What This Category of Data Looks Like
Technical documentation for semiconductor wafer fabrication typically includes process sheets, batch run records, equipment recipes, SPC charts, exception handling logs, defect maps, and yield reports. Data sources include MES systems, equipment FDC, defect detection and metrology tools, wafer maps, process recipe libraries, and experimental datasets. Updates are triggered by each wafer batch production run, real-time equipment parameter collection, and daily report generation. Document fields include wafer batch identifiers, nanoscale process parameters, defect spatial coordinates, yield values, and some documents include visual charts and structured numerical data.
What Constraints Do These Characteristics Impose on Tool Calling and Plugins
Multi-source, real-time data requires tool calling to pull data across systems, and adapt to interface formats of different data sources such as MES and equipment FDC. Document structures with nanoscale parameters and visual charts require tools to parse unstructured text and images. A single wafer batch links to many process documents, and rework windows are narrow. This means tool calling delay must stay within a reasonable range, and redundant data must not interfere with review results. Cross-module yield attribution requires linking multi-batch data, so tools must support batch data queries and association verification.
How to Configure the Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Workflow/API tool selection | auto | Adapt to cross-review calling requirements for multiple types of process documents, automatically match tools for different data sources |
| Workflow/API function-call configuration | Enabled | Support structured calling of dedicated query tools for equipment parameters, yield data, and similar content, to improve review accuracy |
| Single tool call return limit | Top 6–10 entries | The number of process documents associated with a single wafer batch is limited, avoiding redundant data interfering with cross-review logic |
| Tool call timeout period | 90–150 seconds | Need to pull multi-source real-time equipment data and defect maps, reserve sufficient interface response time |
| Result format validation switch | Enabled | Prevent unstructured search results or array-formatted data from being unable to participate in the cross-review process |
Three Common Mistakes
- Phenomenon: Tool call concurrency exceeds limits, 503 status codes appear or requests are discarded. Cause: Concurrency configuration is not adjusted based on the production rhythm of a single wafer batch, and the concurrency queue upper limit matching real-time data pulling is not set.
- Phenomenon: Image display incomplete errors occur when calling MCP tools, and visual content of defect maps cannot be loaded. Cause: Tool plugins that support image parsing are not configured, or image path and format parameters are not passed correctly.
- Phenomenon: Google Search returned
array<object>formatted results cannot participate in cross-review, fields are empty or format errors occur. Cause: The result format validation switch is not enabled, and a formatting tool to process structured array data is not configured.
How to Confirm the Configuration Is Correct
- Based on FastGPT v4.17.0, submit a process cross-review request for a single wafer batch, check whether the number of returns in the tool call log matches the configured upper limit.
- Simulate parallel review requests for multiple batches, confirm there are no 503 status codes or request discard errors, and the concurrency configuration matches the production rhythm.
- Import the specified embedding model, test embedding recipe documents containing nanoscale process parameters, confirm no adaptation errors occur.
- Simulate test data returned in
array<object>format from Google Search, submit a review request, confirm the results can be parsed normally and participate in cross-review. - Call the MCP tool to load visual content of defect maps, confirm images are fully displayed without errors.
All provided configuration values are common starting points, and should be evaluated against the reader's own samples.