What data for this category looks like
Data for this category is generated by manufacturing enterprise change management teams. It includes ECN/ECR forms, change assessment reports, drawings and BOM version documents, process validation reports, transition plans, inventory disposal documents, and customer approval files. Data sources cover PLM, ERP, MES, supplier materials, validation test records, and customer-specific requirement documents. Updates are triggered irregularly for each individual change. Fields include change number, associated material/process version, effective date, and other relevant details. There is no unified fixed unit, and some fields must match the format specifications of their corresponding systems.
Constraints imposed on the tool calling and plugins workflow
The scattered sources and varied formats of data for this category require that tool calling must support integration with interfaces from multiple system types including PLM, ERP, and MES, and adapt to reading both structured forms and unstructured documents. Irregular updates triggered by individual changes require that tool calling must pull the latest data in real time, avoiding static caching that could lead to version mismatches. The chain effects of changes involving design, materials, and processes require that tool calling must support associative queries across multiple related fields. A failure retry and exception alert mechanism must also be configured to prevent production risks caused by data pulling errors. Additionally, customer approval files contain sensitive content, so permission verification rules must be set for tool calling to ensure only authorized users can access relevant interfaces.
How to configure settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Workflow/API tool selection | auto | Adapt to the need to call multi-system tools on demand in change scenarios, automatically matching corresponding interfaces |
| Single tool call timeout | 120–180 seconds | Change data may include cross-system associative queries, reserve sufficient time to pull data from multiple sources |
| Maximum returned results | Top 10 entries | Avoid returning excessive redundant data that exceeds context limits, while covering most core documents linked to changes |
| Number of failure retries | 2 times | Balance the impact of interface fluctuations and system resource consumption, adapting to unstable cross-system call scenarios |
| Data cache duration | Calibrated based on actual testing | Since change updates are irregular, do not set a fixed long cache; adjust based on actual change frequency |
This configuration guidance is based on FastGPT v4.17.0.
Three common configuration mistakes
- Symptom: After configuring plugins, URL image addresses in ECN documents fail to display. Cause: Static resource reading permission was not enabled in the tool calling configuration, or whitelist verification rules for image addresses were not configured.
- Symptom: A connection timeout error occurs when calling a locally created MCP service. Cause: The correct local service port and access path were not entered in the FastGPT plugin configuration, or cross-domain access permission for the local service was not enabled.
- Symptom: After enabling tool calling, the model does not output thought processes. Cause: The tool calling mode is set to forced function calling (
functionCall), and the switch to simultaneously output thought processes was not configured, so the model only executes tool calling logic.
How to verify successful configuration
- Initiate a test query that includes an ECN number, and confirm whether the tool successfully pulls BOM version data from the corresponding PLM system.
- Simulate a call to the local MCP service, and check whether the preset change assessment report content is returned normally.
- View the tool calling logs, and confirm that the failure retry count and timeout settings match the expected configuration.
- Enable the thought process switch, initiate a test, and confirm that the model outputs corresponding thinking logic both before and after calling tools.
The values provided are common starting points and should be evaluated against the user’s own samples.