What data looks like for this category
The data for this category comes from ERP material master data, melting furnaces, spectrometers, MES heat records, mold management systems, and customer drawing technical specifications. Update frequency is per heat, per batch, or daily. Daily documents include foundry process cards, batching lists, heat records, mold histories, first article inspection records, material reports, and nonconforming product review forms. Data fields include alloy batching parameters, furnace temperature values, pouring temperature, mold cycle numbers, defect types, and others. Units include degrees Celsius, kilograms, millimeters, and others.
What constraints these characteristics impose on vector models and indexes
The high-frequency updates per heat require indexes to support incremental synchronization, avoiding full rebuilds that consume computing resources. The field differences in multi-source heterogeneous data require vector models to adapt to mixed input of structured numerical values and unstructured text. Documents contain numerous associations between process parameters and defects, so indexes must accurately match semantic associations between parameter thresholds and defect types. The batch update frequency of data requires index batch processing capabilities to match daily or per-batch update volumes, preventing index delays from reducing cross-review efficiency. Additionally, casting quality traceability requires linking heat records, mold cycle numbers, and inspection records, so indexes must retain association relationships between multiple documents to avoid information silos.
How to set up configurations
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-chunk content length of process cards and heat records in foundry documentation, avoiding semantic fragmentation within chunks |
| Overlap Length | 50–80 characters | Retains cross-chunk process parameter association information, preventing parameter splitting that breaks matching |
| Similarity Threshold | 0.75–0.85 | Matches semantic similarity for associations between alloy batching, furnace temperature, and other values, filtering invalid matching results |
| Single Search Limit | Top 10 results | Covers associated documents for the same heat and same mold cycle, meeting the multi-document verification needs of cross-review |
| maxToken | 8192 | Adapts to vector encoding requirements for long documents, covering complete content of process cards and inspection reports |
| defaultToken | 1024 | Adapts to the base vector encoding length for short text, meeting the needs of short documents such as batching lists |
| Reranking | Enabled | Optimizes matching ranking for defect types and process parameters, improving the accuracy of cross-review |
Three common mistakes
- A 401 error is returned when calling the vector model. The cause is incorrect configuration of the vector model API key, or the key does not have permissions for the corresponding vector service.
- Duplicate documents remain after knowledge base index merging. The cause is that document deduplication configuration is not enabled, or the deduplication threshold is set too high to recognize heat records and inspection forms with similar content.
- The vector model deployed via Ollama is rejected when called in FastGPT, while the curl test works normally. The cause is that the FastGPT vector model call timeout setting is too short, or the proxy configuration does not correctly map the Ollama local service port.
How to confirm the configuration is correct
- Upload a single heat record document, check the vector encoding completion status, and confirm there are no timeout or encoding failure logs.
- Enter a query statement for alloy batching ratios, verify that the similarity scores of returned results fall within the preset threshold range.
- Import multiple associated documents for the same heat, check whether the index merging process retains the association relationships between documents.
- Switch to different vector models, verify that calls work normally, and confirm that API keys and configuration parameters are correct.
The values given are common starting points and should be measured against the reader's own samples.