What data for this category looks like
This category’s data includes foundry process cards, batching lists, heat records, mold histories, first-piece inspection records, material test reports, and nonconforming product review forms. Data sources cover ERP material master data, melting furnace thermometers, spectrometers, MES heat records, mold management systems, customer drawings, and technical requirements. Some data includes numeric fields for melting temperature and composition testing, plus defect photographs. Data is generated per heat, per batch, or daily. Fields include alloy batching ratios, pouring temperature, defect type, heat/mold identification numbers, and others. Units include temperature (℃), composition percentage, weight (kg), and additional standard units.
What constraints these characteristics impose on model access and configuration workflows
This category has high-volume multi-source data with structured numeric fields and unstructured images. Model access must support both text and visual model capabilities. High-frequency updates per heat or batch require configuration to support real-time or near-real-time batch data processing. Traceability requirements that rely on multiple field associations require models to accurately match identifying tags such as heat and mold numbers. Configurations must adapt to locally deployed open-source models, including Qwen3 series embedding and reranking models. Configurations must include correct API endpoints and authentication parameters. Configurations must also support custom large model access to adapt to platforms such as AiHubMix. Traceability workflows must not be interrupted by data format mismatches or incorrect interface configurations.
How to set configuration values
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
vision | Enabled | This category requires defect photographs and spectrometer temperature images for judgment, so visual model capability must be enabled |
maxContext | 8000–12000 token | The total length of documents associated with a single heat, such as process cards and heat records, is relatively high, so multi-document context splicing must be supported |
temperature | 0.1–0.3 | Consistency of defect judgment and traceability results must be ensured, and random output deviation must be reduced |
AIPROXY_API_ENDPOINT | Locally deployed address (such as http://192.168.1.100:11434/v1) or AiHubMix interface address | Adapts to locally deployed ollama, Qwen3 series models, or AiHubMix platforms, avoiding public network latency and data security risks |
AIPROXY_API_TOKEN | Empty or authentication key generated by the corresponding platform | Some locally deployed models do not require additional authentication, and third-party platforms require the corresponding key to be filled in |
maxTokens | 2000–3000 token | The output length of single-round defect analysis and traceability must be controlled within a reasonable range to avoid redundant results |
Three common configuration errors
- Issue: Defect traceability results returned by the model do not associate heat or mold identification numbers. Cause: The configured context does not include heat/mold number fields, or context parameters are not correctly set.
- Issue: The visual model returns incorrect formatting or no results. Cause: The
visionconfiguration item is not enabled correctly, or the input image resolution and format do not meet model requirements. - Issue: Token consumption statistics do not match actual usage, or context over-limit errors occur. Cause:
maxContextandmaxTokensparameters are not correctly configured, leading to excessively long context splicing that exceeds model limits, or correct token statistics configuration is not enabled.
How to verify successful configuration
- Upload a single heat’s process card, defect images, and heat records, trigger model analysis, and verify that returned results include identifying tags such as heat and mold numbers.
- Review the FastGPT v4.17.0 model call log, confirm that the API request address matches the configured
AIPROXY_API_ENDPOINT, and check for no connection timeouts or authentication errors. - Test different types of defect image inputs, confirm that the visual model returns normal judgment results with no format errors.
- Simulate multi-batch data input, verify that the model response after context splicing covers all process parameters and batching information for the corresponding batch.
The values provided are common starting points and should be evaluated against local deployment samples.