What this category of data looks like
Data for iron and steel smelting declaration materials and case review mainly comes from raw material quality inspection ledgers, smelting process parameter logs, finished product inspection reports, compliance declaration documents, and approval case files. Data updates follow the declaration cycle, with batch updates per single project or on a monthly/quarterly basis. Document types include structured Excel tables with fields such as carbon content, furnace temperature, and batch number, using units like %, ℃, and tons. They also include PDF-format process specifications and Word-format approval forms. Some cases contain associated cross-document data, such as the correspondence between raw materials and finished products for the same batch.
What constraints these characteristics impose on knowledge base retrieval and recall
Mixed structured Excel tables and unstructured text require the retrieval system to support both exact field matching and semantic retrieval. Professional terms such as furnace number and alloy composition have high semantic consistency requirements, to avoid low-quality false recalls. The existence of batch-related data requires retrieval to recall multiple associated documents for the same batch. The batch update feature requires knowledge base index updates to support batch synchronization, avoiding resource consumption from full reindexing.
How to Set the Configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | There are many professional terms in iron and steel smelting. A threshold that is too low will mix in irrelevant process documents, while a threshold that is too high will miss qualified matching results. Calibrated based on actual measurements |
limit | Top 8–12 results | Case review requires association of multiple materials from the same batch. Too many results will exceed the context window, while too few will fail to cover associated documents |
searchMode | hybrid | Balances exact field matching for structured tables and semantic retrieval for unstructured text, adapting to mixed-format business data |
usingReRank | Enabled | Semantic similarity of professional iron and steel terms is easily misjudged by base models. Reranking filters low-quality recall results |
datasetSearchUsingExtensionQuery | Enabled | Allows query expansion via fields such as batch number and furnace number to recall associated documents for the same batch, improving retrieval completeness |
collectionIds | Bind the knowledge base collection ID for the current business | Limits the retrieval scope to avoid mixing in documents from other categories, improving retrieval accuracy |
Three Common Misconfigurations
- After importing an XLSX-format raw material quality inspection table, searching for content from a specific batch returns no matching results. The cause is that no field mapping rule for table parsing was configured, so structured data in the Excel file was not split and indexed correctly.
- Selecting an embedding model other than ada 002 when creating a knowledge base triggers the error message
undefined model must match "^(text. The cause is that FastGPT v4.17.0 requires the embedding model to match the specified prefix, and selecting an incompatible model triggers a validation error. - Non-iron and steel smelting documents appear in retrieval recall results. The cause is that the retrieval knowledge base collection was not limited via
collectionIds, leading to recall of cross-business data.
How to Confirm the Configuration is Correct
- Upload a test XLSX raw material quality inspection table, run a search containing batch number keywords, and verify that the corresponding table content is matched.
- Navigate to the knowledge base configuration page, confirm that the embedding model is set to ada 002, and no model validation-related errors are present.
- Check whether
collectionIdsin the retrieval configuration is bound to the exclusive knowledge base collection for the current iron and steel smelting declaration cases. - After enabling
usingReRank, compare the order of recall results before and after enabling to confirm that low-quality matching results are filtered out.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.