What This Category’s Data Looks Like
Data for this category originates from wafer fab MES systems, equipment FDC modules, defect detection and metrology equipment, process recipe libraries, experimental databases, and more. Daily documents include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, and yield reports. Fields include process number, process parameter thresholds, defect spatial coordinates, yield values, and more. Units include nanometers, degrees Celsius, batch numbers, and others. Data updates occur synchronously after each wafer batch is generated. Device operating parameters are collected in real time. Batch and yield data are summarized daily.
Constraints on Tool Calling and Plugins From These Characteristics
Data sources are scattered, and update rhythms differ. Tool calling must adapt to real-time pulling and batch synchronization of multi-source data. It must support parsing non-standard format documents such as structured recipe files, SPC charts, and wafer bitmaps. Ultra-long process chains and nanoscale parameters require tool calling to match micrometer-scale threshold precision. It must also support cross-process associated queries. Limited rework windows require tool calling response delay to stay within a reasonable range. This avoids impacting production scheduling. In addition, associated queries for exception handling sheets and defect maps require tools to support semantic matching and spatial coordinate association. This places requirements on plugin context integration capabilities.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Workflow/API tool selection | Automatically match tools, support manual tool specification | Adapt to scenarios including process parameter query, abnormal root cause analysis, yield attribution, etc. Automatic mode reduces manual configuration costs |
| File reading plugin | Support parsing PDF, CSV, PNG formats, maximum 1000 MB per file | Cover core document types such as process flow cards, batch run records, wafer defect maps |
| Workflow/API function-call configuration | Enable forced function calling, input parameters must include process ID and batch number | Accurately associate wafer batch data, avoid generalized queries that deviate from production scenarios |
| API call timeout | Set to 600 seconds | Adapt to delay requirements for real-time device FDC data pulling and cross-process data summary |
| Number of returned entries per tool call | Top 10 entries | Cover cross-process associated data volume, avoid result overload that affects inference efficiency |
| HTTP plugin import | Fill in HTTPS interface address, configure Bearer token authentication | Connect to internal production systems such as MES, ensure data transmission security |
This document is written based on the configuration logic of FastGPT v4.17.0.
Three Common Misconfigurations
- Symptom: An error indicating the model is not supported is returned when calling
toolChoice, and the interface prompts "Invalid tool selection parameter". Cause: The currently used model has not been confirmed to be compatible with thetoolChoicefunction. Some open-source models require additional configuration of tool calling adaptation files. - Symptom: An "Interface format error" prompt appears when importing the HTTP plugin, and configuration cannot be completed, or no results are returned when calling the API after publishing the application. Cause: The interface address was not filled using the HTTPS protocol, or the authentication token and intranet IP whitelist were not configured correctly, resulting in interface verification or access being blocked.
- Symptom: The application returns the original result directly after calling the tool, and does not use it as context for the large model. Cause: The application's direct output switch was not turned off, or the context splicing logic for tool calling results was not configured, causing the tool return data to not participate in large model inference.
How to Confirm Successful Configuration
- Initiate a query that includes a specific process ID and batch number. Check whether the tool correctly pulls the corresponding batch’s run records and process parameters.
- Upload a PNG file of a wafer defect map. Confirm that the plugin can parse defect coordinates and type information from the image.
- Call the API interface. Check whether the returned results include production data pulled by the tool, and do not directly expose original interface fields.
- Test the automatic matching of
functionCallin abnormal query scenarios. Confirm whether the exception handling sheet query tool is called.
The values provided are common starting points and should be measured against the reader's own samples.