What This Category’s Data Looks Like
The primary entities covered by this category are process and quality departments of foundries, die casting workshops, and equipment component suppliers. Daily documents include casting process cards, batching lists, heat records, mold history logs, first article inspection records, material test reports, and nonconforming product review forms. Data sources cover ERP material master data, melting furnaces and thermometers, spectrometers, MES heat records, mold management systems, and customer drawings and technical specifications. Update frequency is per heat, per batch, or daily. Documents contain professional fields and their corresponding units, such as alloy batching ratios, pouring temperatures, and defect types.
What Constraints Do These Characteristics Impose on the Knowledge Base Retrieval and Recall Link?
High-frequency updated documents require retrieval to support incremental synchronization, to avoid performance loss caused by full reindexing. Documents contain a large number of professional parameters with units and precise numerical values. Matching must cover both semantics and field accuracy to avoid false recalls caused by fuzzy matching. Multiple heterogeneous data sources lead to documents being scattered across different collections. Retrieval scope must be precisely limited to prevent irrelevant documents from mixing into results. Retrieval results directly affect process judgment and quality control. The accuracy and relevance of recalled content must be ensured to avoid incorrect parameters being introduced into subsequent workflows.
How to Set the Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the precision requirements for matching casting professional terminology and precise parameters, filters low-relevance non-process documents |
limit | Top 8–12 entries | Covers the number of documents associated with a single heat process, avoids large model processing pressure caused by overly long context |
searchMode | mixedRecall | Adapts to document matching needs that include both professional terminology and precise numerical values, improves recall accuracy for batching lists and heat records |
usingReRank | Enabled | Performs secondary sorting on recall results, filters mismatches in defect descriptions such as shrinkage cavities and porosity, improves result relevance |
datasetSearchUsingExtensionQuery | Enabled | Expands the associated retrieval scope for batching lists and heat records, covers cross-document process-related content |
collectionIds | Bind dedicated collections for foundry processes, molds, and quality inspection | Isolates non-foundry documents, reduces interference from irrelevant recalls |
Three Common Misconfigurations
- Retrieval response time exceeds the preset threshold. The cause is that
collectionIdsis not configured, and a large volume of non-foundry documents is recalled, leading to excessive computational load for vector retrieval and reranking. - Unexpected large model call error logs appear in the workflow. The cause is that after enabling
datasetSearchUsingExtensionQuery, the context length of retrieval results is not limited, triggering additional large model call quota consumption. - The preset directory structure of uploaded process cards is automatically deleted. The cause is that the default automatic document content formatting configuration is enabled, and this option is not disabled, leading to accidental directory cleanup.
How to Verify the Configuration Is Correct
- Upload a casting process card with a preset directory, check that the directory structure is not automatically modified after upload.
- Initiate a search for alloy batching ratios, check that the recall results include corresponding batching lists, heat records, and material test reports.
- View the retrieval configuration panel in FastGPT v4.17.0, confirm that
usingReRankanddatasetSearchUsingExtensionQueryare in an enabled state. - Simulate a retrieval trigger, check that the response time meets expected standards, and there are no unexpected large model call logs.
Note: The values listed are common starting points, and should be validated against respective samples.