This page is written for FastGPT v4.17.0.
What the data for this category looks like
The data for this category comes from MES systems, online inspection equipment, laboratory reports, process parameter libraries, customer quality standards, and historical defect archives. Data is updated per production batch, during shift rotation, and via continuous collection. Daily documents include control plans, patrol inspection records, SPC reports, final product inspection reports, nonconforming product review forms, release forms, and quality monthly reports. Fields include inspection item number, defect type, judgment result, process parameter value, associated batch number, and customer order number.
What constraints these characteristics impose on model access and configuration
Data sources for this category are scattered and updated in real time. Model access configuration must support synchronous pulling of multi-source data and appropriate context truncation to avoid redundant data interfering with generation. Document formats vary, including structured reports and unstructured review records. Configuration must adapt to text parsing logic for different formats to ensure key fields are accurately extracted. Bid documents must be associated with historical defect archives and customer quality standards. Configuration must reserve sufficient reference token quota to ensure content traceability. The quality judgment link requires rigor, so the randomness of generated results must be controlled to avoid misjudgments.
How to set the configuration
| Configuration Item | Recommended Setting Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the context length of multi-source quality documents for this category, preventing key inspection records and customer standards from being truncated |
maxTokens | 2000–3000 token | Meets complete generation requirements for quotation items and inspection items in bid documents |
quoteMaxToken | 1500–2000 token | Ensures complete display of reference content such as historical defect reports and customer standards |
temperature | 0.1–0.3 | Controls the rigor of quality judgment and document generation, reducing random errors |
AIPROXY_API_ENDPOINT | Fill in the actual deployed proxy address | Adapts to proxy configuration for multi-source data access, ensuring normal synchronization of data from different systems |
| Workflow/API tool selection | auto | Allows the model to call tools as needed to process structured data, improving the accuracy of inspection item conversion |
Three common misconfigurations
- When calling non-official large models such as DeepSeek-R1, the interface returns
{"object":"error","message":"Only allowed now"}. This occurs because custom model access permissions have not been enabled, or the proxy configuration has not correctly bound the allowed model list. - When adding a locally deployed vLLM service to OneAPI, a connection refused error is prompted. This occurs because the proxy address is not configured with the correct port, or a local firewall blocks the communication port.
- The inspection item fields are empty in the bid document generated after importing quality documents. This occurs because the
maxContextquota is not set sufficiently, causing inspection item parameters in the original documents to be truncated and lost.
How to verify successful configuration
- Upload a single final product inspection report. Check whether inspection items and judgment results are accurately extracted in the generated bid document, and confirm that the context truncation logic meets expectations.
- Configure the custom model access address, then initiate a test call. Check whether the returned results include expected quality judgment content, and verify that the proxy configuration takes effect.
- Import historical defect archives and customer standard documents. Check whether generated quotation items are associated with corresponding quality requirements, and confirm that the reference token configuration is reasonable.
- View system logs. Confirm that there are no timeouts or connection errors when pulling multi-source data, and verify the stability of the access configuration.
The values provided are common starting points and should be measured against the reader's own samples.