What the data for this category looks like
Data for this category comes from ERP material master data, melting furnace thermometers, spectrometers, MES heat records, mold management systems, and customer drawing technical specifications. Daily documents include foundry process cards, batching lists, heat records, mold histories, first article inspection records, material reports, and nonconforming product review forms. Data updates are generated on a per-heat, per-batch, or daily schedule. Each individual document includes fields such as alloy batching ratio, melting temperature, pouring parameters, and defect types, plus quantified values with units such as weight, temperature, and component proportion.
What constraints do these characteristics impose on knowledge base retrieval and recall
The high-frequency updated heat and batch data for this category requires knowledge base retrieval to support real-time or near-real-time incremental recall, to avoid using outdated process parameters. Multiple types of structured documents, such as multi-column batching lists and parameterized process cards, have fixed fields and numerical units. Retrieval must accurately match field dimensions; matching only text will lead to association errors. Format differences across multi-source heterogeneous data require the retrieval link to unify data standardization processing logic, to ensure documents exported from different systems can be recalled consistently. Additionally, defect tracing requires association with unique identifiers such as heat and mold times, so recalled results must carry clear associated field information.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches high-precision numerical queries such as foundry batching ratios and temperatures, filters low-relevance results |
limit | Top 8 entries | Covers multiple relevant documents per heat, including process cards, heat records, and mold histories |
searchMode | mixedRecall | Balances structured field matching and semantic understanding, adapts to multiple types of foundry documents |
usingReRank | Enabled | Re-ranks initial recall results, improves accuracy of defect tracing and parameter matching |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands query terms, supplements industry terminology such as alloy grades and defect types, adapts to professional documents |
collectionIds | Specified by foundry business domain | Isolates knowledge bases of other categories, prevents recall of unrelated forging or machining documents |
Three common mistakes
- Phenomenon: Multi-column batching lists are split into scattered fragments after automatic segmentation, and complete row-level alloy batching ratio data cannot be retained. Cause: No row-wise segmentation rule is configured for structured table documents, and the default automatic segmentation logic does not adapt to the structural characteristics of multi-column tables.
- Phenomenon: An undefined model error prompt appears when creating a knowledge base with an embedding model other than ada-002. Cause: The embedding model verification rule is not adjusted in the system configuration; the default verification only supports ada-002 models.
- Phenomenon: Forging or machining documents are mixed in retrieval results, interfering with parameter matching and defect tracing. Cause: Business domains are not isolated via collectionIds, leading to recall of cross-category knowledge base data.
How to confirm the configuration is complete
- Upload a test multi-column batching list to trigger segmentation verification, confirm that each row of data is treated as an independent segmentation unit.
- Enter a query containing an alloy grade and temperature parameters, check the similarity scores of retrieval results, confirm they fall within the preset matching range.
- After enabling
usingReRank, compare the ranking of retrieval results before and after re-ranking, confirm highly relevant documents are displayed first. - Check the associated fields of retrieval results, confirm unique identifiers such as heat and mold times are fully carried.
The values provided here are common starting points, and should be evaluated against the reader's own samples.