What the data for this category looks like
Data sources for PCB manufacturing quality traceability include customer Gerber files, drill files, BOM lists, CAM parsing records, drilling and lamination equipment parameters, bath solution test results, AOI inspection reports, electrical test data, and slice analysis reports. Data is collected and updated by production batch, per shift, or continuously. Documents include production instructions, process parameter records, inspection reports, and other materials. Core fields include the number of graphic layers, hole position coordinates, etching compensation values, continuity test results, and other related fields. Some documents must be linked to multiple production batches to form a complete quality traceability chain.
Constraints imposed on model access and configuration by these characteristics
Multi-source, frequently updated professional data from PCB manufacturing requires model access to support long-context processing for linking multiple production batch records. Diverse document types, ranging from pure parameter tables to graphical inspection reports, require configuring parameters for multimodal input. Fields with dense professional terminology, such as etching compensation and hole copper abnormality, require configuring prompt templates that match industry terminology for the model. The high-frequency data update rhythm requires configuring context refresh parameters to adapt to real-time or batch-level data synchronization. Some inspection reports contain long text content. Appropriate text splitting rules must be configured to adapt to the model's input length limits and avoid call failures.
How to configure parameters
The following configurations are based on FastGPT v4.17.0, and are adapted for domestic-deployed large model access scenarios.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to associated queries for multiple production batch records, covers complete process and inspection data for a single batch |
temperature | 0.2–0.4 | Reduces randomness, ensures consistency of quality traceability results, and meets the rigor requirements of professional analysis |
| Workflow/API tool selection | auto | Allows the model to independently select tools, adapting to the associated analysis needs of multi-source data |
AIPROXY_API_ENDPOINT | Fill in the deployment address of the self-built model, such as http://192.168.1.100:3000/v1 | Adapts to domestically deployed models, replaces the official OpenAI endpoint, and supports domestic models accessed via OneAPI |
AIPROXY_API_TOKEN | Access key for the self-built model | Used for identity verification, adapting to the access requirements of domestic models |
quoteMaxToken | 1500–2000 token | Limits the text length of a single reference, avoids exceeding the model input limits, and adapts to long text content from AOI and electrical test reports |
Three common configuration errors
- When creating a knowledge base in the interface, the locally deployed ChatGLM2 model cannot be selected, and a prompt indicates that no available model is found. The cause is that
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENare not correctly filled in the FastGPT configuration file, or the model accessed via OneAPI is not correctly mapped to the FastGPT model list. - Long inspection reports returned via the HTTP API cannot be passed to the model in batches, and an input length overrun error occurs. The cause is that appropriate text splitting rules are not configured, and long text is not split into fixed-length segments for sequential submission.
- Tools cannot be triggered during model calls, and the returned results do not include expected parameter queries. The cause is that the
toolChoiceparameter is incorrectly set to false, or function call support is not enabled in the configuration.
How to confirm successful configuration
- Enter the model management page, view the configured model list, confirm that the target domestic model is displayed normally, and click test call to verify connectivity.
- Upload the long text content of a single AOI inspection report, test whether the model can correctly extract abnormal points and link to corresponding batch process parameters, and observe the completeness of the returned results.
- After configuring the text splitting rules, import a long document that exceeds the model input limits, verify that the system automatically splits segments and processes them sequentially, with no input length overrun errors.
- View the model call logs, confirm that the actual values of parameters such as
maxContextandtemperaturematch the configured values, and no default parameters override the set values.
The values provided here are common starting points, and should be evaluated against the relevant samples in actual use.