What Data for This Category Looks Like
Data for this category is primarily collected from wafer test maps, Manufacturing Execution System (MES), bonding and molding equipment, ATE testers, X-ray/SAM inspection equipment, and customer specification documents. Daily maintained documents include packaging process cards, bonding and molding parameter tables, incoming wafer maps, test program version files, yield reports, and failure analysis reports. Data update frequencies cover per batch of materials, per chip test, and daily summary. Fields include chip batch number, test bin distribution, equipment parameter thresholds, and more. Adaptation to multi-source heterogeneous data structures is required.
How These Characteristics Create Constraints for Tool Calling and Plugins
Multi-source heterogeneous data sources require tool calling plugins to support integration with industrial device APIs, MES interfaces, and document parsing tools, adapting to different communication protocols and data formats. The high-frequency update frequency requires that tool calling context length be strictly limited, to avoid model context overflow caused by excessive data volume. Diverse document types and fields require tool parameters to cover specific fields such as chip batch number, test program version, and yield indicators, to ensure accurate extraction of required information during audits. Additionally, support for alignment operations between wafer maps and physical batches is needed, to match the association between incoming material information and test results.
Configuration Settings
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Maximum Context Length | 8000–12000 token | Single batch of packaging and testing data has a large volume, to avoid truncation caused by exceeding the model's bearing capacity |
| Tool Calling Retry Count | 2–3 times | Industrial device calls may experience temporary network fluctuations, reducing invalid process interruptions |
| Tool Calling Timeout | 30–60 seconds | Equipment such as bonding machines and ATE testers require a certain amount of time to respond, and excessive time will block the overall audit process |
| Workflow/API tool selection | auto | Automatically select the tool suitable for the current task, no need to manually specify the calling object |
| Workflow/API function-call configuration | strict | Strictly verify the parameter format of tool calls to ensure compliance with the field requirements of packaging and testing documents |
| Plugin Data Format Validation | Enabled | Align the field specifications of structured data such as incoming wafer maps and test program versions |
Three Common Misconfigurations
- Context overflow error occurs during tool calling, with logs showing content length over limit. The cause is failure to limit maximum context length; data such as yield reports and failure analysis reports for a single batch of packaging and testing has a large volume, exceeding the model's bearing capacity.
- Tool call parsing fails, returning Tool call Parser related errors. The cause is that the content returned by the inference model is wrapped in think tags and placed in the content field, which does not meet the preset tool call format requirements.
- Forwarding error occurs after configuring the MCP server, returning a long error code. The cause is that the aiproxy key forwarding rules are not configured correctly, causing the model request to fail to be forwarded normally to the target service.
How to Verify Proper Configuration
- Submit simulated data for a single batch of packaging and testing, check whether the tool call can automatically trigger the corresponding data extraction or device query action.
- View the tool call log, confirm that the parameter format meets the preset field requirements of the packaging and testing documents, and there are no parsing errors.
- Test the timeout and retry logic of tool calls, simulate temporary network fluctuations, and check whether the retry can be completed according to the configuration and valid results are returned.
- Compare the configured context length with the actual size of a single batch of data, confirm that no out-of-limit situation occurs.
The values provided are common starting points and should be evaluated against the reader's own samples.