Data Profile for This Category
This category covers multi-source heterogeneous information across the entire foundry process. Data sources include ERP material master data, melting furnaces and thermometers, spectrometers, MES heat records, mold management systems, customer drawings, and 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. Fields covered include alloy composition ratio, melting temperature, pouring parameters, heat number, mold number, defect type, and more. Update frequency is per heat, per batch, or daily.
Constraints on Vector Models and Indexing
Multi-source heterogeneous data structures require indexes to support vector conversion of mixed data types. This includes structured batching parameters, semi-structured heat records, and unstructured defect review documents. High-frequency update cycles require indexes to support incremental updates, avoiding performance loss from full reindexing. Professional field terminology, such as alloy composition and melting temperature, requires alignment with foundry industry terminology systems. Generic vector models may fail to accurately capture professional semantics. Strong data relationships exist, such as the binding between heat numbers, corresponding molds, and defects. Flat entity indexes cannot fully represent these business relationships.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Aligns with the average length of foundry heat records, process cards, and other documents, and matches the compatible range of maxToken |
| Overlap Length | 150–200 characters | Preserves critical cross-chunk relational information, such as heat numbers and corresponding defect descriptions |
maxToken | 4096 | Complies with the default compatible limit of FastGPT v4.17.0, avoiding timeout during single-batch vector generation |
defaultToken | 150 | Fits the token count of key fields for single structured parameters such as alloy batching lists, ensuring semantic completeness |
| Similarity Threshold | 0.72–0.80 | Balances semantic matching accuracy and recall rate for foundry professional parameters, filtering irrelevant general documents |
| Single Search Limit | Top 8 results | Matches the typical number of associated molds and inspection records per heat, ensuring relevance of search results |
Three Common Misconfigurations
- Issue: Results from different vector models cannot be consistently matched, and search results have semantic deviations. Cause: Vector models with inconsistent semantic dimensions are mixed, and no model adapted to foundry professional text is used uniformly.
- Issue: No results are returned when searching for the relationship between heat numbers and defects. Cause: Only flat entity indexes are used, and no entity relationship index is configured. Business binding relationships between heat numbers, mold numbers, and defects cannot be represented.
- Issue: Batching list fields are empty after parsing PDF-format heat records. Cause: Table extraction for enhanced PDF parsing is not enabled, so structured batching data cannot be correctly extracted and converted into vectors.
How to Verify Proper Configuration
- Upload a single PDF heat record and structured batching data, check if the vectors generated by the index include key business fields such as alloy composition and heat number.
- Enter a query containing professional terminology, such as "The shrinkage defect of a certain heat corresponds to melting parameters", check if search results cover associated mold histories, inspection records, and other data.
- View index task run logs, confirm that incremental update trigger frequency matches the business update cycle per heat.
- Test multiple queries with different semantics, confirm that search result ranking aligns with foundry business relationship logic.
The values provided are common starting points and should be measured against the reader's own samples.