What Data Looks Like for This Category
Core data for this category originates from quality and process teams at foundries, die casting workshops, and equipment component suppliers. It includes casting process cards, batching lists, heat records, mold histories, first-piece inspection records, material test reports, nonconforming product review forms, and similar documents. Data sources include ERP material master data, melting furnace thermometers, spectrometers, MES heat record systems, mold management systems, and customer technical drawings. Updates occur per heat, per batch, or daily. Fields cover alloy ratios, temperature parameters, composition test results, pouring parameters, defect types, and additional metrics. Some data is structured values transmitted directly from equipment. Other data is unstructured text filled out manually.
Constraints Imposed on Vector Models and Indexes by These Data Characteristics
High-frequency data updates require indexes to support incremental synchronization. This avoids performance losses from full index rebuilding. Mixed multi-source data formats (equipment-transmitted values, manually filled text) require vector models to handle semantic encoding for both structured fields and unstructured text. Strongly associated attributes of single-heat data, such as binding between a heat and its corresponding mold or batching list, require retrieval results to accurately link matching identifiers. This prevents cross-batch confusion. Defect traceability business demands require retrieval results to cover all fully associated documents. This ensures complete mold and heat information for internal defects. Additionally, high data sensitivity requires index construction to prevent data leaks. Retrieval result accuracy must meet strict quality audit standards.
Configuration Guidelines
This guide is compatible with FastGPT v4.17.0. Specific configuration items and recommended values are listed below:
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the single-paragraph text length of casting process cards and heat records. Prevents splitting continuous critical parameters such as batching and temperature values. |
| Overlap Length | 100–150 characters | Covers continuous temperature curve and batching ratio data in heat records. Prevents truncation of associated information. |
maxToken | 4096 token | Adapts to the input limit of mainstream embedding models. Supports text input requirements for batch heat records. |
| Similarity Threshold | 0.75–0.85, calibrated via actual testing | Distinguishes semantic differences between qualified and unqualified inspection records. Avoids mismatching low-correlation historical data. |
| Single Search Limit | Top 10 results | Covers all associated heat and mold data for a single casting batch. Meets retrieval requirements for defect traceability. |
| Reranking | Enabled | Corrects semantic bias in basic retrieval. Accurately matches defect types with corresponding mold and heat information. |
Three Common Misconfiguration Issues
- Phenomenon: After calling the create training order interface, no new casting documents are synced to the knowledge base collection. Cause: The training order is not bound to the target knowledge base collection, resulting in data not being correctly written to the index.
- Phenomenon: When retrieving inspection records, associated spectrometer test images and mold appearance photos cannot be returned. Cause: Image vectorization configuration is not enabled, and image embedding vectors are not extracted during index construction.
- Phenomenon: The specified embedding model cannot be selected from the platform embedding model list. Cause: The access key and interface address for the corresponding model are not entered in the platform configuration file, resulting in the model not being loaded correctly.
How to Verify Successful Configuration
- Upload a single heat record document to trigger index construction. Check the platform index progress log to confirm no format errors or data loss prompts are present.
- Enter a query statement containing a specific heat number and defect type. Verify returned document types and field matching levels meet business expectations.
- Call the create training order interface. Confirm the interface returns a completed order status, and the document count in the corresponding knowledge base collection matches the number of uploaded documents.
- Upload a spectrometer test image and perform a retrieval action. Confirm returned results include associated inspection record text and image metadata.
The values listed above are common starting points. They should be validated against the samples used in actual deployment.