What data for this category looks like
This category covers all process quality-related information from work-in-progress to finished products. Sources include MES systems, online inspection equipment, laboratory test records, process parameter logs, customer quality standards, historical defect libraries, and more. Daily generated documents include control plans, patrol inspection records, SPC reports, finished product inspection reports, nonconforming product review forms, release forms, quality monthly reports, and other materials. Data updates per batch or shift. Some parameters are collected continuously. Fields include batch number, inspection item name, defect type, judgment result, process parameter value, detection time, and additional relevant fields. Structure varies based on document type.
Constraints imposed on model access and configuration
Data sources for this category are scattered. Update frequencies are inconsistent. Document structures vary. These factors create multiple constraints for model access and configuration. Multi-source data requires separate access path configuration to ensure unified parsing of fields from different systems. Data updated per batch or shift requires index refresh strategies that match the update rhythm, to avoid data lag. Diverse document structures require models to support unstructured text parsing and field extraction. Context length must be limited to prevent call timeouts caused by loading excessive batch data. Quality analysis needs to link historical defects with current batch parameters. This requires models to accurately recall relevant source data and avoid interference from irrelevant information.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | This category requires linking multiple inspection reports, process parameters, and historical defect data. Context length must cover the full batch analysis scope |
AIPROXY_API_TOKEN | Valid key bound to the corresponding large model | Used to verify model call permissions, ensuring interface requests are legal and valid |
temperature | 0.1–0.3 | Quality problem traceability and analysis require rigorous conclusions, reducing random output deviation |
maxTokens | 2000 token | Defect analysis output needs to include defect causes, impact scope, and rectification suggestions. Limiting output length avoids redundancy |
quoteMaxToken | 1500 characters | Specific inspection records and process parameters must be cited as traceability basis. Limiting quoted text length ensures accurate information |
| Workflow/API tool selection | auto | Allows the model to independently judge whether to call retrieval tools to obtain historical defect library data, improving analysis accuracy |
Three common misconfigurations
- Configuring
AIPROXY_API_TOKENresults in a "no available channel" prompt: Call permissions for the corresponding model are not enabled in the system group permissions, or the model bound to the key does not match the actual accessed model type. - Inconsistent model display between chat and workspace: Global model configuration and application-level model configuration are not synchronized, or browser cache has not been cleared, causing configuration changes to not take effect in real time.
- Excessively long search response time and high hard disk IO usage: The value range of
maxContextis not reasonably limited, causing excessive historical batch data to be loaded during each retrieval, or indexes are not sharded by batch dimension, increasing disk read pressure.
How to confirm successful configuration
- Access the model management page in FastGPT v4.17.0. Check if
AIPROXY_API_TOKENis correctly filled and the corresponding model is in an enabled state. - Initiate a simulated defect analysis request. Verify that returned results include cited source data fragments such as inspection records and process parameters.
- Review system operation logs to confirm there are no model call failure error messages, and that response times meet preset business requirements.
- On the application configuration page, confirm that the model bound to the current application fully matches the model parameters set in the global configuration.
The values provided here are common starting points. They should be measured against the reader's own samples.