What this category’s data looks like
The data sources for this category include MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Data update frequencies cover each wafer lot transfer node, real-time collected equipment operating parameters, and daily summarized yield reports. Daily documents include process flow cards, lot run records, equipment recipes, SPC statistical charts, exception handling sheets, defect maps, and yield analysis reports. Fields include lot identifiers, process parameter values, defect location coordinates, equipment operating durations, and more. Some parameters involve nanometer-level precision and millisecond-level time units.
What constraints these characteristics impose on the tool calling and plugins workflow
The data update frequencies for this category cover real-time, lot-level, and daily updates, and include multiple types of structured and unstructured content. This imposes constraints on the timeliness and compatibility of tool calling. High-frequency real-time equipment parameters require tools to support low-latency pulling, to avoid delayed data affecting exception determination. Multi-source heterogeneous data formats, including structured SPC charts and equipment recipes, as well as unstructured wafer maps and defect maps, require plugins to adapt to multiple types of data parsing logic. The cross-process and cross-module data association requirements brought about by the ultra-long process chain require tool calling to support joint queries across data sources. At the same time, it is necessary to match the limited rework window and control the response duration of a single call.
How to set the configurations
The following configurations are based on FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Tool calling mode | functionCall forced mode | Adapt to scenarios where supplier audits require strict adherence to rules, to avoid unintended free-form generated content |
| Single data pull limit | Top 10 lot records | Match the basic data volume of a single wafer lot, to avoid pulling too much historical data leading to timeouts |
| Response timeout threshold | 600 seconds | Match the response window of cleanroom production, to avoid interrupting the audit process due to timeout |
| Multi-source data association switch | Enabled | Support joint queries across MES and equipment FDC, to meet cross-process yield attribution requirements |
| Plugin data parsing rule | Group by wafer lot | Adapt to the lot-based structure of daily documents, to facilitate subsequent exception determination and report generation |
Three common mistakes
- A 422 error code is returned during calling. The cause is that the
functionCallforced mode is not correctly configured, causing the tool calling parameter format to not comply with the verification rules of the third-party API. - A 400 invalid image error is returned during multimodal calling. The cause is that the non-standard binary format of wafer maps is not adapted, and the plugin fails to correctly parse this type of image file.
- The data link tool call fails. The cause is that the written SQL statement does not comply with the field naming and table association specifications of the wafer manufacturing database, triggering database permission or syntax verification.
How to confirm the configuration is correct
- Initiate a tool call test, check whether the returned lot data is grouped by wafer lot and includes associated fields from MES and equipment FDC.
- Simulate a scenario exceeding the single pull limit, verify whether the system automatically truncates excess data.
- View the plugin running logs, confirm that the retry logic takes effect normally when the call timeout threshold is triggered.
- Check that the tool calling mode is locked to
functionCall, to avoid generating free-form content that does not comply with audit rules.
The values provided are common starting points and should be measured against the reader's own samples.