Looks Like
Data sources for this category include basic hot metal and scrap data, real-time L2 process control data, laboratory composition test reports, MES production management data, energy medium system parameters, and equipment status records. Daily documents include heat plans, batching lists, converter smelting records, continuous casting process records, and slab judgment sheets, among others. Data update frequencies cover per-heat generation, real-time synchronization, and shift summary updates. Fields include heat identifiers, process parameters, composition test items, and equipment operating status, among others. Units are adjusted based on corresponding business scenarios.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
The heat-bound update feature of the data requires retrieval and recall to accurately match the context of the corresponding heat or casting sequence, to avoid cross-heat data confusion. Multi-source heterogeneous document and data formats, including structured reports, process logs, and test reports, require retrieval modes to support mixed-type content recall. Batch data summarized per shift requires limiting the retrieval scope, to avoid redundant information interfering with the accuracy of bid documents and quotations. The strong correlation between composition indicators and process parameters requires retrieval to prioritize recalling historical data from the same process scenario, improving matching accuracy. The high-frequency update of real-time data requires the knowledge base index to support incremental synchronization, to avoid retrieving expired data.
Configuration Settings
These configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the strong correlation characteristic of iron and steel smelting process parameters, filtering low-match irrelevant data |
limit | Top 6 | Balances the content comprehensiveness and length limits of bid documents, avoiding excessive redundant information |
| Chunk length | 800–1200 characters | Adapts to the length characteristics of single documents such as smelting records and composition reports |
| Retrieval timeout | 30 seconds | Adapts to the retrieval response requirements of real-time process data, avoiding timeout interruptions |
| Maximum context length | 4000 token | Adapts to the overall length requirements of bid documents |
usingReRank | Enabled | Improves the relevance of retrieval results for high-semantic-complexity process data |
datasetSearchUsingExtensionQuery | Enabled | Covers multiple expression forms of process terms, improving recall coverage |
collectionIds | Bound to the knowledge base collection of the smelting and continuous casting business domain | Limits the retrieval scope to exclusive documents, avoiding interference from cross-category data |
Three Common Misconfigurations
- Formulas in retrieval results display as garbled text or plain text fragments. This occurs when formula parsing adaptation is not enabled during knowledge base document import, causing mathematical expressions to fail correct recognition and rendering.
- Returned results are inconsistent after multiple retrievals with the same query. This occurs when the retrieval and recall result scope is not fixed, or the locally deployed model does not have deterministic generation configuration enabled.
- Retrieval and recall covers all business table data without limiting the valid scope. This occurs when
collectionIdsis not configured to bind the exclusive knowledge base collection, resulting in cross-category or redundant data being included in retrieval results.
How to Verify Proper Configuration
- Perform a single retrieval, check that the number of returned results matches the value set for the
limitconfiguration. - View the similarity scores of retrieval results, confirm they fall within the range set for the
similarityconfiguration. - Import a test document containing process formulas, perform a retrieval, confirm that formulas display correctly.
- Perform the same query multiple times, confirm that the top entries in returned results do not change significantly, verifying retrieval and generation determinism.
The values given are common starting points and should be measured against the reader's own samples.