What the Data for This Category Looks Like
Core data for this category comes from ERP material master data, melting furnaces and thermometers, spectrometers, MES melt batch records, mold management systems, and customer drawings and technical specifications. Daily documents include casting process cards, batching lists, melt batch records, mold histories, first article inspection records, material test reports, and nonconforming product review forms. Data is updated per melt batch, per batch, or daily. Fields include alloy batching parameters, melting temperature, melt batch identifiers, mold numbers, customer technical requirement items, and more. The data covers two formats: structured business documents and unstructured process documents.
Constraints Imposed on Vector Models and Indexes by These Characteristics
Data updates are frequent and triggered per melt batch. Indexes must support incremental updates to avoid resource consumption from full reconstruction, to meet the real-time requirements of foundry production. Document types include structured documents and unstructured process content. This requires adapting text chunking rules for different lengths, to ensure key business identifiers are not split and lost. The matching accuracy between customer technical requirements and internal process parameters directly affects the accuracy of bid quotations. Matching rules for vector recall must be optimized for specific fields such as alloy batching and melting temperature, to avoid irrelevant data interfering with retrieval results.
How to Set Configurations
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the typical paragraph length of casting process cards and batching lists, prevents key process parameters from being split |
| Overlap Length | 100–150 characters | Preserves process-related information across paragraphs, prevents loss of the correspondence between melt batches and molds across chunks |
| Similarity Threshold | 0.75–0.85 | Matches the precision requirements for casting process parameters, filters low-correlation historical melt batch data |
| Single Search Limit | Top 10 entries | Covers the typical process reference range required for bids, avoids excessive redundant data interfering with quotation calculations |
defaultToken | Calibrate with representative process documents | This is the default token count used for direct text splitting; inspect sample chunks. See the v4.17.2 model-field reference |
maxToken | Set to the selected embedding model/service input limit | This is the configured embedding model's maximum token count, not retrieval or LLM context. Check the selected model or provider documentation. See the v4.17.2 model-field reference |
Three Common Mistakes
- A 503 status code appears, prompting that no available text-embedding-v3 model channel is available for the current group. Cause: The corresponding call channel for the text-embedding-v3 model has not been configured, or the channel quota has been exhausted.
- An API key verification failure prompt appears after starting the FastGPT container. Cause: The CHATAPIKEY environment variable has not been correctly configured in docker-compose.yml, or the container has not been rebuilt for the configuration to take effect.
- Index parameters cannot be adjusted directly through the retrieval test interface. Modifications must be made in the content management page of the corresponding knowledge base. Cause: The platform has removed the direct edit button in the retrieval test interface, and index parameter adjustments must be completed through the knowledge base configuration entry.
How to Confirm Configurations Are Correct
- Upload a casting process card and batching list, run a retrieval test, check whether the recall results include matching melt batch parameters and mold information.
- Check the vector index update log, confirm that newly uploaded melt batch records complete index synchronization within the specified time.
- Check text splitting with a representative process document to validate
defaultToken, then comparemaxTokenwith the selected embedding model/service input limit. - Enter keywords from customer technical requirements, confirm that the recall results include corresponding material test reports and nonconforming product review forms.
The values provided here are common starting points, and should be evaluated against the reader's own samples.