What data looks like for this category
Daily data for this category includes foundry process cards, batching lists, heat records, mold histories, first article inspection records, material test reports, and similar documents. Data sources include ERP material master data, melting furnaces and thermometers, spectrometers, MES heat records, mold management systems, and customer drawings. Update frequency follows per heat, per batch, or daily schedules. Heat records and batching lists update in real time with production pace. Mold histories and material test reports are archived per batch. Document structure includes structured fields and unstructured text. Fields cover alloy composition percentage, pouring temperature, heat number, mold cycle, defect type, and similar metrics. Units include percentage, degrees Celsius, and kilograms.
Constraints for Model Access and Configuration
Multi-source, high-frequency updating foundry data requires model access to adapt to real-time synchronization and batch processing needs. Structured batching lists and heat records must maintain field integrity, avoiding fragmentation during chunking. Cross-system associated data such as heat records and mold histories requires a sufficient context window to carry associated information. Quality audit judgment results must maintain consistency, avoiding excessive randomness. Some scenarios require processing image data generated by spectrometers and flaw detectors, so image recognition configuration requirements must be supported. Model calls in internal network environments require resolution of proxy and permission issues.
How to Set Configurations
These configuration guidelines apply to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Carries associated data of heats, molds, and batching lists, avoids truncating cross-document quality audit associated information |
maxTokens | 2000 token | Controls single-round reply length, adapts to structured defect judgment and proportion verification output formats |
quoteMaxToken | 3000 token | Retains sufficient original text fragments, ensures traceability of verification results for alloy batching proportions and temperature parameters |
temperature | 0.1–0.3 | Reduces output randomness, ensures consistency of composition judgment and defect classification results |
AIPROXY_API_ENDPOINT | Fill in according to deployment environment | Adapts to internal network access restrictions, resolves issues with failed external model connections |
AIPROXY_API_TOKEN | Fill in according to authorization key | Ensures permission validity for model calls, avoids permission errors |
Three Common Configuration Mistakes
- Phenomenon: After uploading a structured Excel batching list, the automatic chunking result splits a single heat record into multiple fragments, and cannot fully match the corresponding fields. Cause: No row-wise chunking is set for Excel data, using the default automatic segmentation logic causes field fragmentation.
- Phenomenon: Model calls return connection timeout or
403 Forbiddenerror, unable to complete quality audit tasks. Cause:AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENare not configured correctly, or the internal proxy network does not allow the corresponding access path. - Phenomenon: After uploading a porosity defect image taken by a spectrometer, the model does not recognize the defect features in the image and returns no relevant judgment results. Cause: The
visionconfiguration is not enabled, or image parameters are not passed in the required format.
How to Verify Valid Configuration
- Upload a single heat record Excel file. Verify that chunking results fully retain core fields including heat number, alloy composition, and pouring temperature for each row.
- Initiate a test request for alloy batching proportion verification. Check that the returned result references original text fragments and has no critical information truncation.
- Call the model interface to upload a standard foundry defect image. Confirm that the returned judgment result includes feature descriptions of the corresponding defect.
- Check model call logs. Confirm there are no permission error or connection timeout records, verifying the validity of the proxy configuration.
The values provided are common starting points and should be measured against relevant samples.